3 ms·
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
by zurfer 3y ago
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 https://www.nature.com/articles/s41566-022-01082-z
- bayindirh 3y agoThe 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 https://en.wikipedia.org/wiki/PCI_Express#Comparison_table [1]: https://en.wikipedia.org/wiki/InfiniBand#Performance https://en.wikipedia.org/wiki/InfiniBand#Performance
- magicalhippo 3y ago> 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-ddr4-4000-c17-2x16gb-review https://www.tomshardware.com/reviews/gskill-trident-z-royal-...
- bayindirh 3y agoThat's per channel. An AMD Epyc processor have 8 channels per socket, topping at 246GB/sec per socket.
- magicalhippo 3y agoTrue, 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.
- mplewis9z 3y agoYeah, 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.
- ksec 3y ago>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.