gpuOwL: an OpenCL program for Mersenne primality testing

[Magellan3s]

That Radeon Instinct M250X-Aldebaran XT is stupid fast compared to anything else on the market!

I hope the 7900 XTX offers performance that is at least somewhat close to it on GPUOWL.

[axn]

Originally Posted by Magellan3s

That Radeon Instinct M250X-Aldebaran XT is stupid fast compared to anything else on the market!

I believe those numbers are estimated base on FLOPS scaling which is not realistic. They will probably be 20-30% slower in real life.

Originally Posted by Magellan3s

I hope the 7900 XTX offers performance that is at least somewhat close to it on GPUOWL.

Assuming a 2.7x (SWAG) improvement over 6950, it would be about 193 microseconds, which would be significantly faster than A100.

[frmky]

Also the MI250X (and MI250) are really two GPUs packaged together, like the nVidia K80. AMD hides this a bit in their marketing. So expect only half of that performance in a single gpuOwl run. And AMD is a bit behind nVidia in getting data from GPU memory to the cores, so the actual performance is closer to the A100 than the raw FLOPS numbers would suggest for most workloads.

[moebius]

Originally Posted by axn

I believe those numbers are estimated base on FLOPS scaling which is not realistic. They will probably be 20-30% slower in real life.

Assuming a 2.7x (SWAG) improvement over 6950, it would be about 193 microseconds, which would be significantly faster than A100.

The yellow values ​​are really only roughly extrapolated on the Mi100, but with the 6800 XT I was very close to the real value at the time, extrapolated on a 5700 XT. https://mersenneforum.org/showpost.p...5&postcount=24

And then there's the famous Bottleneck probably too for MI50 and MI60.

[moebius]

Originally Posted by Magellan3s

That Radeon Instinct M250X-Aldebaran XT is stupid fast compared to anything else on the market! I hope the 7900 XTX offers performance that is at least somewhat close to it on GPUOWL.

This one is rather close to the Instinct MI210. I've entered my predictions for the 7900 graphics cards into the list too.

[frmky]

A MI250 is really two MI210's in the same package. Since gpuOwl doesn't support multiple GPUs, a single run on a MI250 will run on one of them and give the same times as a MI210. You can just run two simultaneously, one on each GPU. Same for the MI250X.

[moebius]

Originally Posted by frmky

A MI250 is really two MI210's in the same package. Since gpuOwl doesn't support multiple GPUs, a single run on a MI250 will run on one of them and give the same times as a MI210. You can just run two simultaneously, one on each GPU. Same for the MI250X.

The technical basis of the Instinct MI250(X), codename Aldebaran, is two dies linked by a 400 GByte/s fast Infinity Fabric (via Elevated Fanout Bridge, EFB). They are manufactured with TSMCs N6, i.e. 6 nm EUV, and have 29.1 billion transistors. The two chips are called GCD (Graphics Compute Die) , each of which has four shader arrays for 112 compute units with a total of 7,168 ALUs at 1.7 GHz when fully expanded. The L2 cache still holds 8 MB, but its bandwidth has been doubled - for good reason: the matrix cores, similar to Nvidia's tensor cores, deliver double or quadruple the rate per clock. The Instinct MI250X as the top model achieves around 96 teraflops with double precision (FP64), without the matrix cores it is still 48 teraflops.

I conclude that mfakto will scale just as well as on a MI100, but gpuowl much better.

But I'm only human, I can also be wrong.

[frmky]

But to the user they appear as two distinct GPUs with separate memory spaces. Data transfer between the two over Infinity Fabric is much slower than HBM2, and applications must be coded to support multiple GPUs to use both.
https://chipsandcheese.com/2022/09/1...-architecture/ (5th paragraph)
https://twitter.com/projectphysx/sta...623746?lang=en

[moebius]

Originally Posted by frmky

But to the user they appear as two distinct GPUs with separate memory spaces...

In any case it would be interesting to know if these two discrete GPU's are as fast in wavefront prp-test as on the card with only one discrete GPU, then they would be at least twice as good for the project.

[axn]

Originally Posted by moebius

In any case it would be interesting to know if these two discrete GPU's are as fast in wavefront prp-test as on the card with only one discrete GPU, then they would be at least twice as good for the project.

These beasts (Instincts, Teslas, Quadros, etc.) are only of theoretical interest to the project. They are f***ing expensive! You're better off building one or more mulit-GPU PCs with that kind of money.

[frmky]

32GB MI60's are available for under $1000. But they are passively cooled so you'd have to deal with rigging a cooler for it.