Open-Source Radeon Tools Updated With Expanded RDNA(2) Support, Other Features

[phoronix]

Phoronix: Open-Source Radeon Tools Updated With Expanded RDNA(2) Support, Other Features

In addition to NVIDIA releasing new open-source GameWorks projects this week for the Game Developers Conference, AMD with their GPUOpen initiative has released several updated Radeon Windows/Linux tools...

https://www.phoronix.com/scan.php?page=news_item&px=Radeon-GPUOpen-GDC21

[hoohoo]

Support the HIP based toolset on RDNA2 and I'll buy a 6900XT. If not, not.

[hoohoo]

Radeon Pro W6800 with 32gb vram isn't that the product exactly for this ?

is HIP in ROCm not yet done ?

W6800 = 3840 shaders, 17.83 TFLOP FP32, $2249 launch price.
6900XT = 5120 shaders, 23.04 TFLOP FP32, $999 launch price.

I need FP32 TFLOP.

Which card would you want to use? The card AMD wants you to use or the card your wallet wants you to use?

HIP is an integral part of ROCm.

AMD does not support ROCm on RDNA1/2 products.

AFAIK ROCm will not work on a W6800 either, since it is an RDNA2 card just like the rest of the 6000 series cards.

If it did support ROCm on RDNA2 then you could get 46 TFLOP FP32 for less money than the price of the Expensive Edition card by running a pair of 6900XT cards. Or you could spend more money on the Expensive Edition card and get 17.83 TFLOP and the patronizing approval of AMD.

Yes, I know about the price bubble for GPUs, but AMD prices have been falling and probably we will see MSRP or nearly so before the end of the year.

[jrch2k8]

W6800 = 3840 shaders, 17.83 TFLOP FP32, $2249 launch price.
6900XT = 5120 shaders, 23.04 TFLOP FP32, $999 launch price.

I need FP32 TFLOP.

Which card would you want to use? The card AMD wants you to use or the card your wallet wants you to use?

HIP is an integral part of ROCm.

AMD does not support ROCm on RDNA1/2 products.

AFAIK ROCm will not work on a W6800 either, since it is an RDNA2 card just like the rest of the 6000 series cards.

If it did support ROCm on RDNA2 then you could get 46 TFLOP FP32 for less money than the price of the Expensive Edition card by running a pair of 6900XT cards. Or you could spend more money on the Expensive Edition card and get 17.83 TFLOP and the patronizing approval of AMD.

Yes, I know about the price bubble for GPUs, but AMD prices have been falling and probably we will see MSRP or nearly so before the end of the year.

ROCm is a utter mess and if i were you i would not wait until that mess is sorted if you need it for work, this is the one case where you have to go nVidia or you are dead on the water.

My recommendation would be a cheap AMD primary for desktop and RTX secondary for compute

[hoohoo]

ROCm is a utter mess and if i were you i would not wait until that mess is sorted if you need it for work, this is the one case where you have to go nVidia or you are dead on the water.

My recommendation would be a cheap AMD primary for desktop and RTX secondary for compute

I have used ROCm with good results on a pair of Radeon VII I used to own. Would you care to explain how ROCm is "an utter mess"?

Sure, CUDA is a better product overall, no question. Better docs, better samples and examples, wider ecosystem. OTOH if one is compiling code for ROCm or for CUDA, and the code already supports both, then currently if RDNA2 was supported then AMD would have better bang for the buck. Certainly Radeon VII did for FP32 vs nVidia's cards.

[hoohoo]

you really really missunterstand me.
i am not here to tell you buy a W6800 because i have stock market shared of AMD to rip off your 2249dollars...

no its not about this. but amd first release an architecture "RDNA2" and then there is not PRO card like W6800
the reason for this is simple: the driver is not ready to fill into the role of a PRO card.
then 1-2 years later the driver becomes ready and then they launch the PRO card.
and this is the time the driver also become stable for non-pro cards like the 6900XT...

the release of the RDNA2 W6800 just means the driver is done and very well aged like a good wine.



last time i checked this topic in the forum the result was the release note did claim ROCm does not work for RDNA2 but people in the forum did say it work for them. and this was even before the launch of the RDNA2 W6800

after the launch of the RDNA2 W6800 you can be 100% sure ROCm works on RDNA2...



really in the past i could easily say the PRO cards is not made in a way i could justify the price
but for the W6800 .,.. man a RDNA2 6800XT has 16gb vram and the W6800 has 32gb vtam and i found many people in the forum who did run out of vram memory with 16gb vram... it is also ECC ram and has other features like Single Root I/O Virtualization (SR-IOV)
so i can say for sure if someone need more than 32gb vram and advanced VM features the price of a W6800 is higher but it also brings you a lot of more features...

right now i have a Vega64 with 13TFLOPs and with AMD Super Resolution Fidelity FX the card pushes nice high FPS with very good visual quality. in my point of view there is no quality lose if you use the 90%/super high quality mode of SRF but you get a nice FPS uplift.

I did not think you were trying to rip me off. I spoke too strongly, sorry about that. I was trying to point out that for my use case the 6900X is a much better card if it were supported by ROCm.

ROCm works on RDNA2 only via it's OpenCL binding. The HIP part of ROCm does not work with RDNA2. The problem is that all the stuff I want to use either was written for CUDA & thus is a good candidate to be HIP-ified and run on ROCm, or already supports ROCm directly.

AFAICT AMD is spending it's resources making sure the mi200 will work as promised in the upcoming US national labs' supercomputers it has the contracts for. This is fine. I just think it is important to keep reminding AMD that there is money to be made selling consumer cards to people who need FP32 compute and who are not working for the US Fed.

[MadeUpName]

AMDs moto: Maybe our shit works and maybe it don't.

Graphics cards are way to expensive and critical for me to gamble with any more.

[jrch2k8]

I have used ROCm with good results on a pair of Radeon VII I used to own. Would you care to explain how ROCm is "an utter mess"?

Sure, CUDA is a better product overall, no question. Better docs, better samples and examples, wider ecosystem. OTOH if one is compiling code for ROCm or for CUDA, and the code already supports both, then currently if RDNA2 was supported then AMD would have better bang for the buck. Certainly Radeon VII did for FP32 vs nVidia's cards.

1.) Only works on a handful of GPU most of the time
2.) Very hard to build consistently
3.) Tend to be very fragile after build or may require of out tree patches
4.) Very little community input

Honestly, i would only use the closed RHEL/Ubuntu packages for the exact version they are built on a Radeon VII or VEGA 64 , on anything else specially rolling distro you will end up with a broken tool set 90% of the time and if you try it on Polaris like me is even worse.

Your best hope is to wait until AMD give up and decide to support Karol's work on Clover over NIR implementation for compute and someone take the usable parts of HIP to make a CUDA state tracker

[StillStuckOnSI]

AFAICT AMD is spending it's resources making sure the mi200 will work as promised in the upcoming US national labs' supercomputers it has the contracts for. This is fine. I just think it is important to keep reminding AMD that there is money to be made selling consumer cards to people who need FP32 compute and who are not working for the US Fed.

At the very least, they are actively working towards Navi support in many of the HIP ecosystem projects. For example, here's one tracker issue for Navi 21 on the DNN library: https://github.com/ROCmSoftwarePlatf...en/issues/1022. To your point though, I would not be willing to pull the trigger on expensive consumer or prosumer cards until an actual stable release materializes.

[StillStuckOnSI]

Your best hope is to wait until AMD give up and decide to support Karol's work on Clover over NIR implementation for compute and someone take the usable parts of HIP to make a CUDA state tracker

Though I would much, much rather see this than a vendor-specific solution, realistically we're going to see broader ROCm support before https://gitlab.freedesktop.org/mesa/..._requests/7231 and any follow-up PRs make their way to Mesa stable. My understanding is that Clover + compute stuff is at best a side project for most of the Mesa devs, so progress is going to be slow in the absence of additional people or money being thrown at the problem. Who knows though, maybe the stars will align and we'll have https://www.youtube.com/watch?v=KfDQb6xOkXg by the end of next year.