Tiny Corp Puts Their AMD-Powered Compute Boxes "On Hold"

[Termy]

Considering nVidea because AMD is not open enough must be one of the most ironic things this year...

[pharmasolin]

Considering nVidea because AMD is not open enough must be one of the most ironic things this year...

Not because it's not open enough, but because it's buggy and AMD doesn't address those issues, and don't open firmware so Tiny Corp can trace the problem and suggest changes (or add their workarounds)

[sobrus]

I disagree, GCN and Vega were both GPGPU architectures. AMD tried to copy Nvidia’s success with Fermi which was the first ever complete compute architecture. With the release of GCN 1.0 but lacked a proper compute stack which they resolved with Vega/ROCm. CDNA is still based on Vega. While RDNA was a gaming only part. Which has also added inference accelerator with the addition of AI accelerators.

Edit: Review of the 7970: https://www.anandtech.com/show/5261/...7970-review/25

There is nothing wrong with Radeon hardware, and up to CDNA the architecture was shared between consumer and data center parts (just like in nvIdia, until Hopper arrived recently).
The problem is, that there is no software stack. I had RX570 and I had to install OpenCL driver from AUR, and only one older "Orca" version actually worked. Back then this card was still in production, popular and well supported, not to mention RX590 that was just been announced and used similar chip.

Radeon VII is actually nice for GPGPU, but support for it is being dropped from ROCm 6. Please note that it is high-end card from 2019, It came out just one year before RDNA2. And it's already being dropped when AMD says it brings support to more hardware.

Even the best hardware is worth nothing if you can't use it. And with AMD, GPGPU always feels like pure coincidence, a byproduct, and their plan to fix it is to have no plan at all. I seriously don't know why Tinycorp even considered Radeon for their product.

[ET3D]

where were they taking orders on a project they canceled?

Preorders. It's still on their site. The full specs were on Twitter. Perhaps they will change the orders to Intel GPUs now, although the specs will change, and price will be lower. And that will last until they decide that Intel isn't good enough either.

I personally don't like the way they're communicating and their inconsistency. It feels very high risk. I contrast this with Bernhard Guentner from GPTshop.ai, who I find quite trustworthy.

[WannaBeOCer]

There is nothing wrong with Radeon hardware, and up to CDNA the architecture was shared between consumer and data center parts (just like in nvIdia, until Hopper arrived recently).
The problem is, that there is no software stack. I had RX570 and I had to install OpenCL driver from AUR, and only one older "Orca" version actually worked. Back then this card was still in production, popular and well supported, not to mention RX590 that was just been announced and used similar chip.

Radeon VII is actually nice for GPGPU, but support for it is being dropped from ROCm 6. Please note that it is high-end card from 2019, It came out just one year before RDNA2. And it's already being dropped when AMD says it brings support to more hardware.

Even the best hardware is worth nothing if you can't use it. And with AMD, GPGPU always feels like pure coincidence, a byproduct, and their plan to fix it is to have no plan at all. I seriously don't know why Tinycorp even considered Radeon for their product.

Nvidia split their consumer and data center cards back with Pascal. Even though the architecture name is the same on some architectures their SMs are completely different.

GP100 vs GP102
V100 vs TU102
A100 vs A102
H100 vs AD102

They considered AMD’s Radeon cards since the RX 7000 series has good inference performance.

[Quackdoc]

Preorders. It's still on their site. The full specs were on Twitter. Perhaps they will change the orders to Intel GPUs now, although the specs will change, and price will be lower. And that will last until they decide that Intel isn't good enough either.

I personally don't like the way they're communicating and their inconsistency. It feels very high risk. I contrast this with Bernhard Guentner from GPTshop.ai, who I find quite trustworthy.

as far as I know, the goal of the tinybox was always just going to be a perf tier, I didn't think they had locked in AMD at all for this.

[ET3D]

as far as I know, the goal of the tinybox was always just going to be a perf tier, I didn't think they had locked in AMD at all for this.

The specs match the AMD hardware they had. They can't have 144GB of RAM using six Intel cards with 16GB.

[pWe00Iri3e7Z9lHOX2Qx]

People have to be able to use AMD's software in more places than where AMD supports. It doesn't matter if you're on Arch, Fedora, or Debian, NVIDIA and Intel's software will work. AMD's software doesn't even support Arch, Fedora, or Debian. Both "open source" and "available to use" aren't always mutually exclusive.

Then add in the whole codec challenge that's a shitshow on AMD cards with some big distros like Fedora and OpenSUSE. 3rd party repos for Mesa and codec packages always end up having issues at some point.

[jeisom]

Honestly, if I had a compute workload that I'd want run on a enterprise cloud compute server, I'd want to be able to test it locally first. From all I've heard rocm is buggy as hell on radeon hardware. On Nvidia, you can test just fine locally on consumer hardware and move it when ready. I wouldn't want to be debugging while paying by the hour. If AMD wants to be truly taken seriously in compute they need to get there compute stack working from top end to bottom end.

There ryzen hardware work everywhere, why is radeon such a mess?

[deusexmachina]

I don't get why loads of AMD employees get to LARP on social media like LinkedIn without ever a single comment from the actual geeks and power users who limit themselves to hacker news, reddit, phoronix, etc (cower in a corner of the internet). Are the geeks so scared of not getting work in the future that they won't write the same critiques using their own real names in the LinkedIn comments and corporate emails? I did this in the internal news sites of large companies I worked for - and I was basically the only one doing it - and while it did "cause trouble," I actually was able to create significant change that way. It was great actually: I found corruption, they called HR on me, then I was like "great, I'm glad HR is involved, perhaps they can help me with this corruption I found." Fear is the mind killer.

People are so incredibly meek now yet expect their anonymous comments on random websites to cause change - while 99% refuse to do something as easy as leave a comment in the right place or bother their ex-colleague that now works for that company. All of the nonsense posting from AMD on LinkedIn (and other companies, of course) should be full of comments with your supposedly good ideas and critiques! At this point the "they" who are making all of these incompetent moves we complain about is actually "us."