At this price I would rather buy a Raptor TALOS II or Blackbird.

With Talos II, you'd get a workstation with only 4 cores at this price. Blackbird board + 8-core CPU bundle is a bit better, but now you're limited by the small form factor.

I'm not sure what performance looks like on 4 POWER cores vs 32 ARM cores; seems like it might win. Genuinely curious; would love to see benchmarks.

Those 4 POWER9 cores (16 threads with its 4-way SMT) will obliterate this ARM. Depending on the benchmark, this 32 core ARM box will do 1/2 the single thread performance and about 2.5x the multithreaded performance of my venerable Sandy Bridge i7 desktop (so it does 2.5x the MT performance with 8x the amount of physical cores). This is a hardware class from 2011, which you can usually pick up for under $250 from eBay in second-hand corporate desktops. On the other hand, the POWER9 in the Talos II can go toe-to-toe with Intel's Skylake in both single thread and especially multithreaded use cases.

The reason to buy a workstation like this is not performance, but rather being able to natively develop ARM software locally. That's the only real killer feature it has over Intel/AMD/IBM processors, but a legitimate reason to buy it given the availability of ARM servers at cloud providers.

https://lemire.me/blog/2019/03/26/hasty-comparison-skylark-a...

Daniel Lemire's single-threaded Mandelbrot benchmark: 15s on the Ampere eMAG (Skylark), 24s on a 4GHz Skylake. (Also wins in bitset_count.) It's definitely faster 1/2 of Sandy Bridge.

eMAG will have a disadvantage in SIMD, but for normal workloads (make -j32 on a huge project :D) it should be plenty fast.

I don't think that that Mandelbrot benchmark is single threaded. It would mean that this processor is much faster than Skylake in computational instructions per clock, which would be a huge game changer (and it clearly isn't). That 25s vs 18s figure makes a lot more sense if the benchmark is multithreaded, since Skylake is around 1.7x to 2x faster than Sandy Bridge.

Moreover, the benchmark source code [0] clearly uses OpenMP to parallelize the benchmark tasks.

[0]: https://benchmarksgame-team.pages.debian.net/benchmarksgame/...

Different benchmarks favor different processors… Here's another one, now multi-threaded

My quick run of `sysbench cpu` (something with prime numbers) shows:

Skylark 3.3GHz 32core (at Packet; Ubuntu 18.04):

    events per second: 44287.87

    events per second: 14632.61