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

[dragonn]

And bugs too, this is why Torvalds gave them the middle finger some years ago.

Torvalds did give Nvidia the middle finger because they refused to support VirtualGL in they driver with made dual GPUs on laptop with nvidia impossible to use with they closed driver at the time. It had nothing to do with any bugs.

[aufkrawall]

And bugs too, this is why Torvalds gave them the middle finger some years ago.

Torvalds giving them the middle finger has exactly nothing to do with how reliable cuda is.
Nobody is switching away from cuda to something else because some bugs force them to. (Seriously, who would truly believe that...)
This is another 100% self-inflicted AMD fail.

[skeevy420]

Humm, TinyCorp is saying exactly the opposite. They WANT to debug the software stack, all the way to the firmware, but they cannot!

All the way to the firmware is not likely to happen with an AMD GPU. Perhaps tiny could sign an NDA to get access to the firmware's code, but they'd still have to release it as a closed blob. I wouldn't be too surprised if AMD has that kind of deal with Sony and Microsoft's game consoles.

At the same time tiny is also saying that they don't mind a black box if it just works. Because of that, I can't put too much stock in their "One of these days Alice, all the way to the firmware" comment.

And bugs too, this is why Torvalds gave them the middle finger some years ago.

I don't think bugs was the case.

This is another can of worm. AMD is shit on the software front and we know it. Open or not, it will not change. This is why openness would give the community the power to make it better. If they cannot, at least let the community do it. This is the question TinyCorp is raising.

But that's the thing, the driver has been open since 2014-ish and ROCm has been open since 2016-ish (don't quote me on those times, I'm going on memory). The community has had 10 years and has done a lot with the AMDGPU driver and open MESA drivers like RADV while ROCm and AMDVLK have improved at a much slower rate. AMDGPU, however, is available everywhere while ROCm isn't. RADV has an open and easy to follow commit history. AMDVLK gets code dumps. They're all open, but they're not the same kind of open.

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.

[jindam]

Both "open source" and "available to use" aren't always mutually exclusive.

its ironic, however, unfortunately its a reality

[WannaBeOCer]

Tenstorrent Grayskull would be a good solution but imagine Intel Arc A750, a cheap $240 GPU. Their margins would be higher.

[ET3D]

TinyGrad is a joke. It's one thing to regularly move between "will do" and "won't do" when you don't actually have a product, but when you're taking orders for a product and then you decide out of the blue to cancel it, that's different. It's clear that this isn't a company that others should work with. Even if it does get a product on the market, this kind of behaviour suggests that there's no way to trust TinyGrad for support over time.

[sobrus]

The truth is, Radeon and Ryzen has always been gaming-only parts for AMD. The only performance metrics AMD uses for both of them are FPS.
GPGPU may or may not work. Doesn't matter. "They are working on it".

I keep my fingers crossed for Intel Battlemage and Celestial parts. Lack of competition is really bad. nVidia has no competition in GPGPU segment, AMD has (had?) no competition in open-source segment. That's why we've got what we've got.

Once we get decently performing ARC GPU with mature open-source drivers, nobody will have to ask AMD to fix or enable anything.

[varikonniemi]

WTF is wrong with AMD? They couled have had it all, but instead they decided to move over to a GFX card architecture where a proprietary (and buggy) OS runs on their card so having open drivers means nothing.

[Quackdoc]

TinyGrad is a joke. It's one thing to regularly move between "will do" and "won't do" when you don't actually have a product, but when you're taking orders for a product and then you decide out of the blue to cancel it, that's different. It's clear that this isn't a company that others should work with. Even if it does get a product on the market, this kind of behaviour suggests that there's no way to trust TinyGrad for support over time.

where were they taking orders on a project they canceled?

[WannaBeOCer]

The truth is, Radeon and Ryzen has always been gaming-only parts for AMD. The only performance metrics AMD uses for both of them are FPS.
GPGPU may or may not work. Doesn't matter. "They are working on it".

I keep my fingers crossed for Intel Battlemage and Celestial parts. Lack of competition is really bad. nVidia has no competition in GPGPU segment, AMD has (had?) no competition in open-source segment. That's why we've got what we've got.

Once we get decently performing ARC GPU with mature open-source drivers, nobody will have to ask AMD to fix or enable anything.

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