The State Of ROCm For HPC In Early 2021 With CUDA Porting Via HIP, Rewriting With OpenMP

[bridgman]

to be honest it sounds like AMD is near bankrupt...

Not today, but a few years back things were very tight and obviously the state of the ROCm stack today is a function of the funding we had at the time. The fundamentals are all there but unfortunately no money for anything past core functionality targetting core business plans.

We were losing money right up to 2018, and 2020 was the first solidly profitable year in the last decade - arguably closer to 2 decades since the 2010/2011 profits were small and short lived. We did manage to develop a lot of technology without a lot of money (or sleep), but now that we are finally profitable again we can start putting those profits back into funding growth.

to sell a ~1500€ 6900XT and its out of stock all the time and then tell people: "We were not able to get funding/staffing at the time to implement on newer consumer cards" is a hard pill to smallow for people who buy 6900XT...

We aren't selling 1500 euro 6900XT's, we are selling $999 US 6900XTs (we actually sell just the chips but you know what I mean) and the supply chain is marking them up. They don't actually send us "oh Mr. AMD we charged too much for your card so here is the extra money" letters.

your company should try some very different business practice if you can't fund/staff something but the people really want something maybe do Crowdfunding campain. this really would be much better than what you do.

Again, if you can suggest something we could do today that would give us hundreds of millions of dollars in R&D funding a few years ago that would be wonderful, and I'm sure we would be all over it. I'm not sure how big crowdfunding projects get but (a) I don't think they get that big and (b) I don't think they have the time travel part figured out yet:

https://www.statista.com/statistics/...nt-since-2001/

[bridgman]

well at least it is nice to see that your $12,425.99 per card datacenter customers do a little subsidize the rest of the products line.

https://www.dell.com/en-us/work/shop/amd-mi100300w-pcie-32gb-passive-double-wide-gpu-customer-install/apd/490-bggd/graphic-video-cards

could AMD also do some cheaper CDNA "Arcturus" cards to ? for miners or other non Large Datacenter customers ?

Hah... if it makes you Europeans feel any better the same card is nearly $25,000 (18,500 US) in Canada:

https://www.dell.com/en-ca/work/shop/amd-mi100300w-pcie-32gb-passive-double-wide-gpu-customer-install/apd/490-bggd/graphic-video-cards

[extremesquared]

At the time, 3-4 years back, yeah it was very tight and obviously the state of the ROCm stack today is a function of the funding we had at the time. The fundamentals are all there but unfortunately no money for anything past core functionality targetting core business plans. We were losing money until 2018, and 2020 was the first solidly profitable year in the last decade (arguably almost 2 decades since the 2010/2011 profits were small and short lived). We have managed to develop a lot of technology without a lot of money, and now that we are finally profitable again we can start putting those profits back into funding growth.

My main concern is that AMD leadership could take a step back, look at the market, and decide that they are comfortable permanently conceding compute to nvidia. It's not a simple equation -- whether ROCm (on consumer cards) will be profitable. The amount of money lost to student and amateur developers stuck buying affordable nvidia cards that already have the necessary functionality isn't obvious without market research, and is plausibly a small percentage of purely-gaming-oriented purchase decisions. And future profitability of getting a "foot in the door" among comp sci students is utterly impossible to measure.

[qarium]

The only compute-specific functionality is a block in the Command Processor (CP) called MEC (Micro-Engine Compute), and even without a compute API the MEC block supports all the compute queues from DX and Vulkan. I don't think there is any hardware sitting idle.

we can say that there is compute only hardware in CDNA the 64bit floatpoint units...
games don't use this games use 32bit or lower...

i don't know any games who use 64bit fp...

[qarium]

We aren't selling 1500 euro 6900XT's, we are selling $999 US 6900XTs (we actually sell just the chips but you know what I mean) and the supply chain is marking them up. They don't actually send us "oh Mr. AMD we charged too much for your card so here is the extra money" letters.

you are right. but i as a consumer i use geizhals.de price search and the cheapest 6900XT is 1550€

https://geizhals.de/?cat=gra16_512&xf=132_16384%7E9809_05+11+-+RX+6900+XT

sure they do not give AMD money because they charge an extra 550€ but this shows there is a lot of money in AMD chips right now.
it can be seen like this: AMD did a exelent job with this generation of hardware if people charge 550€ extra on top of the AMD price.

Again, if you can suggest something we could do today that would give us millions of dollars in funding (a lot of millions) a few years ago that would be wonderful, and I'm sure we would be all over it.

i talked with you about crowdfunding long time ago. you said something like the community should do it instead of AMD- but for everthing what does not generate new hardware sales like anti-obsolecense stuff should be done by crowdfunding and for sure by AMD. if the people want to reduce electronic waste like the AMDGPU GCN1.0/1.1 VGA analog output stuff the people should pay money to such a crowdfunding campain. and anyone should be happy about such crowdfunding because its clear stuff like this does not generate any new hardware sales.

[Sonadow]

Willing to bet tens of thousands of Windows guys are just as frustrated with ROCm as any FOSS Einstein. Ultimately ROCm is not fundamentally FOSS. It is support software for very expensive hardware. People pay for this. It's not FOSS even if it's open source.

But then you didn't come here for logic, did you.

I'm being perfectly logical here. It's the common so-called community's boast that they can do anything with the sources available. Well all the sources are available, so they better get cracking on ways to patch and build ROCm and its components for various kernels and distributions all by themselves without AMD's help.

If they can't do that. then it's high time to permanently dispense with the 'give us the sources and we'll do everything ourselves' nonsense.

Besides, ROCm isn't even available for Windows, so your comparison is totally void.

[wallcarpet40]

If you run clinfo what does it say? clinfo is a seperate package.

clinfo for the 6800XT --> https://pastebin.com/JXjihVCN

[seesturm]

The kernel source code goes upstream quite aggressively (with the exception of a couple of non-upstreamable bits like RDMA) and the binary packages are organized so that you can either install userspace only (with a newer kernel) or userspace and kernel drivers (with an enterprise distro's older kernel).

Why is the RDMA part not upstreamable? Or what does RDMA mean in this context?

I'm asking because it would be nice to have this in upstream: https://github.com/RadeonOpenCompute...210b529715f8ce

[gsedej]

I am asking why I have to pay for the expense of the compute units if they can't be used due to the drivers. It just runs up the cost of the card. I want working compute but I don't want to pay for it if it doesn't work.

Just don't pay for AMD GPU if you depend on compute. Once AMD GPUs gets known for solid compute support like nvidia's CUDA or AMD's mesa 3D drivers, then go for it.

[vegabook]

I'm being perfectly logical here. It's the common so-called community's boast that they can do anything with the sources available. Well all the sources are available, so they better get cracking on ways to patch and build ROCm and its components for various kernels and distributions all by themselves without AMD's help.

If they can't do that. then it's high time to permanently dispense with the 'give us the sources and we'll do everything ourselves' nonsense.

Besides, ROCm isn't even available for Windows, so your comparison is totally void.

That ROCm isn't available for windows actually reinforces my point, if you use your brain. Windows users are paying for GPUs with not even any chance of compute. With Linux we at least have a chance. Compare that with CUDA where even a 99$ Jetson Nano has Cuda 10 support working right out of the box, as does every single retail GPU Nvidia sells.

And your main point also misses the mark. You criticize Linux guys' contention that they can compile their own code and fix things, by showing that this ROCm situation proves that they can't. But that's besides the point here (even if your comment was in good faith, which it isn't). The point is that compute capability is PAID FOR by the GPU buyer and therefore should work without the need for user to fix the software themselves. I'll remind you that AMD made a big deal out of Radeon VII FP64 capability so AMD clearly markets the compute capability of retail cards.

But look, I understand. You didn't come here to talk about the issues in the thread. You came here to dump on Linux. Well done you've succeeded. Now please could you do this in a Windows forum or wherever else you hang out instead of just coming here to rile people.