AMD launches the 11.5TFLOPS (FP64) Instinct MI100 accelerator

Quote:

---

Claims that it is "the world’s fastest HPC GPU accelerator for scientific workloads".

---

Read more.

But can it run Crysis?!


...(getting my coat)

WHAT?? A100 specs: FP32 Training 312 Tflops, INT8 Inference 1248 Tops & FP64 19.5 Tflops.....who is fooling who?. These data centre AI specs are becoming fuzzy.

Quote:

---

Originally Posted by lumireleon

WHAT?? A100 specs: FP32 Training 312 Tflops, INT8 Inference 1248 Tops & FP64 19.5 Tflops.....who is fooling who?. These data centre AI specs are becoming fuzzy.

---

Tflops doesn't necessarily translate between platforms/hardware, we only need to look at the ps5 and xbox series x for proof in that. It's entirely possible that in this exact field of usage that the current AMD card is better than the nvidia one.

In real terms, unless one us wins the lottery and knows how to code (or has access to them), we'll never really know or have any need to in all honesty.

edit: and which a100 are you referencing, the one with the article today with 4 cards inside or the card on it's own... they both have a100 in the name lol.

Quote:

---

Originally Posted by lumireleon

WHAT?? A100 specs: FP32 Training 312 Tflops, INT8 Inference 1248 Tops & FP64 19.5 Tflops.....who is fooling who?. These data centre AI specs are becoming fuzzy.

---

Those are the figures for the tensor cores which only have a specific compute focus whereas AMDs figures are for the general purpose compute which smokes the A100.

Quote:

---

Originally Posted by tharfcake

But can it run Crysis?!

---

Given the MI50 ended up as the Radeon VII after some tweaking, I would have to say "eventually" ;)

I am slightly surprised they don't slap a couple of DP ports on them and at least sell them as workstation video cards tbh. They would be awful value, but that doesn't seem to stop some people buying stuff.

Quote:

---

Originally Posted by DanceswithUnix

Given the MI50 ended up as the Radeon VII after some tweaking, I would have to say "eventually" ;)

I am slightly surprised they don't slap a couple of DP ports on them and at least sell them as workstation video cards tbh. They would be awful value, but that doesn't seem to stop some people buying stuff.

---

Reading the STH article, it looks like graphical output hardware is actually removed/missing/not baked into the die so it can't actually output if it wanted to...?

I think Nvidia may have been sleeping, as in so did intel....

Wouldn't it be great if you could link one of these to your upcoming 6800XT for a performance boost in gaming. Sigh.

Quote:

---

Originally Posted by LSG501

Quote:

---

Originally Posted by lumireleon

WHAT?? A100 specs: FP32 Training 312 Tflops, INT8 Inference 1248 Tops & FP64 19.5 Tflops.....who is fooling who?. These data centre AI specs are becoming fuzzy.

---

Tflops doesn't necessarily translate between platforms/hardware, we only need to look at the ps5 and xbox series x for proof in that. It's entirely possible that in this exact field of usage that the current AMD card is better than the nvidia one.

In real terms, unless one us wins the lottery and knows how to code (or has access to them), we'll never really know or have any need to in all honesty.

edit: and which a100 are you referencing, the one with the article today with 4 cards inside or the card on it's own... they both have a100 in the name lol.

---

And To make matters worse are they comparing it to the 40 or 80GB A100

Quote:

---

Originally Posted by Tabbykatze

Reading the STH article, it looks like graphical output hardware is actually removed/missing/not baked into the die so it can't actually output if it wanted to...?

---

Thanks for the pointer, that was quite an in depth article.

I find it odd that they would omit a couple of displayport outputs and yet include a multimedia engine which is labelled as being on the die. There is a possibility that the bulk of the XGMI inter card links are serdes units, and there may be a mux hidden in there to re-purpose the block or part of it to talk to a couple of monitors as they are also a big serial link block but with a different driving block.

I can understand not selling these things for gaming, though MI50 did eventually make an appearance, but Firepro cards are quite a lucrative business and it seems odd to not allow for workstation use for these things.

https://www.servethehome.com/wp-cont...ck-Diagram.jpg

Quote:

---

Originally Posted by DanceswithUnix

Thanks for the pointer, that was quite an in depth article.

I find it odd that they would omit a couple of displayport outputs and yet include a multimedia engine which is labelled as being on the die. There is a possibility that the bulk of the XGMI inter card links are serdes units, and there may be a mux hidden in there to re-purpose the block or part of it to talk to a couple of monitors as they are also a big serial link block but with a different driving block.

I can understand not selling these things for gaming, though MI50 did eventually make an appearance, but Firepro cards are quite a lucrative business and it seems odd to not allow for workstation use for these things.

https://www.servethehome.com/wp-cont...ck-Diagram.jpg

---

Tbh, I'm willing to bet bottom dollar there will be a graphic output version in the future but for now they're probably saving a a fair buck per unit not including those as I would hesitate to guess that there is a large back order of these already. Considering all the HPC wins AMD has had announced over the past 6 months which include AMD GPUs, there may be a lot of datacentre orgs that'll want to try this out

Quote:

---

Originally Posted by Tabbykatze

Tbh, I'm willing to bet bottom dollar there will be a graphic output version in the future but for now they're probably saving a a fair buck per unit not including those as I would hesitate to guess that there is a large back order of these already. Considering all the HPC wins AMD has had announced over the past 6 months which include AMD GPUs, there may be a lot of datacentre orgs that'll want to try this out

---

The thing is, even if it is only a small design change it can easily cost a million dollars to tape out a new chip like this. That's why I expect something like this to hit as many target markets as possible with just a few fuses lasered in production to set how it behaves.