Intel's Open-Source Compute Runtime Appears To Be Ready For DG2/Alchemist dGPUs

[qarium]

Thanks a lot for your insightful comment, I appreciate all the details. Let's hope that with their recent acquisitions AMD can come up with something sane for all of their chips. And I also was a vocal critic of their GPU architecture split, we got to see what this leads to for end users (being the least priority, RDNA1/2 compute support being absent for years).

technically the architecture split is genius it saves a lot of transistors on these compute cards and make the completive. AMD just need to hire more software people and make sure the drivers and software works well.

but bridgman did tell us this many times it was ~2017 AMD was near-bankrupted they had no money to invest their stock shares where low so it was not possible to just print more stock shared to earn money to invest. it was the time the first ryzen cpus and first threadrippers did come into market and the time of the RX480 and vega64 ….. all my stuf i have threadripper 1920x and vega64 is from this time means this saves their company or else AMD would has been gone forever.

in this time 2017 was the time they developed the "architecture split" they only had money to develop ROCm for the CDNA cards and no money to develop ROCm for the RDNA cards and one is clear without the architecture split the compute cards would not be competive in compute and the gaming cards would not have been competive in gaming... because the Vega64/radeon7 lost the compute market and also lost the gaming market.

you have no idea what would have happened if there was no architecture split... AMD would had lost both markets the compute markets and also gaming markets.

today 2022 and 2021 AMD had no money problem they now invest money in doing all this but as we see from 2017-2020 it need 3-4 years to develop new hardware--- this means you can expect good hardware support in 2-3 years...

also blame the LAW it is agaist the law to support CUDA on AMD hardware on binary level...

this means the CUDA monopoly is protected by the law.

[hoohoo]

as fahr as i know ROCm supports vega64/Radeon7 and the amd 6800WX (what is a radeon 6800XT) so your claim that you need a 5000 dollar mi 100 is wrong.
but sure the vega64 and radeon7 is only to buy on ebay used and very expensive in comparison of other new cards.
the jump to the amd 6000series is also not optimal because this means amd does not really support the 5000cards in ROCm but they claim they work on this.

and i have to say this to you: do not blame AMD instead blame the "LAW"

Nvidia has a monopoly with CUDA and the LAW makes it agaist the law that AMD supports binary compatible CUDA api on their hardware.

ROCm/HIP goes over the source-code to make compatibility with CUDA because to support the CUDA apu directly on AMD hardware is agaist the law. this is seen in US law Oracle/Java vs google/andorid in this court case google only won because they did not sell a product that is 100% java compatible they did sell a incompatible version and had only similar functions in source code.

AMD can not solve any of these problems only the LAW maker could solve these problems.

the government should make the law in a way that makes monopoles agaist the law or else make CUDA binary api forced-opensource that any company can use it. or else the government need to spend 1000 of billions of dollars to make webassembly/WebGPU the new gpu compute standard.

Show me a listing to sell a Radeon 7.

HIP is a source level shim that compiles much or most CUDA to opcodes that execute atop the COMc runtime.

As for monopolies (also as for correct grammar and capitalization), I suggest you educate yourself further.

[hoohoo]

technically the architecture split is genius it saves a lot of transistors on these compute cards and make the completive. AMD just need to hire more software people and make sure the drivers and software works well.

The split put ROCm only within reach of people with deep pockets. The split hugely limited ROCm's mass appeal. The split made it such that the huge body of millions of programmers in the world were locked out of ROCm.

And before you start babbling about Vega64 & Radeon VII: the VII was not made in huge numbers in the first place and both vanished into the void of crypto mining once ppl found out they were the best Ethereum cards. Vega64 and Radeon VII were not mass market cards. Sh!t I sold my two VIIs to a miner for double(!) the purchase price a year after they were EOL. And you know why I sold them? The ML toolkit I was invested in stopped working on ROCm after a ROCm minor point release (it wouldn't even compile from source anymore), and I couldn't even get the ROCm devs to respond to pleas for help or pointers to fix the code myself. Real genius in the dev side.

Real genius in the marketing too:

"Hey! Lets make our new GPGPU stack as FUCKING DIFFICULT AND EXPENSIVE AS POSSIBLE for all the people who might use it and drive mass adoption. Lets make it only really available to the scientists at the National Labs in the USA".

That's a fail. That's AMD's classic fail. A marketing dept that still thinks it live in the Big Tin era not the mass market open source era.

[ms178]

That's a fail. That's AMD's classic fail. A marketing dept that still thinks it live in the Big Tin era not the mass market open source era.

I would argue that it wasn't only a marketing fail but that it became a bad technical decision to stick to ROCm over the years. As Intel announced oneAPI, Level 0 etc. they should have taken notice and abandon ROCm sooner rather than later, Intel even pays the University of Heidelberg to make AMD cards run on that stack. I also agree with you totally on the hardware side of things, AMD clearly has still not the ressources to support both architectures fully, hence you want as much re-use the work between client and HPC offerings, especially on the software side with a lower barrier to entry on the hardware side.

[qarium]

Show me a listing to sell a Radeon 7.

right its near to impossible to buy one. i already have a vega64 ...
on the other side many replace a vega64 with lets say a 6800xt and sell the vega on ebay.

HIP is a source level shim that compiles much or most CUDA to opcodes that execute atop the COMc runtime.

right source level... but the market do want binary compatible cuda api what is agaist the law.

As for monopolies (also as for correct grammar and capitalization), I suggest you educate yourself further.

i am a german with dyslexia i already did spend 20+ years in lerning english.

be sure i know what a monopole looks like and what it is. and nvidia has a monopole in the gpu compute sector.

the "LAW" is part of the problem why we have this problem with nvidia.

[qarium]

The split put ROCm only within reach of people with deep pockets. The split hugely limited ROCm's mass appeal. The split made it such that the huge body of millions of programmers in the world were locked out of ROCm.

thats all because of the lag of money in the AMD company not because of the split.

remember this: before the split happened AMD lost both markets the gaming and the compute market.
Nvidia 1000 and 2000 series annihilated the vega64 and also the radeon7 in gaming
Nvidia cards on CUDA did annihilate all the AMDs card in any market but crypto mining.

the split did make it possible for AMD to stay competive in the compute market and also become competive in the gaming market. thats because the split of CDNA saved more than 50% of all transistors on the chip.
the core count because of this did go from 4096 to 8182 without any thermal problems just because the split did save 50% of all tranistors.

all the negative effects you tell us is not because of the split but because of the lag of money at amd.

And before you start babbling about Vega64 & Radeon VII: the VII was not made in huge numbers in the first place and both vanished into the void of crypto mining once ppl found out they were the best Ethereum cards. Vega64 and Radeon VII were not mass market cards. Sh!t I sold my two VIIs to a miner for double(!) the purchase price a year after they were EOL. And you know why I sold them? The ML toolkit I was invested in stopped working on ROCm after a ROCm minor point release (it wouldn't even compile from source anymore), and I couldn't even get the ROCm devs to respond to pleas for help or pointers to fix the code myself. Real genius in the dev side.

i already know your story. and yes it is a sad story. amd was not ready to enter the compute market outside of the crypto boom.

Real genius in the marketing too:
"Hey! Lets make our new GPGPU stack as FUCKING DIFFICULT AND EXPENSIVE AS POSSIBLE for all the people who might use it and drive mass adoption. Lets make it only really available to the scientists at the National Labs in the USA".
That's a fail. That's AMD's classic fail. A marketing dept that still thinks it live in the Big Tin era not the mass market open source era.

right it is sad but how to deal with an unbeatable monopole like cuda ?
what would you do in AMDs place without money to invest in software developers ?
without money to invest you can do nothing about this.

to my knowelege without the architecture split AMD would had go out of business in the GPU market
because they lost on gaming and compute the same time.

with the architecture split it was possible for AMD to survive in the supercomputer market.
also the radeon 5000 and 6000 series did gain much performance for games because of the architecture split
it was like 25% for the 5000 series at the same tranistor count and like 50% for the 6000 series.

[hoohoo]

right its near to impossible to buy one. i already have a vega64 ...
on the other side many replace a vega64 with lets say a 6800xt and sell the vega on ebay.



right source level... but the market do want binary compatible cuda api what is agaist the law.



i am a german with dyslexia i already did spend 20+ years in lerning english.

be sure i know what a monopole looks like and what it is. and nvidia has a monopole in the gpu compute sector.

the "LAW" is part of the problem why we have this problem with nvidia.

Nvidia has competiton from several sources. It is not a monopoly.

[qarium]

Nvidia has competiton from several sources. It is not a monopoly.

like the intel ARC gpus who fail on any benchmark ?