AMD Launches The Accelerator Cloud To Try Out EPYC CPUs, Instinct GPUs + ROCm

[vegabook]

I don't think you are a minority. The "consumer first or datacenter first" subject was debated hotly inside the company but the decision at the time was to focus entirely on datacenter.

I felt that we could maintain a focus on datacenter while still doing enough to keep the consumer/workstation customers supported, but that was very much a minority view at the time. I think there is a better understanding these days about the importance of grassroots developer support (which implies consumer products, not just workstation) but there is still more work to be done.

We did make another important internal change a couple of months ago - integrating the release processes for graphics and ROCm stacks to the point that we twin the releases and use the same ROCm code for both. We haven't fully integrated the testing yet but what this should give us for the first time is testing on "GUI apps" (ie apps using compute/graphics interop) before releasing. That is not manifesting as happy AMDGPU-PRO users yet but we are trying hard to pull our internal testing into alignment with what our customers are doing with the drivers.



This part I don't agree with - I think I have always been pretty clear that we were working on RDNA support but I couldn't say when it would arrive because we we had to prioritize other work (the big datacenter parts) higher.

I'm not sure, but it seems like you think we were doing nothing for a couple of years and then suddenly got off our a**es and ported the entire compute stack to RDNA2 in a few weeks. That's the only explanation I can think of for you thinking I was obfuscating. As far as I know the only change in what I have been saying is...

- we're working on it but I don't know how long it will take
- we're working on it and everything in the ROCm stack up to OpenCL is shipping as the compute solution for AMDGPU-PRO on RDNA and RDNA2
- we're working on it and RDNA2 is pretty close to official support right up to the ML frameworks (and people are using it already albeit with some tweaking)
- (hopefully soon) we're working on it, we have official support for RDNA2, <this subset> works for RDNA, and we're fairly close to finishing RDNA

To me that's the same message all the way modified only by the passage of time.



Don't think so. It does mean that I don't come on here and rant about internal problems but it doesn't mean I say things that are either untrue or misleading. I don't get paid anywhere near enough to do that.

Here's what somebody as privy to the internal machinations of AMD, as you but who wasn't actually being paid a penny by AMD, would have said.

"Fellas, forget about AMD compute for the next 3 years. We hate the situation, but it's what we gotta do.". Okay that would probably have been suicide in the market. More realistically, at your weekly engineering/marketing/PR/whatever meeting: "guys, were taking an absolute smackdown in devland on compute. Let's give these guys somethin', yeah?"

What did we get? Lots of @brigman stoical talk. Zero actual product.

Instead what we got, from Lisa Su, undenied by you, was "Vega 7 is a compute beast" (or words do that effect).

And don't argue: you purposefully de-gimped FP64 to make it look promising; thereafter to make sure we couldn't actually use it!!!

Such a mess.

@brigman your credibiltiy on compute is less than zero with me. I can't speak for others.

[vegabook]

Here's what somebody as privy to the internal machinations of AMD, as you but who wasn't actually being paid a penny by AMD, would have said.

"Fellas, forget about AMD compute for the next 3 years. We hate the situation, but it's what we gotta do.". Okay that would probably have been suicide in the market. More realistically, at your weekly engineering/marketing/PR/whatever meeting: "guys, were taking an absolute smackdown in devland on compute. Let's give these guys somethin', yeah?"

What did we get? Lots of @brigman stoical talk. Zero actual product.

Instead what we got, from Lisa Su, undenied by you, was "Vega 7 is a compute beast" (or words do that effect).

And don't argue: you purposefully de-gimped FP64 to make it look promising; thereafter to make sure we couldn't actually use it!!!

Such a mess.

@brigman your credibiltiy on compute is less than zero with me. I can't speak for others.

One More Thing (TM).

Nvidia is worth 4x what AMD is. What's the difference between Nvidia and AMD. No, it's not Gaming. It's COMPUTE!! Six. Hundred. Billion. Dollars. Toss that around in your mind.

[smitty3268]

@brigman your credibiltiy on compute is less than zero with me. I can't speak for others.

I feel like bridgman has always been fairly careful in what he says. The line was always "it's coming, but i can't say when" which I took to mean it's a long ways away and won't be here for years. I think some people may have read something else into that and assumed it was always right around the corner, but he never said that. I do agree he has very PR-like responses to some questions. You're always going to get an optimistic take from him, rather than a more realistic one. But that's ok.

That said, AMD has zero credibility on compute overall and it's going to take a long time with a lot of success to even get back into the conversation for me. I can't imagine anyone involved with gpu compute is excited about rocm when they already have cuda support working. Probably their best bet is to achieve hardware supremacy for the next few years, if they can. People still won't love going to AMD for compute, but if they can get faster performance that way than on nvidia hardware they'll end up going anyway.

[bridgman]

Here's what somebody as privy to the internal machinations of AMD, as you but who wasn't actually being paid a penny by AMD...

Those people don't exist. Please don't compare me to that hypothetical ideal... you're always going to be disappointed.

...would have said.

"Fellas, forget about AMD compute for the next 3 years. We hate the situation, but it's what we gotta do.". Okay that would probably have been suicide in the market. More realistically, at your weekly engineering/marketing/PR/whatever meeting: "guys, were taking an absolute smackdown in devland on compute. Let's give these guys somethin', yeah?"

What did we get? Lots of @brigman stoical talk. Zero actual product.

I'm not sure how to interpret that, given that Vega10 and Vega20 have been fully supported for those three years. If you just mean "forget about ROCm on RDNA for three years although you will have the usual OpenCL support and RDNA2 will be more like a year" that sort-of works, although I hope you understand that "I don't know" actually meant "I don't know" not "the schedule is XYZ and I won't tell you".

Instead what we got, from Lisa Su, undenied by you, was "Vega 7 is a compute beast" (or words do that effect).

Why would I deny that (assuming you mean Radeon VII (Vega20) rather than the 7 CU Vega 7 iGPU in some of our APUs) ? I have a Radeon VII in my development system at home (bought personally) and it has been supported since launch (it came out a couple of months after Instinct MI50/50). It was a compute monster then (although the datacenter and workstation versions were even faster) and still is pretty good today.

And don't argue: you purposefully de-gimped FP64 to make it look promising; thereafter to make sure we couldn't actually use it!!!

Why couldn't you use it ? Vega20 had full ROCm support from launch.

@brigman your credibiltiy on compute is less than zero with me. I can't speak for others.

OK, I am sorry to hear that, but I would appreciate you making a bit more effort to explain what I said that was not correct. So far you seem to be saying that instead of "I don't know how long it is going to take" I should have said "it's going to take three years" even though I had no idea how long it would take at the time.

[qarium]

Those people don't exist. Please don't compare me to that hypothetical ideal... you're always going to be disappointed.
I'm not sure how to interpret that, given that Vega10 and Vega20 have been fully supported for those three years. If you just mean "forget about ROCm on RDNA for three years although you will have the usual OpenCL support and RDNA2 will be more like a year" that sort-of works, although I hope you understand that "I don't know" actually meant "I don't know" not "the schedule is XYZ and I won't tell you".
Why would I deny that (assuming you mean Radeon VII (Vega20) rather than the 7 CU Vega 7 iGPU in some of our APUs) ? I have a Radeon VII in my development system at home (bought personally) and it has been supported since launch (it came out a couple of months after Instinct MI50/50). It was a compute monster then (although the datacenter and workstation versions were even faster) and still is pretty good today.
Why couldn't you use it ? Vega20 had full ROCm support from launch.
OK, I am sorry to hear that, but I would appreciate you making a bit more effort to explain what I said that was not correct. So far you seem to be saying that instead of "I don't know how long it is going to take" I should have said "it's going to take three years" even though I had no idea how long it would take at the time.

its their own fault ... i am on Vega64+TR1920X since 2017 and this combination was always supported by ROCm.
the 7nm Vega20 was even better than the Vega64 was i have.

the people plain and simple do not buy what is supported instead they buy what is the newest whatever i is.
and it is their own fault

we have 2 kind of AMD linux customers around here one group is the "their own fault" group they buy what IS NOT supported...

the other group buy what is supported by the software even if the hardware is 1-2 generations old.

in my point of view the "RX5700XT" series did make no sense compared to Vega64+Radeon7 the 5700 was only around 10 % faster than my vega64 without extra features like raytracing support... and the radeon7 was even faster than a 57000XT,,,

for the RX 6900XT series yes it is faster than vega65/radeon7 and has more features like AV1 and raytracing but in the end the smart people did buy Vega64/radeon7 until the driver and softwarte support catch up...

so we have the "their own fault" people they buy whatever gives them the greatest pain and we have people like me who prever stay on years old hardware to avoid any software problem.

believe it or not soon i buy a RX7800XT or 6800XT and it will run like a charm means without any driver or software problems.
and yes the hardware is soon 2-3 years old RX5000/RX6000 and it needs 3 years to become fine in driver and software...

sure i want AMD to perform a better support task for the "their own fault" fraction but i also know that these people will always blame others for errors they make for themself.

[qarium]

Look I can see that AMD has made some effort on OpenCL, my scepticism on that technology notiwthstanding. Moreover perhaps WGPU will require AMD to provide browser-based (and also through wgpu-rs therefore, native), GPU compute. And there's Vulkan compute. So there's plenty of hope.
That said I work constantly for a bunch of finance guys doing statarb in both bonds and crypto all day long here in London. Nobody ever saw an AMD compute card anywhere in that very big market. I can vouch for that myself.
So maybe the hyperscaler strategy pays off, but I can't help thinking that even TensorFlow/Torch/Jax etc were once a dream on someone's desktop and they reached out for, you guessed it, what was _available_, namely CUDA. Now sure, these guys don't want lockin at their scale so the AMD strategy makes some sense, I suppose. But it still needs to get this stuff onto desktop sharpish because the Next Big Thing, the creator market, which defo uses desktop compute, is running away from them, just when everybody is starting to sprinkle the "metaverse" into their business.
Also Edge. I'd buy an ROCm capable Jetson-like low-end AMD board tomorrow. Or AMD could support ROCm on the APUs and let third parties do it.

to bring this to the masses so normal people can touch the technology they maybe should do some chiplet design mix between RDNA2 and CDNA ... yea right this sounds crazy...
but for some people this could be a big win.
maybe together with an IO chip to share the vram...

[vegabook]

its their own fault ... i am on Vega64+TR1920X since 2017 and this combination was always supported by ROCm.
the 7nm Vega20 was even better than the Vega64 was i have.

the people plain and simple do not buy what is supported instead they buy what is the newest whatever i is.
and it is their own fault

we have 2 kind of AMD linux customers around here one group is the "their own fault" group they buy what IS NOT supported...

the other group buy what is supported by the software even if the hardware is 1-2 generations old.

in my point of view the "RX5700XT" series did make no sense compared to Vega64+Radeon7 the 5700 was only around 10 % faster than my vega64 without extra features like raytracing support... and the radeon7 was even faster than a 57000XT,,,

for the RX 6900XT series yes it is faster than vega65/radeon7 and has more features like AV1 and raytracing but in the end the smart people did buy Vega64/radeon7 until the driver and softwarte support catch up...

so we have the "their own fault" people they buy whatever gives them the greatest pain and we have people like me who prever stay on years old hardware to avoid any software problem.

believe it or not soon i buy a RX7800XT or 6800XT and it will run like a charm means without any driver or software problems.
and yes the hardware is soon 2-3 years old RX5000/RX6000 and it needs 3 years to become fine in driver and software...

sure i want AMD to perform a better support task for the "their own fault" fraction but i also know that these people will always blame others for errors they make for themself.

Oh please. You trying to say it's our own fault for not making AMD work nice on compute? haha. Hahahahahah

[qarium]

Oh please. You trying to say it's our own fault for not making AMD work nice on compute? haha. Hahahahahah

if you buy the product and you know it does not work nice on compute then yes it is your own fault.

[davide445]

Curious about this initiative since so far I discovered a single cloud provider I can test a Mi100 on.
Only I didn't find how to request access to AAC resources, is this trough AMD sales channels? There is any prerequisite?

[vegabook]

if you buy the product and you know it does not work nice on compute then yes it is your own fault.

Your comment is gratuitously sloppy and uninformed. The Vega VII was specifically pitched as excellent for compute. That's how long ago AMD promised compute, screwed everyone like me up the *rse, and still nothing satisfactory to this day.