←back to thread

Meta's open AI hardware vision

(engineering.fb.com)
240 points GavCo | 6 comments | 15 Oct 24 18:08 UTC | HN request time: 0.808s | source | bottom
1. seydor ◴[15 Oct 24 18:42 UTC] No.41851732[source]▶
They ve already gone after openAi, are they after Nvidia now?
replies(2): >>41851801 #>>41852876 #
2. throwup238 ◴[15 Oct 24 18:47 UTC] No.41851801[source]▶
No, this is a rack built on NVIDIA’s platform, so this is just more $$$ for them.
replies(3): >>41852155 #>>41852222 #>>41853060 #
3. moffkalast ◴[15 Oct 24 19:22 UTC] No.41852155[source]▶
Yeah this is... nothing. At least nothing anyone worth less than a few billion could ever care about.

Would be far more interesting to see MTIA in an edge compute PCIe form.

4. throwaway48476 ◴[15 Oct 24 19:28 UTC] No.41852222[source]▶
There's also an AMD rack and meta is big enough that they won't get blacklisted for it.
5. KaoruAoiShiho ◴[15 Oct 24 20:41 UTC] No.41852876[source]▶
Sort of, while yes this uses Nvidia, one of Nvidia's moats or big advantages is its rack scale integration. AMD and other providers just can't scale up easily, they are behind in terms of connecting tons of GPUs together effectively. So doing this part themselves instead of buying Nvidia's much-hyped (deservedly) NVL72 solution, which is a nonpareil rack system with 72 GPUs in it, and then opensourcing it, opens the door to possibly integrating AMD GPUs in the future and this hurts Nvidia's moat.
6. mhandley ◴[15 Oct 24 21:00 UTC] No.41853060[source]▶
The article talks about NVidia racks, but also about their DSF [0], which is Ethernet-based (as opposed to Infiniband) built on switches using Cisco and Broadcom chipsets and custom ASIC xPU accelerators such as their own MTIA [1] which is built for them by Broadcom. So they are talking more than one approach simultaneously.

[0] https://engineering.fb.com/2024/10/15/data-infrastructure/op...

[1] https://ai.meta.com/blog/next-generation-meta-training-infer...