AMDKFD/ROCm GPU Compute Can Work On POWER Systems Like Raptor's Talos II

[phoronix]

Phoronix: AMDKFD/ROCm GPU Compute Can Work On POWER Systems Like Raptor's Talos II

While NVIDIA graphics in IBM POWER systems have been known to make a powerful combination for supercomputer deployments, for those wanting a libre GPU compute experience can also use POWER with AMD Radeon's open-source driver with a pending patch to the kernel driver...

http://www.phoronix.com/scan.php?page=news_item&px=AMDKFD-Compute-POWERPC

[coder]

Sadly, this is about a year late. That's when AMD started releasing PCIe 4-capable MI60 GPUs and POWER 9 was the only PCIe 4 game in town. Now, Zen 2 CPUs have PCIe 4 and surpass POWER 9 on pretty much all fronts.

Worse, for POWER, the next gen won't even be on 7 nm. I guess they had Global Foundries obligations to fulfill.

https://www.anandtech.com/show/14751/hot-chips-31-live-blogs-ibms-next-generation-power

Well, at least this is good for people who need a fully-libre stack or have other reasons for using POWER.

[madscientist159]

Well, at least this is good for people who need a fully-libre stack or have other reasons for using POWER.

That's actually a growing requirement, especially outside the United States. The required AMD PSP (basically equivalent to Intel ME) makes the Zen CPUs pretty much useless unless your organization can accept the data tampering / theft risks. Even the AMD card-side firmware is a sticking point, but right now it's better than the NVIDIA situation so there is a definite use case for AMD HSA on POWER.

POWER is also an open ISA. x86 is not and will never be. That limits x86's usefulness overall.

Also, my understanding is P10 is 7nm. Don't confuse the P9 14nm refresh with P10.

[bridgman]

Sadly, this is about a year late. That's when AMD started releasing PCIe 4-capable MI60 GPUs and POWER 9 was the only PCIe 4 game in town.

In our defence, the support has been there for a while - this patch is just changing kernel config logic to enable it by default.

[madscientist159]

In our defence, the support has been there for a while - this patch is just changing kernel config logic to enable it by default.

With 5.4 POWER can handle the not-quite-64-bit-bypass logic required for the Vega (and presumably Navi) GPUs, which is why the config change makes sense at this point. Userspace is still a mess though; there's a bunch of pull requests to fix POWER builds that have not been acted on in quite a while (well over a year). Any chance we could get some action on those?

[coder]

Also, my understanding is P10 is 7nm. Don't confuse the P9 14nm refresh with P10.

If you follow my link, the slide they presented @ Hot Chips (August 2019) shows P10 in 2021 and simply states "New Process Technology". For their sake, I hope it's at least 7 nm EUV and that there's no schedule slip.

Still, that leaves a gap of probably 1.5 - 2.5 years where they're not really going to offer any better performance or efficiency.

[coder]

In our defence, the support has been there for a while - this patch is just changing kernel config logic to enable it by default.

I didn't mean to point a finger of blame - just stating that the timing was sub-optimal. Still, you're entitled to state your case.

[R41N3R]

A 100% closed/proprietary Nvidia GPU is the last thing people would stick into an otherwise 100% open system :-) At least I wouldn't invest my money into a Power system just to compromise security later on by such an action.