Intel Compute Runtime 24.45 vs. AMD ROCm 6.3 vs. NVIDIA R565 Linux GPU Compute Benchmarks

[Grinness]

they are probably focusing on AI now - and if that helps with Compute (incidentally), that's the only chance.

What you think AI is based on?
Magic?!

Ai algorithms are based on large scale matrix multiplications -- pure maths
Improving "AI" means improving compute -- specifically hardware level support for matrix multiplications

[Grinness]

For the ppl like Panix that seems to not understand Ai, what enables it and some of the terminologies related, for example "tensor cores":

A tensor may be represented as a (potentially multidimensional) array. Just as a vector in an n-dimensional space is represented by a one-dimensional array with n components with respect to a given basis, any tensor with respect to a basis is represented by a multidimensional array. For example, a linear operator is represented in a basis as a two-dimensional square n × n array.

https://en.wikipedia.org/wiki/Tensor

Tensor == Matrix

AMD calls their hardware implementation "matrix cores", perfectly inline with the meaning of "tensor"
Nvidia uses "tensor cores"
(intel I don't know)

[Opteron 170]

Also AMDs last gen was the worst in the last 10 years probably, Intel don't have that excuse... for them it's more a question if this headless company even produces any cards after this gen, or if they only sell 100 per country and then say "ohh we are sold out because so many people want it"...

How is RDNA 3 worse than RDNA 2 & 1?

[Panix]

For the ppl like Panix that seems to not understand Ai, what enables it and some of the terminologies related, for example "tensor cores":

https://en.wikipedia.org/wiki/Tensor

Tensor == Matrix

AMD calls their hardware implementation "matrix cores", perfectly inline with the meaning of "tensor"
Nvidia uses "tensor cores"
(intel I don't know)

For morons like Grinness who don't realize that ppl who use AI applications use Nvidia gpus.

[oleid]

Yes I have the XT and XTX. But this comparison is mid-range focused as noted in the article given the preparations for Battlemage... Later on will do a full stack comparison for all vendors, mostly time limited at the moment.

Would it be possible to include rusticle as an alternative OpenCL driver for at least some of the benchmarks?
At least the AMD cards should be able to use this OSS OpenCL implementation.

[oleid]

ppl who use AI applications use Nvidia gpus.

Depends. I use AMD since they have more memory / price for my personal LLM needs.

[Grinness]

For morons like Grinness who don't realize that ppl who use AI applications use Nvidia gpus.

Panix,

before insulting people you don't know about topics you demonstrate to know nothing about:

1. wash your mouth with soap, rinse and repeat 100 times
2. study and learn

[mrg666]

Panix,

before insulting people you don't know about topics you demonstrate to know nothing about:

1. wash your mouth with soap, rinse and repeat 100 times
2. study and learn

Yeah, he is known to be an idiot around here.