New Photonic Chip Is the Full Package | Hacker News Reader

New Photonic Chip Is the Full Package

spectrum.ieee.org

61 pointsmfiguiere15 comments

This suggests doing things like connecting RAM through an (optical) cable. There will still be latency, but prefetching into cache can make it worthwhile.

Maybe macs with upgradable memory will appear in the future :-)

The potential application that I'm most excited about is better high bandwidth networking. In data centers e.g. for training on multiple GPUs this could yield a magnitude improvement in shuffling data around. If I read it correctly the promise is that we get high Terrabit or even Petabit per Second bandwidth (also with lower latency, although here I did not find numbers)

https://www.nature.com/articles/s41566-022-01082-z

The problem would be the PCIe itself in that case. According to Wikipedia, PCIe 7.0 x16 is limited to 242GB/sec [0], which makes ~2000Gbit/sec. Current version is around 121GB/sec, which makes 1000Gbit/sec, where an NDR Infiniband 12X connection can already saturate [1].

We currently use 800Gbit links divided to four nodes (200x4) at the switch level.

The first application will probably be core switches with higher total throughput with lower P2P latency, and with massive number of ports without much speed/latency penalty.

[0]: https://en.wikipedia.org/wiki/PCI_Express#Comparison_table

[1]: https://en.wikipedia.org/wiki/InfiniBand#Performance

> PCIe 7.0 x16 is limited to 242GB/sec

Meanwhile the DDR4 memory in my computer is limited[1] to just 41GB/s...

[1]: https://www.tomshardware.com/reviews/gskill-trident-z-royal-...

That's per channel. An AMD Epyc processor have 8 channels per socket, topping at 246GB/sec per socket.

True, and IMHO just highlights my point, which is that PCIe 7.0 has quite insane bandwidth. After all the memory is sitting right next to the CPU, with tons of IO pins to do its work. Meanwhile the PICe device is far away and using only 16 differential pairs.

Yeah, but your cache throughput is already faster than PCIe 7.0 (and way lower latency), and cache replacement algorithms for modern CPUs are really, really good.

>Current version is around 121GB/sec

XD, I wouldn't say PCI-E 6.0 as current version since we are not even anywhere close to releasing actual product yet. PCI-E 5.0 x 16 is at 63GB/sec.

We need a new internet law. All comment threads will now turn into AI/GPU conversations no matter what.

A GPU is also useful in other scientific calculations, AI is just a small part of it.

Im aware. and sure high bandwidth applications related to GPU is fine application- but the researchers specifically call out spectrum over saturation and "security" ... so to me these comments are just trying to inject their relevancy...

We’ll call it “spacecadet’s law”: every online discussion will eventually end up talking about the bubble-du-jour.

Thats self fulfilling...

Well, photonics are mostly useful for communication and SIMD computation, so this one is inevitable.

But how will they help the blockchain?

Reply on news.ycombinator.com