if you’re using a simplistic CPU benchmark to evaluate a server or workstation, you’re making a mistake, or at least probably shouldn’t be considering a server or workstation.
Server-class CPUs like Xeon and EPYC are not optimized for peak single-thread performance, in part because that’s not what the target market runs, and because there are many points in CPU and system design where the trade offs are between single-core performance, multi-core performance, and reliability.
Memory bandwidth vs. latency is an example. A bigger cache increases latency while also increasing bandwidth, but only for applications which can take advantage of the larger cache. Similarly, more memory controllers increase bandwidth at the expense of additional on-chip routing which adds latency. For memory-bound workloads, that’s an easy choice—and GPUs require a lot of both memory and PCIe bandwidth. (At least, for work involving lots of data—but not games or small benchmarks.)
There are more specialized benchmarks suitable for evaluating servers and workstations, for instance those published by SPEC. Obviously the best benchmark is your own application, if you can find a system to test on. Key software companies have likely been optimizing their code for some months already, but probably have more months of optimization ahead, and smaller developers may not have had early access at all. It will take some time before enough users have their hands on the Mac Pro to evaluate how well it serves their needs.
As a software developer, I know the Mac Pro is not for me: compilation doesn’t tax the CPU or memory subsystem enough. An iMac Pro is much cheaper and will serve just as well, and even the Core i9 is fine if I don’t mind the lack of ECC RAM. But if I were working with high resolution video, and had the need to apply filters, transcoding, color adjustments, etc. then the memory bandwidth in the Mac Pro could make a huge difference. If I were training AI models, it’d work well, but I might prefer a more traditional server with multiple Xeons. For photography? I’d have to measure it, or learn from others doing so.
A side note: Desktop applications, and even some server applications, tend not to scale well to multiple sockets; NUMA introduces varying latencies which are hard to work around, so it’s preferable to stick to one socket. AMD‘s first EPYC suffered from a NUMA design as well, one reason why it wasn’t adopted more broadly.
(I don’t work for Apple, but I do work for a large computer manufacturer. I’m speaking purely for myself here.)
