Facebook Still Pursuing "NetGPU" - Working On AMD GPU Support In Addition To NVIDIA

[phoronix]

Phoronix: Facebook Still Pursuing "NetGPU" - Working On AMD GPU Support In Addition To NVIDIA

It was the recent Facebook patches for implementing NetGPU that with one of the NVIDIA-focused patches led to the recent controversy around "GPL condoms" in the kernel and ultimately leading to new protections with Linux 5.9. That NetGPU code is still being worked on by Facebook with upstream hopes but now in addition to the NVIDIA driver support they are also working on AMD GPU support with the open-source driver...

http://www.phoronix.com/scan.php?page=news_item&px=Facebook-NetGPU-LPC2020

[M@GOid]

This is a OFF-Topic but since the article mention AMD GPU and internet, hence Firefox ( ¯\_(ツ)_/¯ ), I like to take a opportunity and tell people that had old GCN Radeon GPUs and are using experimental AMDGPU support (with the "amdgpu.si_support=1 amdgpu.cik_support=1" kernel parameters on Grub), that this can cause instability on the recent video hardware acceleration on Firefox 80. Reverting back to Radeon kernel driver improves stability. Is not perfect yet (if it was it would be enabled by default), but is a step on the right direction.

[bug77]

Wait what? Writing software that works both with Nvidia and AMD drivers? Surely that's not possible

[ALRBP]

Having the AMD GPU support working off an open-source compute stack will also clear an obstacle for NetGPU getting review and approval from other upstream kernel developers rather than being contingent upon the NVIDIA proprietary driver.

Were they seriously hoping to mainline something that is useless without a proprietary driver ?

[Aeder]

This is why it pays off to be stubborn about the GPL with big companies. Since they really want this functionality, they will do what it takes, which means the open AMD driver is a bit closer now to parity as a ML platform, instead of winding up with NVIDIA-only capabilities in kernel.

[Iksf]

Not to be confused with WebGPU. Why are we so shit with names

[Spacefish]

200G Network interconnect. Imagine those switches for such a cluster o.O
Depending on neural net architecture, they might need N to N, so the switch is probably "fully meshed" 200G

Wonder why they did not go the PCIe only route, i mean you could theoretically interconnect those nodes via more PCIe switches, instead of converting to 200GbE and back.. Would benefit the latency as well probably.

[curfew]

This is why it pays off to be stubborn about the GPL with big companies. Since they really want this functionality, they will do what it takes, which means the open AMD driver is a bit closer now to parity as a ML platform, instead of winding up with NVIDIA-only capabilities in kernel.

Personally I see this just as a social hack. They provide a patch that's so so for the AMD hardware just to get their way with the proprietary Nvidia stuff and have it injected into the kernel. Once the first set of patches is approved, they will abandon all support for AMD and instead focus solely on patching whatever bugs are discovered with Nvidia.

Kernel devs should maybe only accept the AMD portion of the code and tell the FB engineers to switch to AMD hardware, hehe.

[cusa123]

Is Facebook doing a "google stadia for VR"?

[microcode]

It is refreshing to see somebody respond with productive development instead of turning up the corporate nose to the upstream, when their patches are not welcomed.