Re: ROCM host architectures
Hi Cory,
On 2023-02-22 08:35, Cordell Bloor wrote:
> I'm very curious if the arm64 build actually works. Does anybody have
> hardware to test it?
that's a great question, actually. Didn't see the forest for the trees,
but practically all arm64 stuff I see is either SBCs, Chromebooks, or
Amazon's Graviton, neither of which are helpful.
Consumer-purchasable options do/did exist, e.g. the Gigabyte MP30-AR0
(which is older, and sold out as it looks), or SolidRun's HoneyComb LX2
(though physical 8x slot).
Browsing our infra machine database [1], there are some hosts with
AppliedMicro and some even with AMD "Seattle" processors (didn't even
know those exist), but those are provided/hosted by companies and I'm
not sure how easy we can get access. We can always ask.
[1] https://db.debian.org/machines.cgi
>> This is hardly surprising, as some of these architectures (most notably
>> armel, armhf, mipsel, i386) are probably not at all what ROCm was
>> intended for. For example, it seems that PCIe 3.0 with atomics are
>> required for anything GPU-related.
>
> I suspect that trying to enable 32-bit architectures would be a very
> painful process given that upstream is not considering them at all.
In the above context -- with the exception of i386, I doubt that systems
even exist that could support a PCIe GPU. Which is just another argument
for dropping those architectures, as building would clearly be pointless.
> Though, the requirement for PCIe 3.0 with atomics is perhaps a bit
> overstated. In September 2021, AMD driver developer Alex Deucher [1]
> clarified,
>
> PCI atomics are not required by the firmware on vega parts. They are
> required by the firmware on navi parts, but we are in the process of
> fixing that. Beyond the firmware, there are no requirements for PCI
> atomics in the greater ROCm stack in general, although they are
> required for certain features (e.g., atomic shader instructions
> writing to system memory).
Oh, good to know, thanks.
> I'm afraid the docs have probably steered you wrong. Alas, there are a
> lot of things in ROCm where the only real way to know if something works
> is to try it and see.
Yeah, this is where upstream's extensive tests will prove to be invaluable.
Best,
Christian
Reply to: