WEBVTT

1
00:00:29.390 --> 00:00:31.980
Angelo Dragone (SLAC): Looks like there's only one person here.

2
00:00:32.540 --> 00:00:33.540
Angelo Dragone (SLAC): Hello?

3
00:00:34.360 --> 00:00:35.570
Angelo Dragone (SLAC): Hey, Gregor.

4
00:00:36.350 --> 00:00:38.150
gdeptuch: Hey, hey, and…

5
00:00:38.150 --> 00:00:41.650
Angelo Dragone (SLAC): It's just the two of us.

6
00:00:41.650 --> 00:00:44.270
gdeptuch: So far! So far!

7
00:00:44.270 --> 00:00:49.560
Angelo Dragone (SLAC): So far. Oh, it's early. Okay, I'm 3 minutes early.

8
00:00:49.560 --> 00:00:52.460
gdeptuch: Two minutes, or 2 minutes, probably.

9
00:00:52.460 --> 00:01:07.049
Angelo Dragone (SLAC): Okay, okay, okay. I thought, I thought for a second that there was something not… not… not working. I'm in the wrong meeting. He said, no, it's me, so I'm early.

10
00:01:07.510 --> 00:01:10.420
gdeptuch: No, you… but… but there was few…

11
00:01:10.570 --> 00:01:24.410
gdeptuch: meetings, I mean, the Zoom addresses, you know, views, so frankly, I don't know if we're 100% on the right one, because it was one for SAPI, and then it was another for, you know, kind of…

12
00:01:24.410 --> 00:01:33.269
gdeptuch: general meeting of everyone, so we have to be sure, because if there are only two of us on this one, maybe there are three. Hello!

13
00:01:33.780 --> 00:01:34.590
gdeptuch: Good ocean.

14
00:01:34.930 --> 00:01:37.859
gdeptuch: So probably we are in the right one.

15
00:01:39.050 --> 00:01:44.499
Oceane Bel (PNNL): Oh, I just clicked on the one that came out with the notification, so I hope it's the right one.

16
00:01:44.740 --> 00:01:48.739
Angelo Dragone (SLAC): I have only this one on my calendar, so must be.

17
00:01:53.120 --> 00:01:57.730
gdeptuch: Okay, so we have to wait for others to join us.

18
00:01:57.980 --> 00:02:00.999
Oceane Bel (PNNL): Today's ELFO presentation, woo!

19
00:02:01.000 --> 00:02:04.459
gdeptuch: Supposed to be! Let's see.

20
00:02:04.460 --> 00:02:05.969
Oceane Bel (PNNL): Come on, positive, right?

21
00:02:06.380 --> 00:02:13.130
gdeptuch: Yeah, you know, if nobody else is joining, we will have a private screening for Anteloch.

22
00:02:13.490 --> 00:02:16.719
Oceane Bel (PNNL): There you go.

23
00:02:23.880 --> 00:02:26.449
Angelo Dragone (SLAC): Well, let me see if that's.

24
00:02:26.450 --> 00:02:34.859
gdeptuch: We are in… we are dangerously close to… It's at 6pm.

25
00:02:38.250 --> 00:02:43.850
gdeptuch: But Angela, I haven't seen any email from Paul, effectively…

26
00:02:43.850 --> 00:02:49.799
Angelo Dragone (SLAC): Then I'll see an email from Paul. I see Jim, okay, at least that is Jim.

27
00:02:49.800 --> 00:03:00.199
gdeptuch: This is our team, right? This is our team, at least. So I'm not surprised they were connected to this meeting, but I haven't seen any email from either…

28
00:03:00.300 --> 00:03:02.070
gdeptuch: Paul, or…

29
00:03:02.450 --> 00:03:04.130
gdeptuch: Michelle, I think, right?

30
00:03:05.760 --> 00:03:12.109
Angelo Dragone (SLAC): Well, I can always at least send a reminder to the outage people. That's easy enough.

31
00:03:13.300 --> 00:03:16.270
gdeptuch: It would be strange, frankly, because…

32
00:03:16.450 --> 00:03:30.329
gdeptuch: last meeting, or no, before the Code D presentation, we're clearly saying that the schedule is 15 with the Code D, and 22nd is helpful. Oh, there are people joining, okay.

33
00:03:32.190 --> 00:03:34.610
gdeptuch: Now he's the… now everybody's booking up.

34
00:03:37.090 --> 00:03:38.850
James (Jim) Ang: Yep, and there's Paul.

35
00:03:38.850 --> 00:03:39.190
jsnelso: Okay.

36
00:03:40.430 --> 00:03:42.789
gdeptuch: So, when we have auto, we have Quorum.

37
00:03:46.680 --> 00:03:49.789
Paul McIntyre: Okay, looks good. People joining?

38
00:03:54.340 --> 00:03:57.899
Angelo Dragone (SLAC): And a reminder for our race, we'll see.

39
00:04:01.960 --> 00:04:06.639
'Pops' / NY CREATES: Hey, Pops. Hello. Sorry, I haven't been able to join earlier, but

40
00:04:07.020 --> 00:04:09.639
'Pops' / NY CREATES: I keep… I keep missing it. Hi, Jim.

41
00:04:13.860 --> 00:04:18.409
Angelo Dragone (SLAC): Why… why we wait for other people to join? Maybe a question.

42
00:04:18.529 --> 00:04:20.580
Angelo Dragone (SLAC): Are we gonna have our,

43
00:04:21.000 --> 00:04:27.090
Angelo Dragone (SLAC): nearcut, annual meeting, because, I know that, the other,

44
00:04:27.410 --> 00:04:33.139
Angelo Dragone (SLAC): gather up, have, have, have started doing that. For instance,

45
00:04:33.250 --> 00:04:36.090
Angelo Dragone (SLAC): China at their first annual meeting.

46
00:04:39.950 --> 00:04:42.580
Paul McIntyre: We, we had talked about it, I think at.

47
00:04:42.580 --> 00:04:44.210
Angelo Dragone (SLAC): I was not… sorry.

48
00:04:44.210 --> 00:04:57.459
Paul McIntyre: Yeah, at least one of our, one of our prior meetings, but we, we didn't… I don't think we quite had a quorum of, of the, of the PI team present for that. We, we talked about,

49
00:04:57.660 --> 00:05:15.190
Paul McIntyre: we didn't really get into timing. We talked a little bit about format, and I think there was some, some consensus, or some, you know, belief expressed that it might make sense to do this

50
00:05:15.440 --> 00:05:24.180
Paul McIntyre: with a subset of PIs, rather than the whole PI team, because it would… that might make it a little unwieldy.

51
00:05:24.270 --> 00:05:36.189
Paul McIntyre: And also, we might want to do it in, proximity to, to Germantown, to, make it easier for

52
00:05:36.190 --> 00:05:48.080
Paul McIntyre: the DOE staff to attend. If, if that's the objective of the review, is to have some interaction with them, because they often find it difficult to travel.

53
00:05:48.180 --> 00:05:51.810
Paul McIntyre: But of course, with the government shutdown.

54
00:05:52.200 --> 00:05:55.460
Paul McIntyre: you know, we… I think we'd probably need to wait

55
00:05:55.570 --> 00:06:03.750
Paul McIntyre: In any case, until that's over, for public fac… for public-facing or quasi-public-facing events.

56
00:06:03.840 --> 00:06:20.119
Paul McIntyre: or events that require a lot of expenditure of government funds to host, that the timing might not be right for that while there's a shutdown. At least that would be my concern, I guess. But that doesn't mean we can't start planning it now.

57
00:06:23.200 --> 00:06:36.079
gdeptuch: But also, we were talking about having pretty much clear directives, maybe, from our program managers, what they would be expecting from us when reporting, because there is no really reason to have such

58
00:06:36.080 --> 00:06:44.010
gdeptuch: meeting without knowing what to report, and having, you know, people from DOE. I mean, this was also mentioned last time that we…

59
00:06:44.270 --> 00:06:50.669
gdeptuch: we… I think someone was supposed to figure it… figure it out. Oh, maybe Valerie was supposed to look at that.

60
00:06:50.800 --> 00:06:52.180
gdeptuch: We, we…

61
00:06:52.180 --> 00:06:55.279
Valerie Taylor: What, what was, what was I supposed to…

62
00:06:56.060 --> 00:07:09.190
gdeptuch: But what DOE program managers are expecting, frankly, from us, is reporting materials when we plan our annual meeting for Meerkat.

63
00:07:09.460 --> 00:07:19.310
Valerie Taylor: Oh, that part. So I, you know, I did talk to Pavel when he… because he was in the area and stopped by, Argonne.

64
00:07:20.180 --> 00:07:21.530
Valerie Taylor: So,

65
00:07:21.960 --> 00:07:38.539
Valerie Taylor: it's unclear when they're going to request, a report from us, and so, you know, and I was telling him that we're just moving forward with our collaboration plan as to what was described in the plan, and he was like, great.

66
00:07:38.640 --> 00:07:44.539
Valerie Taylor: And that was all we said, and then he moved on to the American Science Cloud.

67
00:07:46.100 --> 00:07:59.830
Valerie Taylor: So we had, you know, we did give them an update on what we're doing with BEA, but in terms of, you know, looking also,

68
00:07:59.860 --> 00:08:06.180
Valerie Taylor: At the centers in general, yeah, that was just, like, a 2- or 3-minute conversation.

69
00:08:06.180 --> 00:08:06.760
Paul McIntyre: Yep.

70
00:08:07.740 --> 00:08:23.390
Paul McIntyre: Yeah, I think it's difficult to get their attention. I mean, even with the government shutdown business, they were very… I mean, we've been waiting for quite a while to get… to receive feedback on our original collaboration plan.

71
00:08:23.660 --> 00:08:41.850
Paul McIntyre: in ESTEAM, we've been collecting up papers and… and trying, you know, trying to get things, sort of ready so that we're… we… we… when… if a report request comes, we won't… it won't be that difficult for us to respond to it. But,

72
00:08:41.960 --> 00:08:55.830
Paul McIntyre: Yeah, we ought to think about having, you know, having a plan for a, for a review, or a, at least an, at least for an annual meeting, and, how we want to structure that.

73
00:08:56.490 --> 00:09:03.719
Paul McIntyre: Maybe we can make that the topic of our next PI meeting, is just trying to work through that.

74
00:09:04.020 --> 00:09:14.299
Valerie Taylor: Right, and what about the working groups? Because I think we were all looking at names for the working groups, so that we can at least move forward with the four working groups.

75
00:09:15.030 --> 00:09:25.140
Valerie Taylor: have forward movement there. Such that… and I think what we had discussed previously was to have

76
00:09:25.160 --> 00:09:40.189
Valerie Taylor: One person from each project serve on an initial working group team, and that way, that team can lay out a plan for that working group, maybe over the next 18 months.

77
00:09:40.370 --> 00:09:54.270
Valerie Taylor: And then, using that plan, we get others engaged, and they can use that plan as a starting point, but it's much easier to work with 8 people than 30 people.

78
00:09:54.370 --> 00:09:59.210
Valerie Taylor: When the… when the page is blank. So…

79
00:10:01.400 --> 00:10:17.019
Paul McIntyre: Yeah, so let's, let's have these as our two agenda items for our next, meeting, next, next week. And, as, as you know, we're going to launch a new, meeting, series, with a different, Zoom link.

80
00:10:17.610 --> 00:10:23.380
Paul McIntyre: And and that… you'll… that will be coming in the next day or so.

81
00:10:24.110 --> 00:10:31.600
Valerie Taylor: Okay, and so that Zoom meeting, that's where we have the PIs and also a deputy or alternate…

82
00:10:32.140 --> 00:10:35.030
Valerie Taylor: So, it's two people per project.

83
00:10:35.030 --> 00:10:35.720
Paul McIntyre: That's right.

84
00:10:35.720 --> 00:10:36.110
Valerie Taylor: Okay.

85
00:10:36.110 --> 00:10:36.630
Paul McIntyre: Yeah.

86
00:10:38.350 --> 00:10:39.050
Valerie Taylor: Yes.

87
00:10:39.050 --> 00:10:39.890
Paul McIntyre: Sounds good.

88
00:10:39.890 --> 00:10:46.289
gdeptuch: Okay, so maybe, maybe we switch to our L4, because we are… we are working… we're waiting to start.

89
00:10:46.290 --> 00:10:59.650
Paul McIntyre: Yes, I know, and we want to make sure you have enough time, Gregor, so, please, so the purpose of this meeting is to see, to have an overview of the ELFO team and what their plans are.

90
00:11:00.430 --> 00:11:04.729
gdeptuch: Yes, so hopefully, hopefully you can see my screen.

91
00:11:05.160 --> 00:11:06.680
gdeptuch: Can you see my screen? Yep.

92
00:11:06.680 --> 00:11:07.610
Valerie Taylor: Looks good.

93
00:11:07.610 --> 00:11:11.310
gdeptuch: Yes. So, we'll be switching between,

94
00:11:11.310 --> 00:11:33.720
gdeptuch: tasks during the presentation, but I will start with introduction and covering Trust 1, and then my colleagues will also be covering Trust 2, Trust 3, and maybe let's start with that. So, the outline of the presentation is, first part gonna be about project organization.

95
00:11:33.720 --> 00:11:47.709
gdeptuch: Some words of introduction, motivation, discussing electrophotonic interfaces and silicon photonic platforms that were label polarization of the project. Then, at high level, project objectives and trust will be covered.

96
00:11:47.710 --> 00:11:55.690
gdeptuch: And then project design, stack, and tools, because we are actually covering several layers of, of approaches from,

97
00:11:56.070 --> 00:12:21.049
gdeptuch: materials up to the system level, and then we covered three trusts of, within ELFO, Trust 1, Trust 2, Trust 3, then a bit testing… testing investment in equipment, because it's essential for this project. So I don't have to remind you how this project actually was generated, because we have… we are part of the… of the Meerkat Center. The Meerkat Center, we have

98
00:12:21.050 --> 00:12:24.819
gdeptuch: projects, from the…

99
00:12:24.820 --> 00:12:46.490
gdeptuch: In ELFO, we have collaboration from… collaborators from Brookhaven, Columbia University, CUNY, CCNY, and FIU, Florida International University, and PNNL. We have a great industrial liaison, New York Create, represented here by COPS, and also we have an industrial collaborator with Cadence.

100
00:12:46.490 --> 00:12:55.879
gdeptuch: So you can look a bit, some details at our websites that are under construction, of course, one about ELFO, and another one

101
00:12:55.880 --> 00:13:09.180
gdeptuch: about Newcastle Center. Generally speaking, the second one is under construction. We are actually now getting some help from a company to get it, you know, quicker up to the level that could be good for our center.

102
00:13:09.180 --> 00:13:26.319
gdeptuch: So, maybe let's jump to motivation, because whenever you have any commercialization, you know, training, you're always being told you should address customer needs, and our, you know, working on developments of highly granular detectors.

103
00:13:26.320 --> 00:13:42.459
gdeptuch: for the electron-ion collider that is a flagship project at Brookhaven, we realized that we are suffering in such great-scale systems from congestion in delivering resources and taking data.

104
00:13:42.460 --> 00:14:05.100
gdeptuch: So this… what is shown here in red is all that congestion of wires, cables, delivering power, taking signals out, and basically, some areas have already more… more material than is expected, and this object is very, very complex. We are using large-scale silicon detectors that are bent.

105
00:14:05.100 --> 00:14:29.899
gdeptuch: to make this… make them thin, so the level of complexities that are present here is gigantic. And particularly in the center of the detector, we have so-called monolithic active pixel sensor, silicon vertex tracker, that is made of large-scale detectors as one big part of the silicone wafer. This is a 12-inch wafer, and one sensor is effectively marked

106
00:14:29.900 --> 00:14:48.279
gdeptuch: as this red rectangle, and here there are multiple units producing a lot of data, and they are, from one such long tail, we have six 10 gigabits per second drivers. Okay, so we started scratching our heads how to actually handle all that.

107
00:14:48.300 --> 00:14:53.080
gdeptuch: And the first place we were thinking about jumping into

108
00:14:53.080 --> 00:15:18.029
gdeptuch: photonics, very close to the endpoint of such long slab of monolithic sensor. We are thinking about delivering power through light, and also taking data using photonics approach. So we are thinking about designing a photonic IC that will be, you know, taking the 6-bit… 6

109
00:15:18.030 --> 00:15:39.199
gdeptuch: times 10 gigabit per second link out of the detector. So we realized that there are existing good commercially mature processes where you can do silicon photonics, already. So this was actually our starting point, thinking about the L4 project, or L4 scope.

110
00:15:39.350 --> 00:15:53.160
gdeptuch: And then we realized that, effectively, when you start doing wavelength division multiplexing that is very advantageous for transferring data, you can actually obtain very high

111
00:15:53.160 --> 00:16:17.950
gdeptuch: speeds of transmission, but using multiple links. Here, the plot shows, for example, this is a borrow from our collaborators from Colombia, that if we go with multiple wavelengths, which they typically are multiplex of single photonic fiber, we can actually get speeds that are very, very high. At the same time, where we can save a lot of energy.

112
00:16:17.950 --> 00:16:25.980
gdeptuch: So this is actually what we somehow previously, already we tried to actually address by…

113
00:16:28.350 --> 00:16:46.620
gdeptuch: multiple streams of 10 gigabits per second on… using different wavelengths. And that wavelength division multiplexing will be appearing through our presentation several times, because this is underlying technology for also terraforming infotron computing. We'll be talking later.

114
00:16:46.620 --> 00:16:58.260
gdeptuch: Okay, later… now when you look at the photonic system, how they are constructed nowadays, when you have… you have typically photonic photonic elements sitting a bit on the side of

115
00:16:58.260 --> 00:17:22.590
gdeptuch: of electronic circuits, when you have pluggable optics, fibers for data transmission. But if we think about embedded photonics, which is, you know, subject of our project, we can… we bring the photonics inside the hybrids, having, you know, all the elements ranging from memory, processors, drivers, sensing, all on one stack.

116
00:17:23.670 --> 00:17:24.880
gdeptuch: So…

117
00:17:25.000 --> 00:17:49.909
gdeptuch: knowing that we have some experience already in the past in hybrid 3D integration, I mean, for example, you know, we've been involved in development of 3D stacked, fully fusion-bonded detector that was composed of the sensor, analog layer, digital layer. We started already having, you know, oh, we have something in our pocket that we have a

118
00:17:49.910 --> 00:18:12.229
gdeptuch: hybridization in 3D stacking. Also, when we have, our collaborators working, working on high-speed, multiplexing on data using waiver and division multiplexed principle, we can have processing, added, by having electropical modulators, and then in the, in the

119
00:18:12.230 --> 00:18:31.250
gdeptuch: the flow of data before even reaching the final destination when data is stuck. Okay, so we realized we have components that are available in our hands, so we thought that Alpha project can actually have such a structure. We may have a generic detector that is instead

120
00:18:31.250 --> 00:18:55.350
gdeptuch: reading read out electrically can be read out photonically, but having conversion to the photonics as close as possible to the signal generation part, then we… when we transfer photonic data out, we can perform already significant level of processing, on-the-flight processing in the photonic domain, using advantages of addition of light, and also using the wavelength multiplex, as I said earlier.

121
00:18:55.350 --> 00:19:19.089
gdeptuch: But later, towards the end, we can also equip such a data acquisition system in electrophotonic… high data throughput processing, where you can have neural networks, non-phenolmon processing embedded, and this is why we trust… we created three trusts.

122
00:19:19.090 --> 00:19:29.159
gdeptuch: Trust one is electrophotonic detector, Trust 2 is electro… is on-the-flight data processing, and trust three is electrophotonic processors, and those are the three thrusts that we have in our…

123
00:19:29.160 --> 00:19:53.320
gdeptuch: alpha project. So Trust 1 focuses on devices and interfaces, system characterization and modeling, application-driven device optimization. Trust two is on-the-fly data processing with a focus on enables and interconnects, partitioning of data, because certain functions which will be discussed later, can be optimally realized in electrical domain, certainly in photonic domain. And trust three is

124
00:19:53.320 --> 00:20:16.609
gdeptuch: processor operation beyond von Neumar architectures, where we harness advantages of electrical and photonic signals. And these three trusts are tied together with, you know, a co-design approach that is heterogeneous integration, high bandwidth interfaces between Trust 1 and Trust 2, and co-design architectures between Trust 2, Trust

125
00:20:16.610 --> 00:20:41.410
gdeptuch: and trust three, and all of us, we are aiming towards energy efficiency in an extreme environment, as was our funding proposal plan. So, projects, objectives, and trusts, as I mentioned, we have myself, I'm the leader of the L4, we have great industrial liaison, we have project manager, we have Trust One, led by, you know.

126
00:20:41.410 --> 00:20:55.689
gdeptuch: myself and Alex from Colombia, Trust 2, led by Sumi Ajit, who is present in the room, and Karen, who couldn't join us today, unfortunately, and Trust 3, led by Jim and Prashanta, and also we have Ocean connected to us.

127
00:20:55.690 --> 00:21:13.330
gdeptuch: And we identified major technological components, because without that, it's difficult to convey such scale of projects, ranging from material to system. We work with Infotonics, Global Foundries, Tower Selling Conductor, especially the last one is for the sensing part.

128
00:21:13.910 --> 00:21:22.879
gdeptuch: So, we were thinking significantly, you know, from where we should be actually starting, and we realized that existing

129
00:21:22.880 --> 00:21:46.839
gdeptuch: commercially-grade electrophotonic platforms are already equipped with a lot of elements that could be harnessed by us, and we can start constructing the systems. Also, we have having access to more, like, laboratory-scale silicon photonic platforms, and also beyond silicon photonics platform, we can enhance this portfolio.

130
00:21:46.840 --> 00:21:56.450
gdeptuch: And I think from electrical photonic modulation, certainly we have in existing platforms electro-absorption modulators.

131
00:21:56.450 --> 00:22:21.289
gdeptuch: phase interferometric modulators. We have also some much more advanced hybrid electrical modulators. Going back from photonic to electrical, we have direct photodiodes based on… also on Germanium. We have balanced photodiodes, balanced photodetectors, heterodine, coherent detector, all of that is available on the platform.

132
00:22:21.290 --> 00:22:24.160
gdeptuch: Could be used for our needs.

133
00:22:24.160 --> 00:22:44.640
gdeptuch: And then application in data computation, we have high-speed data links, we have models for unomorphic artificial intelligence accelerations, you know, sensors, so all that enables, actually, actually what we realize, closed-loop electrophotonic systems for transmission, computation, and sensing, and this is where we embedded our tree trusts.

134
00:22:45.480 --> 00:22:55.899
gdeptuch: So, the stack… L4, design stack and tools, in L4 project, we are starting with materials and devices, going up, particularly reaching algorithms and applications.

135
00:22:55.900 --> 00:23:19.739
gdeptuch: And at each level of this stack, we need to have different tools, and the tools are essential, because when we're designing systems, we need to have models ranging from high-precision device, going all the way up to high abstraction levels. So here are listed the tools

136
00:23:19.740 --> 00:23:44.159
gdeptuch: And boundaries between trusts are, you know, shared between certain types of the tools. So we have multi-physic device modeling we exploited, mixed electrophotonic circuit simulators, analysis of EM, ER, signal and power integrity, and then on the top, we have processing architectural modeling, functional simulator for

137
00:23:44.160 --> 00:24:00.809
gdeptuch: simulation for the design. So today, tools versus needed tools, because we… certainly the very mature are tools for electrical design, so there are vendors available… making tools available, but photonics in the AI is still fragmented and immature.

138
00:24:00.810 --> 00:24:11.689
gdeptuch: So this is also what we plan to cover a bit in our Alpha to gap, to cover this gap with our approaches that we'll be aiming at development of the tools.

139
00:24:11.900 --> 00:24:25.400
gdeptuch: So, this, maybe, slide is a bit bulky. I will… I will keep going into details, but this can be taken as a reference. So, we have… we put on the left side existing industry-grade tools.

140
00:24:25.400 --> 00:24:37.960
gdeptuch: So this is for stable, you know, this is very stable for silicon photonic… silicon photonic processes, but what is missing is uncharted territory that we are planning to cover in L4s on the right side.

141
00:24:37.960 --> 00:25:02.760
gdeptuch: So I may probably skip it in terms of time to allow, you know, covering more our trusts, but this is, you know, a good compendium of kind of market research, also on this rich portfolio, and sometimes also quite expensive if it goes to commercial tools packages. But when we want to bridge the gap between photonics.

142
00:25:02.760 --> 00:25:15.320
gdeptuch: circuits and electrical circuits, and going all the way up to high level of abstraction, the tools are largely missing, and we will be… we are having this also hung up addressed in our project.

143
00:25:16.050 --> 00:25:25.560
gdeptuch: So, maybe now going to Trust 1, I will still cover Trust One, and then I will give microphone to my colleagues. So, in Trust One, as I mentioned, this…

144
00:25:25.600 --> 00:25:35.890
gdeptuch: Focused on electrophotonically readout, sensors, and the inspiration for this, development is, was effectively,

145
00:25:35.890 --> 00:25:46.249
gdeptuch: It's coming from the fact that in photonic domain, we can realize quite easily linear operations, where we have

146
00:25:46.250 --> 00:25:52.260
gdeptuch: analog values.

147
00:25:52.260 --> 00:26:17.160
gdeptuch: that are represented in amplitude of light, coupled to other waveguides using couplers, resonant circuits, that we can assign different weights to them, and by doing this, we can actually construct XY readout that can be fully going on the photonic domain. So this is, like.

148
00:26:17.160 --> 00:26:40.639
gdeptuch: kind of transfer a bit of knowledge or experience from computing in photonics to developing fully read out in photonic manner detector, like monolithic active pixel sensors. And of course, we have to go through several layers, starting from the front-end layer, which is detector sensor itself.

149
00:26:40.640 --> 00:27:05.530
gdeptuch: Then the conversion layer to… from electrical to photonic domain, data transport, and application. So, having all these elements achieved in Trust 1, we can develop monolithic active pixel sensor with 3D stacked photonic reader, and our idea is shown maybe on this slide more… in more details. So, when we have the first layer physical of sensor and analog front-end

150
00:27:05.530 --> 00:27:08.499
gdeptuch: That is, shown here.

151
00:27:08.500 --> 00:27:10.650
gdeptuch: That, this, layer

152
00:27:10.650 --> 00:27:34.759
gdeptuch: Of course, detect signals, and it feeds electro-optical modulators that are receiving signals from the wave… from waveguides, where we have multiple wavelengths sent over there, and then we have these electro-optical modulators, sensitive to specific wavelengths, and coupling to the output fiber, where we can read already modulated light.

153
00:27:34.760 --> 00:27:59.169
gdeptuch: So if there is signal, for example, in a given channel, we can encode addresses, and we can read at unprecedented speed, uncomparable to whatever is possible to be achieved in the electrical domain. So the cross-sections are also shown here, that's what we plan, so having, you know, of course, 3D stacked approach with sensor.

154
00:27:59.220 --> 00:28:07.160
gdeptuch: Electrical layer and electro-optical layer is added in the stack.

155
00:28:07.970 --> 00:28:31.919
gdeptuch: how we plan to actually realize this, the tasks. So, we can fabricate, of course, sensors or electrical layers on the full-scale wafers, and this is why we have also tower in our equation, that we can develop sensing as a monolithic system, and add photonic circuits using chip on wafer stacking.

156
00:28:31.920 --> 00:28:34.940
gdeptuch: To have photonic layer attached to

157
00:28:35.320 --> 00:28:41.560
gdeptuch: to the electrical layer. Of course, one part of our Trust One is also development of

158
00:28:41.560 --> 00:29:05.550
gdeptuch: very efficient coupling that is suitable for an extreme environment, including cryogenic temperatures, and also another part of this trust one is multi-physics simulation and modeling, because we are not only merging layers as heterogeneous materials, but also

159
00:29:05.550 --> 00:29:19.019
gdeptuch: two different domains, electrical and photonic, so all that needs to be carefully modeled, and we will… it's also solidly planned in Trust… Trust One. So,

160
00:29:19.070 --> 00:29:42.710
gdeptuch: optimization of the process, of the process, and this is where we have advantage of working in laboratory scale, maybe not laboratory scale, not commercial-grade foundry, where we can optimize also basic components, such as resonant microdisk modulators, and here is an example of work performed by Columbia.

161
00:29:42.710 --> 00:30:07.439
gdeptuch: we're optimizing microdis modulators by having some undercut technique. We can increase sensitivity of tuning. This actually is shown by having plot different powers used for, modulation, how wide the spectrum of modulation, wavelength

162
00:30:07.440 --> 00:30:12.120
gdeptuch: resonance we can obtain, so I think this is…

163
00:30:12.390 --> 00:30:30.059
gdeptuch: advantage of working with a founder that is not only hermetically closed for achieving, you know, data transmission, but also on R&D, and this is part of our Trust One efforts.

164
00:30:30.060 --> 00:30:33.519
gdeptuch: And probably with… with this, I will pass,

165
00:30:33.930 --> 00:30:57.549
gdeptuch: battle to Trust 2, and Sumi Ajit will cover Trust 2. Okay, so to continue, so once the data has been generated, so the thrust 2 is, the… well, the focus of Thrust 2 is to, process the data, while it's moving, so moving away from the idea of the interconnect just being a passive interconnect in, like, a wire, to see if it can do some processing.

166
00:30:57.590 --> 00:31:21.869
gdeptuch: That is useful. So, in the… as Irish already mentioned, the… the main, enabler here is wavelength vision multiplexing, because we have this additional degree of freedom in optics because of the huge bandwidth that we can actually pack thousands of, hundreds to thousands of, wavelengths, each modulated individually at pretty high rates, you know, tens of gawbits per second.

167
00:31:21.870 --> 00:31:46.840
gdeptuch: to a single fiber, and this gives us very, very large bandwidths. And the way this typically works is that you have a continuous wave laser that gets modulated by these micro-resonators, these are group velocity dispersion tweaked micro-resonators, you get this curve effect, and that gives you this comb. The comb can then be individually de-interleaved through this tree network, so each of these gives you a splitter, and then the splitter is going to

168
00:31:46.840 --> 00:31:52.890
gdeptuch: geometry. So then once that has been split up, each of these individual wavelengths can then be modulated.

169
00:31:52.890 --> 00:32:12.940
gdeptuch: And then you can recombine them. So now you have a bunch of modulated tones that are on the same fiber. These can be translated over, I mean, transferred over the fiber, single-mode fiber typically, and then de-interlaced again at the other end, detected, and then passed out further processing, which could be done electrically, or could be further

170
00:32:13.090 --> 00:32:37.340
gdeptuch: optical processing if you don't put the photodiodes. So… so this gives you… this is the generic framework that you can use, but then the work done at Columbia, in particular, Karen's group, has really gone, taken this to the next level, to… to look at how efficiently you can actually transfer data using this, and the results together are pretty promising. You can get less than a picojoule per bit, and multi-bits per link.

171
00:32:37.340 --> 00:32:39.050
gdeptuch: terabits per second per link.

172
00:32:39.300 --> 00:32:48.619
gdeptuch: Next slide, Brian. So, in this context, the additional degree of freedom that the wavelength division multiplexing WDM gives you

173
00:32:48.620 --> 00:32:49.460
gdeptuch: is…

174
00:32:49.460 --> 00:33:14.389
gdeptuch: is that you get this additional degree of freedom for processing data, and the one way to look at this is to compare how you might do a very common operation, such as a convolution operation, in a traditional, let's say, digital accelerator, like a GPU, versus an analog in-memory accelerator, which is shown in the middle, versus a photonic tensor cord, which is shown on the right. And here, the tensor really means that now you have this additional degree of freedom in your vector, which means

175
00:33:14.390 --> 00:33:32.849
gdeptuch: another dimension, which you can use to do multiple matrix vector multiplications in one step, as opposed to a traditional approach, or digital, which is sequential, or even in-memory computing, where you do it in one step. But now you can go beyond that and do n of these operations in parallel. So it's massive parallelism.

176
00:33:32.850 --> 00:33:35.439
gdeptuch: Very, very high computational throughput.

177
00:33:36.820 --> 00:33:49.019
gdeptuch: So something that Karen's group has been working on is to take advantage of this, couple this with another key enabling technology here, which is a waven-selective Switch. So, WSS is a waven-selective Switch.

178
00:33:49.020 --> 00:34:13.760
gdeptuch: And this is basically the analog of a router, but in the optical domain. And the key idea here is that it's, actually, you can take any wavelength at one of the input ports and route it to one of the other output ports. And this gives you switching, arbitrary switching inside an optical network. Now, the interesting thing is this is a fully passive,

179
00:34:13.760 --> 00:34:25.730
gdeptuch: passive structure, so, you know, in terms of power consumption, it doesn't add, but then it's also very high, data rates, because, again, you're doing it with fiber. So, compared to an optical… sorry, compared to an electrical router, you get these advantages.

180
00:34:25.800 --> 00:34:45.340
gdeptuch: And, this lets you conceptually now build up a sort of a hierarchical network where you have this, these, modules, which they call B-cubes, which have a computational agreement, the CU. Right next to it is one of these, is a silicon Photonics I.O. with.

181
00:34:45.340 --> 00:34:51.800
gdeptuch: one of these, WSS, multiplexing, several of these cubes together.

182
00:34:52.020 --> 00:35:03.519
gdeptuch: Now, you can take this structure, which is just a one layer, and recursively extend that to hierarchically as many layers as you like, and this lets you build up these very high bandwidth networks with

183
00:35:03.550 --> 00:35:15.310
gdeptuch: optical back… optical switching of the… optical circuit switching of the connections. And this can… I mean, they have demonstrated experimentally and also chosen simulations that this typically gives you something like.

184
00:35:15.310 --> 00:35:33.980
gdeptuch: 10x more, bandwidth per, say, watt, or 10x more energy efficiency for data transmission than an equivalent of a state-of-the-art InfiniBand or other wired electrical network. So this can be very useful if you're doing operations that require a large amount of,

185
00:35:33.980 --> 00:35:44.380
gdeptuch: Data transfer, particularly, for example, if you're doing distributed training of deep neural nets, when you have operations like, broadcast, you have to go out across the internet, so…

186
00:35:44.380 --> 00:36:01.290
gdeptuch: The… so this is a good example of how the optics, in this case passive optics, passive optical switching is doing some computation, some useful computation on the fly, you know, during transmission, because we have these data getting routed arbitrarily through the network, programmatically through the network.

187
00:36:01.360 --> 00:36:15.079
gdeptuch: Yeah, so like I say, you can do this hierarchically, build it up layer by layer. People, they've demonstrated, in simulations and experiments, you know, several layers of hierarchy, so, and this lets you,

188
00:36:15.290 --> 00:36:22.900
gdeptuch: in principle, go to arbitrary network sizes. And this is called CYPAC, the Silicon Photonic Accelerated Computer Architecture.

189
00:36:23.240 --> 00:36:35.120
gdeptuch: So another example which is also related to this is doing more computation while the data is being transferred between, between computational unit and memory. So instead of the memory.

190
00:36:35.240 --> 00:36:57.649
gdeptuch: bus being just a data transfer bus, it can do some processing while the data is being transferred, and this can reduce… this can do several things. For example, you can use this for compression, or you can use this for doing some reduction and some reduction in… and that can give you system-level advantages, such as reduction in latency or reduction in overall energy use.

191
00:36:57.650 --> 00:37:20.229
gdeptuch: So you can imagine now we have these, as shown on the left, you have these photonic compute engines, which are leveraging 3D integration, as Gresh mentioned, which are now co-packaged with the optics, getting very, very high throughput, and processing data through these WDM buses as they're moving. So, one kind of concrete example of this, which they published recently, is doing

192
00:37:20.230 --> 00:37:26.860
gdeptuch: There's a sort of innovative approach to doing multiply-accumulate operations, which are, as everybody knows, one of the core operations.

193
00:37:26.860 --> 00:37:39.750
gdeptuch: For all these, AI engines. So the way this works is that you have, you have a multi-layered stack, again, 3D integration with, several layers of high-bandwidth memory DRAM stacked on top.

194
00:37:39.750 --> 00:38:03.860
gdeptuch: Below that, you have an electrical IC, and below that is a photonic IC. The data, the coefficients that you want to multiply are stored in DRAM. They get transferred to the electrical IC, which calculates partial products, which are basically just multiplies. But then, when you want to do the addition, you transfer that to the photonic IC, which then does the addition. Moreover, the output light then goes back

195
00:38:04.500 --> 00:38:06.659
gdeptuch: Naturally, and then is used.

196
00:38:06.660 --> 00:38:27.429
gdeptuch: To look, to, to control the, the, the level of light, which can then, can tell you which then, through basically what, what is, what amounts to a flash type, to digital conversion, that says how many of the, outputs of the, of these, photodiodes are actually high or low, so that, that gives you the thermometer-coded output.

197
00:38:27.430 --> 00:38:30.589
gdeptuch: Which means you have done this photonic analog-to-digital conversion.

198
00:38:30.590 --> 00:38:40.080
gdeptuch: Quite efficiently in parallel, again, because we are doing this with, with a bunch of, resonators, which are… which are thermally tuned.

199
00:38:40.430 --> 00:38:48.510
gdeptuch: So… so then we have this… okay, now in terms of performance achieved, this is still not at a high resolution. I mean, the… the,

200
00:38:48.730 --> 00:38:58.039
gdeptuch: typical, like, number of bits is on the order of 3 to 4 as of now. You know, work is going on to increase that, but, but I'll say that already.

201
00:38:58.040 --> 00:39:12.009
gdeptuch: for a lot of applications, even 4 bits of precision is good enough, and so this is a compelling way to actually implement these MAC operations and digitize in analog, and then digitize them efficiently in the organic domain.

202
00:39:13.260 --> 00:39:38.180
gdeptuch: Yeah, so this is more of a blow-up of how that might actually be, implemented. Like I said, the high-bandwidth memory modules are on top. You have the IC, electrical IC in the middle, active photonic interposer, and then everything is sitting on a… and a package. Everything is then interposed on a substrate. And yeah, this is more of a blow-up of the algorithm. Like I said, partial products are actually… this is a good example, actually, of something

203
00:39:38.180 --> 00:39:43.800
gdeptuch: that will be talked about in Thrust 3, which is how different operations might be partitioned between domains. You can see here the

204
00:39:43.800 --> 00:40:00.450
gdeptuch: The multiplications are done, which is nonlinear operation are done electrically, but then the linear addition, which can be efficiently done optically, is done optically, so this might be a good example, or made a good point to actually hand over to the… to Thrust 3, which I think is coming up, and we'll talk more about partitioning.

205
00:40:01.070 --> 00:40:11.419
gdeptuch: Yes. Okay, so the Trust Trust 3 will be shared between Prashant and Ocean, right? Yeah, yeah, okay, so I'll provide an overview, and then Ocean, who's from PNNL, will provide more details.

206
00:40:11.420 --> 00:40:23.340
gdeptuch: So, in Trust 3, our objective really is to develop electrophotonic processing architectures and make use of all the different pieces, all the insights gained from Trust 1 and Thrust 2 in doing so.

207
00:40:23.340 --> 00:40:30.489
gdeptuch: And for building these architectures, we'd like to leverage all the benefits of processing in photonic domain.

208
00:40:30.490 --> 00:40:48.890
gdeptuch: And while we're trying to build these mixed electrophotonic processing architectures, our goal really is to also develop frameworks, tools, and methodologies which allow us to systematically develop these architectures following a top design… a top-down design approach as opposed to a bottoms-up approach.

209
00:40:49.040 --> 00:41:02.309
gdeptuch: And we'd like to automate how we do our architecture search. We'd like to make the architecture search hardware aware, taking into account all the constraints coming from the processing platform that we use.

210
00:41:02.320 --> 00:41:16.340
gdeptuch: And the target application for these processing architectures, one is developing optimized processing architectures for processing of data that's produced by detectors, but also something more generic, such as processing generic AI workloads.

211
00:41:16.530 --> 00:41:29.940
gdeptuch: And one of the key principles that we follow in Trust 3 is that we'd like to optimally partition tasks between electrical and photonic domains, so as to best exploit the properties

212
00:41:29.940 --> 00:41:43.240
gdeptuch: of each domain. And by doing that, we would like to model all our interfaces, both digital analog and electrical-optical interfaces, as well as the conversion costs associated with the interfaces.

213
00:41:43.240 --> 00:41:48.100
gdeptuch: And also try and mitigate the conversion costs as we do the architecture search.

214
00:41:48.590 --> 00:41:53.059
gdeptuch: the kind of architectures and applications that we are targeting.

215
00:41:53.060 --> 00:42:16.799
gdeptuch: So, applications ranges… I mean, it's data processing, processing of the data that's produced by detectors at various points in the detector pipeline, ranging from, say, data compression and filtering, which is closer to the sensor, to developing low-latency feedback, as well as generating trigger signals at low latencies.

216
00:42:17.030 --> 00:42:32.099
gdeptuch: And at the other end, it could be offloading tasks like particle track reconstruction, clustering calibrator hits, and doing this in a more energy-efficient manner than a typical electronic-only hardware accelerator would do.

217
00:42:32.270 --> 00:42:56.430
gdeptuch: And then there are several choices for the processing architectures for doing one or multiple of these tasks. So, trust three, we'd like to explore both. First, the suitability of different architectures for doing these different compute operations, and also how photonics can be leveraged, and how photonics can be leveraged specific to each of these different architectures.

218
00:42:56.600 --> 00:43:03.030
gdeptuch: The architectures range from, deep neural networks, where we implement dense MAC operations, striking neural networks.

219
00:43:03.030 --> 00:43:25.139
gdeptuch: doing combinatorial optimization using icing machines, and also graph neural networks, which are of particular interest here. The reason being that the data that's produced by detectors is… it's often sparse, irregular, restructured, there are complex dependencies between the different pixels, different detector subsystems, and such data maps naturally to graphs.

220
00:43:25.140 --> 00:43:32.829
gdeptuch: So we think graph neural networks can do a good job of processing the data that we typically see coming out of these detectors.

221
00:43:32.940 --> 00:43:51.119
gdeptuch: And as I said, one of the key drivers of Thrust 3 is doing optimal electrophotonic design partitioning, and there's more coming on this from our colleagues in PNNL, but the approach here that we follow is that we try to implement optimization loops within our architecture search engine.

222
00:43:51.120 --> 00:44:08.079
gdeptuch: Where we define performance, cost trade-offs, we define key metrics, which should be taken into account. These are some of the partitioning drivers that are shown here. Could be latency, power consumption, bandwidth throughput, how reliable is the architecture, as well as the noise resilience.

223
00:44:08.080 --> 00:44:18.189
gdeptuch: And in the functional simulator that we developed, we do a hardware-avail behavioral modeling, so we try to model our components as close as possible to what we might realistically see.

224
00:44:18.190 --> 00:44:36.199
gdeptuch: Given the, fabrication process, so we'd like to model non-idealities, crosstalk, losses, thermal drift, as well as conversion efficiencies, between different domains, and make use of all these different parameters as we do our design space exploration, as well as parameter optimization.

225
00:44:36.200 --> 00:44:59.879
gdeptuch: As for task division, our initial starting point is allocating linear operations to photonic domain, because that's what photonic is really good at, doing nonlinear operations in electrical domain for the time being, until material advances allow us to do nonlinear operations efficiently in photonic domain. But this is just a starting point, and we hope to get more insights from our partitioning algorithm.

226
00:45:00.230 --> 00:45:12.179
gdeptuch: And also make use of photonic interconnects wherever possible for doing energy-efficient data transmission, mitigating bottlenecks associated with moving data between the memory and the compute units.

227
00:45:12.250 --> 00:45:29.999
gdeptuch: And as we evaluate these different architectures, wherever possible, we are trying to make use of publicly available physics-based datasets, which are all derived from experiment-like conditions. And our goal is to use these datasets for rigorous benchmarking of the processing architectures. One example is the TradCamel dataset that we're using

228
00:45:30.000 --> 00:45:34.109
gdeptuch: For evaluating GNNs for doing particle trap reconstruction.

229
00:45:34.910 --> 00:45:42.350
gdeptuch: And I think that's it for me. I'll pass over to Ocean. Yep, so we'll navigate, right? Is that okay if we navigate?

230
00:45:43.320 --> 00:45:47.819
Oceane Bel (PNNL): Yeah, you guys can navigate. I'll just say next slide when, I need to get to the.

231
00:45:47.820 --> 00:45:48.840
gdeptuch: Next slide.

232
00:45:48.840 --> 00:46:02.839
Oceane Bel (PNNL): So, yeah, hi, I'm Ocean, I will be talking about what we've been working on at PNNL as part of Alpha. So, the partitioning problem that was mentioned earlier, we took it in two different directions. So, one.

233
00:46:02.840 --> 00:46:19.269
Oceane Bel (PNNL): Is essentially what is the architecture we are looking at? Are we going to be looking at using certain components as electronic components and certain components as photonic components? Bear in mind that there are components that can be either photonic or electronic, so you also have to decide for us which

234
00:46:19.270 --> 00:46:39.039
Oceane Bel (PNNL): portals as flexible components. So one of the ideas with that partitioning is that regular electronic computers are hitting a wall, they need a massive amount of power to run faster, whereas new photonic technology could help, but we don't know how to combine them with the electronic components quite yet in an effective manner.

235
00:46:39.470 --> 00:46:53.529
Oceane Bel (PNNL): So, the tool that we are building, currently uses, Monte Carlo to essentially try different, combinations until it converges to us, a point based on the metrics that Prasancha mentioned earlier.

236
00:46:53.700 --> 00:47:06.170
Oceane Bel (PNNL): And another approach that we're looking at is called hypergraph partitioning. This is a approach that our colleague Jason is bringing to the table and has been working on to integrate as part of our codebase.

237
00:47:06.310 --> 00:47:25.689
Oceane Bel (PNNL): Basically, we're creating algorithms that will decide which computing task or which computing node would be what type. So, we would also work on allocating a certain task to a certain node, and whether that node is a photonic node or electronic node. So, it's a multi-dimensional problem that we're trying to break down and work step by step.

238
00:47:26.300 --> 00:47:27.580
Oceane Bel (PNNL): Next slide.

239
00:47:29.850 --> 00:47:47.499
Oceane Bel (PNNL): So, before I go too deep into the weeds of the technical approach, these are some of the libraries that we've been looking at. So, we've had, like, libraries like Meep that are used to simulate how light travels through photonic chips. We've looked at NG Spice, which simulates electronic circuits.

240
00:47:47.500 --> 00:47:58.989
Oceane Bel (PNNL): Pymtl, which models digital logic, KLAL, PyTorch, GamePy. We've also explored some electronic photonic partitioning approach, such as MEDIS,

241
00:47:59.650 --> 00:48:17.700
Oceane Bel (PNNL): I'm not going to try to pronounce the rest of them, it's on the side. But, yeah, we've also looked at those approaches, because we do want to get some knowledge of a baseline partitioning approach, specifically that has been developed for electronic photonic, so that once our tool has been fully developed, we can then compare it to existing state-of-the-art.

242
00:48:17.880 --> 00:48:19.089
Oceane Bel (PNNL): Next slide, please.

243
00:48:21.460 --> 00:48:33.279
Oceane Bel (PNNL): So, right now, we are working on a version of our partitioning tool that uses Monte Carlo Search to find a good combination of electronic component, photonic components, as you can see on the video here.

244
00:48:33.280 --> 00:48:42.950
Oceane Bel (PNNL): We have both a baseline one, where we are just simply running a streamline of data and then seeing how we want to create the different partitioning approaches.

245
00:48:42.950 --> 00:48:46.539
Oceane Bel (PNNL): We do generate graphs of the different partitioning that is generated.

246
00:48:46.540 --> 00:48:55.350
Oceane Bel (PNNL): We also, give the user the ability to change the power or the latency as constraints, as part of our interface.

247
00:48:55.360 --> 00:49:14.290
Oceane Bel (PNNL): We will be working on integrating more constraints later down the line. The code already has the capability, we just need to bring them up to the interface. We also do have a state of the code where we are essentially running not only Monte Carlo approach, where we're testing out the different performance factors of the different nodes.

248
00:49:14.350 --> 00:49:32.769
Oceane Bel (PNNL): But we are also running and training the neural network model based on, essentially, the TensorFlow library, and then we're incorporating back the performance of the TensorFlow library into the resulting performance of the chip, and then that is taken into account for the loss of the model.

249
00:49:32.770 --> 00:49:40.810
Oceane Bel (PNNL): And we're gathering the loss of the model by taking into account both the performance and the TensorFlow output. So that does take a while to finish.

250
00:49:40.870 --> 00:49:49.869
Oceane Bel (PNNL): In the meantime, I'll also talk about another approach that will not be shown in the video quite yet, which is essentially

251
00:49:50.250 --> 00:49:55.099
Oceane Bel (PNNL): The hypergraph approach, which I'll go more into, the next slide over.

252
00:49:55.690 --> 00:50:00.030
Oceane Bel (PNNL): We will be integrating that one into,

253
00:50:00.170 --> 00:50:03.359
Oceane Bel (PNNL): Oh, I think the video's over. Next slide, sorry.

254
00:50:04.930 --> 00:50:06.400
gdeptuch: Excellent. Okay.

255
00:50:07.020 --> 00:50:20.359
Oceane Bel (PNNL): So our second approach treats the problem like a puzzle. So, essentially, we represent the electronic and photonic components as dots, and their connections as lines linking multiple dots together, which create what we call a hypergraph.

256
00:50:20.780 --> 00:50:34.140
Oceane Bel (PNNL): First, we simplify this complex puzzle by grouping related components together. We can then only group, components that belong to the same category or have a strong connection, kind of like, sorting puzzle pieces up by color or pattern.

257
00:50:34.630 --> 00:50:47.680
Oceane Bel (PNNL): Next, we make a first guess as to how to divide everything between electronic and photonic parts, trying to keep things as balanced as possible. Again, we refer back to those metrics that we talked about at the beginning.

258
00:50:47.690 --> 00:50:59.299
Oceane Bel (PNNL): And finally, we improved our solution by testing small changes. For each component. We asked, would moving this piece to the other side make things better? We focused on important improvements.

259
00:50:59.300 --> 00:51:10.919
Oceane Bel (PNNL): First, using priority weight. This code is, currently in a working phase in the background of the code, of our codebase. We still have yet to connect it to our interface.

260
00:51:10.920 --> 00:51:22.039
Oceane Bel (PNNL): But the goal would be that we would have this and the Monte Carlo approach connected to the interface, so that we can do, actually, a comparison between the partitioning approaches on both, and come up with a

261
00:51:22.170 --> 00:51:35.929
Oceane Bel (PNNL): What we refer to as potentially an ensemble approach that, gives us a overwhelming positive partitioning solution that both includes task distribution and the selection of different components for different partitions.

262
00:51:36.720 --> 00:51:37.470
Oceane Bel (PNNL): Thank you.

263
00:51:38.810 --> 00:51:50.090
gdeptuch: Okay, so what about we now could close? I will share a couple, maybe, more slides, because we… to realize the goals of our Alpha project, of course, we…

264
00:51:50.190 --> 00:52:12.360
gdeptuch: We need huge simulation engines, software, models, but also one cannot forget about investment in testing equipment, because having anything fabricated requires testing, and it's jumping from electrical domain to the photonic domain, electrophotonic domain.

265
00:52:12.360 --> 00:52:18.399
gdeptuch: That has to be addressed, and we are working also on equipping our…

266
00:52:18.400 --> 00:52:31.350
gdeptuch: laboratory spaces, so I think, certainly, this is… this slide is coming from the Columbia University, that, it's an electrophoton probe station for… allowing for dire wafer probing.

267
00:52:31.350 --> 00:52:55.600
gdeptuch: investment of infrastructure for testing. It's also happening in our… at our colleague… colleague's facility at CUNY CCNY, and our colleagues at Florida University are investing in finite element methods simulations… simulation tools, and at Brookhaven, we are also investing in a new laboratory space that is,

268
00:52:55.670 --> 00:53:06.209
gdeptuch: We call it a new test lab facility, located at Instrumentation department, where we already have 12-inch semi-automatic probe stations. Now we're adding

269
00:53:06.210 --> 00:53:28.180
gdeptuch: photonic package on top of it, where we'll be able to test electrical and photonic chips on the wafers, on single dice, with edge coupling of single fibers, ribbon, ribbon fibers.

270
00:53:28.180 --> 00:53:51.100
gdeptuch: So all that is needed later on to explore, especially, particularly at, aimed at 3D integrated stacks, that will be needed to access such stacks with very high precision. So this investment is ongoing. We're also looking into development of in-house capabilities such as photonic wire bonding.

271
00:53:51.100 --> 00:53:55.120
gdeptuch: To have, pop-up, through,

272
00:53:55.120 --> 00:54:14.230
gdeptuch: using polymeric wires to be able to interconnect photonic photonic ICs. So these investments are happening in parallel. They are part of the alpha project, but also generally enhancement of capabilities and instrumentation that Department and the Brookhead National Laboratory.

273
00:54:14.230 --> 00:54:38.959
gdeptuch: But this brings me also to… this brings me into summary, and in summary, I would like to say that ELFO is a very intense project, covers… spans a full chain from concept to… through design, fabrication, finally, to testing. We cover manufacturing of electrophotonic circuit prototypes, we pursue heterogeneous integration, we require advanced test and characterization infrastructure.

274
00:54:38.960 --> 00:54:39.800
gdeptuch: structure.

275
00:54:39.800 --> 00:55:03.749
gdeptuch: We drive development of new CAD EDA tools and methodologies for the designing, and when I was talking about special investment in the testing architecture, or testing infrastructure, certainly it's great to be part of a Meerkat Center, because certainly this is why we are having all these presentations.

276
00:55:03.750 --> 00:55:12.000
gdeptuch: happening when, in front of Meerkat Center members. We seek for… we seek for synergies.

277
00:55:12.000 --> 00:55:36.720
gdeptuch: within Meerkat and also outside, so we enlarge on external foundry facilities. We mix commercial and open source design flows, and we have huge exposure, you know, overlap already we can notice with Deco D project, that is another project within Meerkat. I would like to also stress that was a great experience for me, particularly

278
00:55:36.720 --> 00:55:54.909
gdeptuch: attending, you know, the SINT annual meeting that was… that took place in Santa Fe a bit more than one month ago, when there was a lot of exposure on what is going on within the centers, and also we've invited guests, and thanks to Jeff. So I think the synergies

279
00:55:55.090 --> 00:56:10.220
gdeptuch: and reusing maybe some investments among Meerkat Center members could be a great addition to our L4 project, and as I mentioned, we have very intense plan in front of us, and hope we'll be able to realize it.

280
00:56:10.220 --> 00:56:29.879
gdeptuch: And, hopefully we'll be… we'll be able to grow, next steps out of this, what I'm… what I'm seeing as a planted seed, kind of incubate… incubate it to broader, broader efforts. So I think we managed to be on time, so this is the last slide, and thank you for your attention.

281
00:56:30.980 --> 00:56:33.689
gdeptuch: Or if there are questions, we try to answer.

282
00:56:34.620 --> 00:56:36.400
gdeptuch: Or we take notes.

283
00:56:41.960 --> 00:56:43.930
gdeptuch: Yeah, we lost people, I guess.

284
00:56:46.630 --> 00:56:58.420
Paul McIntyre: Hey there. Yeah, thanks so much, Gruz Gors, that was terrific. Thank you to your team as well. Really a lot of cool stuff in this, project.

285
00:56:58.630 --> 00:57:00.720
Paul McIntyre: Any, any questions?

286
00:57:04.810 --> 00:57:14.449
Paul McIntyre: I guess maybe I'll kick it off with a question. Can you give us an idea of how this effort kind of compares with

287
00:57:14.450 --> 00:57:25.189
Paul McIntyre: others that might be going on in industry that you're aware of, or in other consortia. Clearly, you're addressing some really

288
00:57:25.240 --> 00:57:30.960
Paul McIntyre: Critical, topics that are relevant to many different kinds of

289
00:57:31.090 --> 00:57:46.450
Paul McIntyre: of applications in computing more broadly. So is there a… is there a fairly well-developed set of projects out there that are doing other things, or is this pretty unique?

290
00:57:47.150 --> 00:57:52.710
gdeptuch: Oh, I would maybe partially answer the question, and I later ask my colleagues to add, because

291
00:57:52.710 --> 00:58:15.430
gdeptuch: that everyone may have different… slightly different opinions. I would say that when we started… this is why I brought the motivation slides at the beginning, because a few years ago, when we started talking about reaching for photonics, we were told, or we actually also were thinking that, oh, photonics is… photonics is great for beta transmission, because you cannot stop photons, you cannot do computation. Slowly.

292
00:58:15.430 --> 00:58:40.099
gdeptuch: we start seeing, you know, many efforts emerging or, you know, advancing, especially maybe more at universities, at the first place, to use photonics for computation. You may, of course, as Prashansa mentioned, consider, you know, operating purely on amplitude. You can also work with phase, with interferometric, you know, approaches.

293
00:58:40.180 --> 00:58:50.549
gdeptuch: Of course, you don't stop photons, but you can get these linear operations, multiplication and addition very efficiently and very fast.

294
00:58:50.550 --> 00:58:54.340
gdeptuch: And what is actually exceeding speed of these computations?

295
00:58:54.340 --> 00:59:18.810
gdeptuch: even up to 2 orders of magnitude compared to electrical domain, is wavelength division multiplexing. You cannot… you cannot imagine on a single wire to have, you know, 100 frequencies multiplexed, and then demultiplex them, because, you know, resonance circuit will be gigantic, and will be bad. But in photonics, yeah, it starts to be possible. So you can get a factor of acceleration 10X, 100x.

296
00:59:18.930 --> 00:59:24.259
gdeptuch: in computation. Whether it is something unique or not.

297
00:59:24.410 --> 00:59:43.409
gdeptuch: probably it is not… it cannot be said it is totally unique, because, for example, efforts that Columbia colleagues are doing is oriented on the data center, and they are operating on funding from companies that are either, you know, investing towards the future.

298
00:59:43.410 --> 00:59:50.649
gdeptuch: Or, you know, aiming at new technology development. And we started seeing that in photonic computing.

299
00:59:50.650 --> 01:00:03.369
gdeptuch: is really emerging, and the platforms, particularly silicon photonics, started to be available, where you had all elements needed for linear operations. You start seeing, also, the platforms that are

300
01:00:03.370 --> 01:00:26.719
gdeptuch: non-silicon-based, where you can mix nonlinear operations. I will not be going into these details, because probably it will take a bit more time than needed. You have also heterogeneous embedment of, for example, lasers in the silicon photonics, so these things are trends, I think, in the industry, but we try to have our L4 project

301
01:00:26.720 --> 01:00:31.139
gdeptuch: aimed at what DOE could be interested in.

302
01:00:31.160 --> 01:00:37.130
gdeptuch: And of course, if something happens that we are able to sell this, our findings.

303
01:00:37.150 --> 01:00:59.089
gdeptuch: to industry, it would be… it would be a great, greatest achievement. I don't know, you want to add anything? The only other thing is about… so part of TELFO is also focused on developing tools and methodologies, and for photonics specifically, commercially available EDA tools are not as mature as electronics, so we are hoping that some of the R&D that we do can eventually become part of commercial tools.

304
01:00:59.090 --> 01:01:00.299
gdeptuch: Is it good? Yep.

305
01:01:00.630 --> 01:01:01.220
Paul McIntyre: That makes a lot of sense.

306
01:01:01.220 --> 01:01:06.400
gdeptuch: or what Prashant was covering in her part, you know, partitioning.

307
01:01:06.400 --> 01:01:29.320
gdeptuch: you know, what to cover in photonic domain and what to cover in electrical domain. We are at the very early stage of that, because we are still discussing how to define weight function, you know, to… for the automated optimization. But when we get it somehow, I mean, this certainly can trigger, I hope, huge interest from industry to get… to embed it in regular tools. And of course, now we are focusing on

308
01:01:29.320 --> 01:01:43.100
gdeptuch: partitioning between photonic and electrical domain using known components that exist in the silicon photonic platform, but we can expand it. And, you know, the most important thing is that we skip handmade

309
01:01:43.100 --> 01:02:03.970
gdeptuch: photonic layouts, photonic implementations, because now when you go into Cadence and you start using, you know, even the very sophisticated tool, like Kirby Core to avoid drawing waveguides, you have to do it by hand. But we were able to automate the process from top to bottom, from the concept of GNN, like Prashansa was describing.

310
01:02:04.100 --> 01:02:23.949
gdeptuch: through partition in between domains, and then, like, synthesis in the electrophotonic domain. I think this can be breakthrough. Whether we'll achieve it, we'll see. Maybe in collaborations with others, this can be a huge addition to automate, effectively, this type of efforts. I don't know…

311
01:02:24.560 --> 01:02:26.470
gdeptuch: Anyone wants to add?

312
01:02:27.130 --> 01:02:33.280
gdeptuch: We can ask our advisor, industrial advisor, if you have any… Bob, do you have anything to add?

313
01:02:33.280 --> 01:02:40.139
'Pops' / NY CREATES: No, no, no, I don't think I have much to add. I know you're working closely with Infotonics and

314
01:02:40.460 --> 01:02:42.609
'Pops' / NY CREATES: Nope, nothing much to add.

315
01:02:45.050 --> 01:02:46.540
Paul McIntyre: Okay, any other questions?

316
01:02:53.290 --> 01:03:11.049
Paul McIntyre: Well, if not, we're pretty much at time, so thanks so much to the Alpho team for a great presentation. There was a lot there, and we're going to closely, review the recording of this. I want to share this with some of our colleagues in my team, because I think they'll really,

317
01:03:11.050 --> 01:03:17.700
Paul McIntyre: benefit from seeing it, and there's some… I see some opportunities for collaborations that could emerge

318
01:03:17.700 --> 01:03:27.259
Paul McIntyre: And so it's been great to see this, and really exciting to see what your team does during this, this Meerkat Center project.

319
01:03:28.370 --> 01:03:35.520
Valerie Taylor: Yes, indeed. Yes, we have some, someone that's doing work in,

320
01:03:36.070 --> 01:03:41.790
Valerie Taylor: optical interconnects, so I definitely want to share, the recording.

321
01:03:42.040 --> 01:03:43.580
Valerie Taylor: There, as well.

322
01:03:44.550 --> 01:03:54.830
gdeptuch: So, I think we're happy to, you know, go maybe in more details, if someone is interested in kind of going in the details, because we actually do high-speed jet… jet…

323
01:03:54.910 --> 01:04:00.719
gdeptuch: fly over what we plan, but we wanted to cover the whole…

324
01:04:00.720 --> 01:04:17.029
gdeptuch: I, you know, set part of ideas that we are having, going, as I said earlier, from materials, modification, but needed, to all the way to high-level abstraction randomization of models for computing.

325
01:04:17.950 --> 01:04:22.050
jsnelso: Horace, I'm sure you're still talking with Rakdam here, just making.

326
01:04:22.050 --> 01:04:25.920
gdeptuch: Yes, yes, we are planning to have what I call a workshop.

327
01:04:25.920 --> 01:04:49.419
gdeptuch: invite him. He shared with us school staff what he's doing, and effectively, we already found… I mean, at least I found synergies, because tackle… tackling all the problems of on-the-flight processing in photonic domain, he was showing certain things for images that he's interested, you know, some objects, identification, and this is exactly what,

328
01:04:49.420 --> 01:04:57.350
gdeptuch: It's, common between, you know, his work and, for example, what we have in particle physics, even, that we have

329
01:04:57.810 --> 01:05:15.380
gdeptuch: something is being… something's changing quickly, and you would like to catch that. I think what he was… particularly what he was showing to us was great achievements in compression of sparse data sets. This is… this is something that we would like to look closer, working with, you know, with him.

330
01:05:15.570 --> 01:05:16.939
jsnelso: Okay, great. Thanks.

331
01:05:16.940 --> 01:05:17.510
Paul McIntyre: Cool.

332
01:05:18.130 --> 01:05:36.820
Paul McIntyre: Thanks so much. I've got to jump onto another call, but this was terrific, and we'll have a PI-only discussion next week. We can talk about those two agenda items, the, you know, initial planning for an annual meeting.

333
01:05:36.820 --> 01:05:42.679
Paul McIntyre: And also, let me see… oh, the, the, the, our,

334
01:05:42.780 --> 01:05:49.789
Paul McIntyre: our subcommittees, or our sub-teams, and how we want to populate those. So we'll, we'll,

335
01:05:49.790 --> 01:06:04.859
Paul McIntyre: We'll pick that up, next time, and be on the lookout for a new… I think we'll… you'll see… you should see these, the current calendar invitations, deleted and replaced by, new ones, starting, for next Wednesday.

336
01:06:05.660 --> 01:06:08.999
gdeptuch: Very good. I'm glad that we didn't go over time too much.

337
01:06:09.000 --> 01:06:11.549
Paul McIntyre: Okay, thank you.

338
01:06:11.550 --> 01:06:12.120
'Pops' / NY CREATES: Thank you again.

339
01:06:13.060 --> 01:06:13.630
Valerie Taylor: Thank you!