AMD Cloud Platform Makes It Easy To Try Out AMD's Latest CPUs, Accelerators & ROCm Software

[phoronix]

Phoronix: AMD Cloud Platform Makes It Easy To Try Out AMD's Latest CPUs, Accelerators & ROCm Software

Last week at Intel's Innovation conference the Intel Developer Cloud "DevCloud" was announced, while on the AMD side there is already something similar: the AMD Cloud Platform. At the tail end of 2021, AMD announced the Accelerator Cloud as a way for trying out the latest EPYC CPUs and Instinct accelerators complete with a pre-configured ROCm compute software stack. The AMD Cloud Platform is a currently parallel effort to the Accelerator Cloud with the former intended more for developers while the latter is more customer-oriented. After trying out the AMD Cloud Platform, it's indeed an easy way to evaluate the latest AMD data center wares while having a easy-to-deploy, pre-configured software environment.

https://www.phoronix.com/review/amd-cloud-platform

[skeevy420]

Too much bureaucracy is involved in creating an account on AMD Cloud Platform. It is absurd that a person first needs AMD's permission in order to be allowed to pay AMD for any kind of use of ACP - there is no such loop on other cloud platforms (AWS, DigitalOcean, Scaleway, etc). Does a customer first need to acquire the seller's permission when buying a chocolate in a supermarket?

If you read the fine print you'll see it is specifically down to the the 3a Restriction:

use AMD Accelerator Cloud for any purpose other than your own evaluation of the AMD software and hardware made available to you by AMD;

The fine print converted to Layman is basically, "This is in a beta/trial phase so only use it to see if it'll work for your needs, nothing production yet, as well as don't do any shady shit like use it for mining coins or spreading viruses."

I'd wait until the service is fully up and running, when it leaves the current evaluation phase, to see if their policies change and if they're worth criticizing.

[Michael]

Too much bureaucracy is involved in creating an account on AMD Cloud Platform. It is absurd that a person first needs AMD's permission in order to be allowed to pay AMD for any kind of use of ACP - there is no such loop on other cloud platforms (AWS, DigitalOcean, Scaleway, etc). Does a customer first need to acquire the seller's permission when buying a chocolate in a supermarket?

AMD isn't trying to be a public loud provider. Mostly for potentially interested customers wanting to gauge the hardware performance / software support with ease. Or for developers wanting to test their own workloads / models / etc on AMD hardware they might not otherwise have access to. So obviously they don't want a bunch of small random people to sign up.

[skeevy420]

AMD isn't trying to be a public loud provider. Mostly for potentially interested customers wanting to gauge the hardware performance / software support with ease. Or for developers wanting to test their own workloads / models / etc on AMD hardware they might not otherwise have access to. So obviously they don't want a bunch of small random people to sign up.

Exactly. This is for those wanting to run sample data on AMD hardware and services they might buy and for people like you who will run a barrage of tests and benchmarks just to see the results. All of that and how it runs in a multi-user environment is information that AMD is interested in obtaining before their cloud goes live.

[Vistaus]

Too much bureaucracy is involved in creating an account on AMD Cloud Platform. It is absurd that a person first needs AMD's permission in order to be allowed to pay AMD for any kind of use of ACP - there is no such loop on other cloud platforms (AWS, DigitalOcean, Scaleway, etc). Does a customer first need to acquire the seller's permission when buying a chocolate in a supermarket?

It's a trial, so basically a beta. If there's someone handing out free samples in the supermarket and they don't notice you, are you just going to quickly grab one from their hands or do you ask politely if you can have a sample too?

[reza]

How can I obtain an invitation code? On ACP form, it says invitation code is required.

[sobrus]

I'd rather like to test their accelerator on my own machine with stable and mature GPGPU drivers, not the mess that we have currently. Nvidia users just install cuda, while AMD users try to compile Rocm from sources, it takes around 14GB of packages just to get pytorch-rocm installed, and ... still it's not working. Polaris users are happy when they manage find working OpenCL driver. What a joke after all these years, especially since AMD cards are usually well suited for these tasks.

[Jabberwocky]

I'd rather like to test their accelerator on my own machine with stable and mature GPGPU drivers, not the mess that we have currently. Nvidia users just install cuda, while AMD users try to compile Rocm from sources, it takes around 14GB of packages just to get pytorch-rocm installed, and ... still it's not working. Polaris users are happy when they manage find working OpenCL driver. What a joke after all these years, especially since AMD cards are usually well suited for these tasks.

+1 Official support has been an absolute joke even just on Linux alone, nevermind crossplatform solutions.

In retrospect it's understandable given AMD's approach to target CUDA. IMO AMD should have just said we are never going to support these cards because of X/Y. Instead we got it's working for me on Z cards and some of us were expecting support... it's over 5 years later and meanwhile we have https://github.com/ROCm-Developer-To...agent/issues/1 and https://gitlab.freedesktop.org/tomst...umr/-/issues/2 ... I know the unsupported cards do not support 2:1 FP16 but some can do over 16 TFOPS in FP32. I'm very glad that I did not buy a Radeon Pro Duo / AMD Radeon Pro Vega II Duo / Radeon VII. Those were capable cards, just not capable of anything most people want to compute today! xD Heck even Navi 1 took years to get support in rocm. People who bought those cards got burned big time.

Looking forward thing's are still a little bit demoralising compared to AMD's competitors https://github.com/RadeonOpenCompute/ROCm/issues/1776 ... maybe Intel can do better than AMD for consumer products? LoL, we'll see.

[qarium]

+1 Official support has been an absolute joke even just on Linux alone, nevermind crossplatform solutions.

In retrospect it's understandable given AMD's approach to target CUDA. IMO AMD should have just said we are never going to support these cards because of X/Y. Instead we got it's working for me on Z cards and some of us were expecting support... it's over 5 years later and meanwhile we have https://github.com/ROCm-Developer-To...agent/issues/1 and https://gitlab.freedesktop.org/tomst...umr/-/issues/2 ... I know the unsupported cards do not support 2:1 FP16 but some can do over 16 TFOPS in FP32. I'm very glad that I did not buy a Radeon Pro Duo / AMD Radeon Pro Vega II Duo / Radeon VII. Those were capable cards, just not capable of anything most people want to compute today! xD Heck even Navi 1 took years to get support in rocm. People who bought those cards got burned big time.

Looking forward thing's are still a little bit demoralising compared to AMD's competitors https://github.com/RadeonOpenCompute/ROCm/issues/1776 ... maybe Intel can do better than AMD for consumer products? LoL, we'll see.

to target cuda on the source code level is the hard way... to target it on binary level would be easy but maybe agaist the law.

the support of ROCm/HIp on consumer cards is inconsidtent
Polaris does not work
vega64/radeon VII should work
RDNA1 does not work
RDNA2 does work...

i have blender 3.3 installed on fedora 37 with a vega64 and blender claims no supported card

it claims it need vega or rdna gpu with mesa 22.10 .. looks like fedora 37 is not on mesa 22.10 right now.