AMD Instinct MI300A data center APU underperforms against mainstream CPUs in Geekbench — MI300A submissions show lower performance than a Core i5-14600K

Posted by BarKnight@reddit | hardware | View on Reddit | 18 comments

[-]

SmashStrider@reddit

The main power is gonna be coming from the MI300 GPU, I don't really see what's the problem here. CPU performance of a 14600K is more than enough to support the GPU, considering that the i5-14600K is already a really powerful processor.

[-]

ttkciar@reddit

I think perhaps the author doesn't understand what the MI300A is.

The CPU only takes up a small part of the total die space. The lion's share is taken up by the GPU, which isn't exercised by Geekbench at all.

Given that the CPU part is only really there to provision the GPU with tasks, performance comparable to i5-14600K seems like overkill.

[-]

Alarchy@reddit

CPU is important for heterogenous compute (one example: HPC/supercomputers). Grace Hopper is over 4x the multicore of MI300A, if these Geekbench scores are legit. That's a non-trivial difference, and would impact selection of AMD over Nvidia for such workloads.

[-]

Strazdas1@reddit

Its important, but its job is to manage GPU clusters, not do compute.

[-]

Alarchy@reddit

It can absolutely be a choke point if not powerful enough for the density of GPU/memory. Being 4x slower per unit than Grace Hopper is concerning.

[-]

ResponsibleJudge3172@reddit

Nvidia, who I assume know what they are doing, has once outlined why they had chosen Sapphire Rappids for Hopper. That is due to single core performance. The more the better. That's why I do not believe 14600K is overkill in anyway.

Grace CPU is not much if at all better than Sapphire Rappids but at least it has the massive advantage of bandwidth over NVLink vs PCIe. Which if I am not mistaken, AMD uses here.

[-]

reddit_equals_censor@reddit

Nvidia, who I assume know what they are doing, has once outlined why they had chosen Sapphire Rappids for Hopper. That is due to single core performance.

oh thank goodness intel is going to regain single core performance leadership eventually once royal core comes in to have high performance/core server chips available to people with such needs....

....

...

WHAT DO YOU MEAN THEY CANCELLED THE JIM KELLER ROYAL CORE PROJECT???

[-]

ttkciar@reddit

PCIe. Which if I am not mistaken, AMD uses here.

You are mistaken. The MI300A integrates the CPU and GPU. They are on the same die. There is no external bus linking them together, because they are the same part.

[-]

the_dude_that_faps@reddit

In other news, a truck loses to sports car on a drag race.

[-]

ConsistencyWelder@reddit

Tom Shardware, why are you like this?

[-]

W0LFSTEN@reddit

Posted by Tom’s Hardware writer Aaron Klotz. 😊

[-]

BlueGoliath@reddit

Is that the "Just buy it" guy?

[-]

Real-Human-1985@reddit

Proof that geekbench is not a useful benchmark.

[-]

SomeKindOfSorbet@reddit

It's a useful benchmark for consumer devices., but it doesn't scale well across multiple cores, which is usually what datacenter workloads are about. GB6 specifically cares a lot more about the speed of your cores (even for the MT score) than the number of them, which is why a datacenter CPU with 128 low-clocked cores can get beat by a consumer desktop chip that can boost much higher in low multithreading scenarios

[-]

rgknz@reddit

Maybe the author doesn't know what the MI300A is.**

CPUs occupy a small portion of die space. The GPU dominates, which Geekbench doesn't do.

Performance like i5-14600K seems overkill since the CPU only provides GPU tasks.

[-]