The Technical Workloads Where AMD Ryzen 9 7900X3D/7950X3D CPUs Are Excellent

[coder]

Exactly its not exactly the same, in order for thr CPU without vCache to access the L3 cache of the CPU with vCache it has to go via the interconnect.

I think the main case where one compute die needs data from the L3 cache of another is when accessing shared state used for thread-communication. If the threads doing most of the communication are all scheduled on the same compute die, it's a non-issue.

[unwind-protect]

Is anyone here on Phoronix doing performance sensitive CPU inference? Big Blender renders on CPU? CPU CFD? And are you doing them on a $500+ CPU?

I am open to use cases, but AFAIK those workloads are not sane in 2023, which is something Michael should consider.

For molecular dynamics there is the Amber toolkit. Their GPU acceleration is CUDA only. What is worse is that the CUDA support is a re-implementation of a CPU program, and it does not support the full feature set of the CPU version. So for some calculations you fall back to CPU.

I scheduled these jobs on AWS machines, but the virtual farm also had small local machines (AM4 with ECC).

[brucethemoose]

For molecular dynamics there is the Amber toolkit. Their GPU acceleration is CUDA only. What is worse is that the CUDA support is a re-implementation of a CPU program, and it does not support the full feature set of the CPU version. So for some calculations you fall back to CPU.

I scheduled these jobs on AWS machines, but the virtual farm also had small local machines (AM4 with ECC).

Couldnt yall run it with ROCM? I am sympathetic to being unwillingly stuck with CUDA though.

coder's use case is also interesting, as I figured appliance type devices used ARM SoCs with small integrated accelerators, but IDK anything about that world.

[coder]

coder's use case is also interesting, as I figured appliance type devices used ARM SoCs with small integrated accelerators, but IDK anything about that world.

I use the word "appliance" to mean a rackmount PC that ships with specialized software pre-installed. To end customers, they treat it as a fixed-function device, but it's really just a fairly vanilla PC with specs & components installed that enable it to fulfill its intended purpose.

[Ricard]

In some tests of this article, 7950x has less performance than in previous test, for example (but not the only one) the ONNX Runtime 1.14 Model: fcn-resnet101-11. In 7950x3d’s prelaunch review, 7950x has 3.64 vs 2.22 in this article.

3.64 are more inferences per second than the achieved by 7950x3d in this article’s test which are 3.53.

Why this results?