Talos II Lite

Ignore the 4-core and 8-core options. But they seem to exist as a potential cheap option to draw in some eyes (maybe??). But they're not worth it IMO.

But the 18-core Power9 intrigues me and seems very well priced. Threadripper 1950x is 16-cores / quad-memory channel for $900 (sometimes $699), but is way "smaller" than Power9 18-core / 72 threads. And holy moly, is that 10MB L3 cache per pair of chips??

So really, the 18-core Power9 compares favorably to the 24-core / 48-thread 7401P EPYC (8-channel RAM / 64 PCIe lanes).

Intel's 14-core i9-7940x is straight up inferior (similar to Threadripper). 4x channel RAM and way smaller core-counts. Xeon Gold 6150 is comparable but is lol-wut $3200 per chip.

-------------

If we're looking at a system of total-costs of ~$5k to ~$10k (depending on RAM / SSDs), then Power9 Talos II Lite actually looks like an extremely good option.

Only AMD's EPYC 7401p has similar price/performance, but the Power9 is a better designed chip with regards to inter-process communication and highly-threaded tasks. EPYC 7401p is a quad-NUMA design on a single socket, with relatively slow interprocess threads.

EPYC 7401p is great for lots of "server-like" tasks: web server, render farms, compile-box, and the like. But Power9's superior thread-communication architecture means you get both high-cores and high-interprocess communication. No infinity fabric or NUMA nodes slowing you down.

This seems like an incredibly good deal in my book. At least, if I was in the market for $5000 to $10,000 computers. If you're running a singular big application that needs superior inter-process communications, only Power9 or Xeon's mesh will do. Xeon is 3x more expensive, while EPYC has those infinity fabric slowdown issue (although EPYC is still cheaper).

EDIT: I double-checked the Power9 architecture and it seems like its very similar to AMD's Zen design. There are 2-cores per unit, and those cores share a 10MB L3 cache (instead of AMD's 4-cores per CCX sharing 2MB L3 cache). So I guess my assertion about "superior threading" requires more discussion (Afterall, IBM's "internal fabric" looks very similar to AMD's infinity fabric)

EDIT: Power's architecture / ISA has one advantage: and that is it has a looser interpretation of "cache coherency" than AMD / Intel. So the caches are more free to operate more independently. Optimal Power9 coding under this "looser" definition requires mastery of C / C++'s memory_order_consume mechanics (lol @ all 5 people in the world who know how to use those). But Acquire / Release semantics will still be faster on Power9 than on Intel/AMD, due to the looser definition of "cache coherency" on Power9 systems.

So I'd imagine (although this assumption is untested), that Power9 has superior multithreaded performance. Assuming proper use of relaxed / acquire / release / consume mechanics, which can speed up inter-thread communications on Power9 systems.

Of course, if you have bad programmers who use "global locks" or whatever, Power9 communications would not be able to operate very quickly under those circumstances. You'd likely need to really take advantage of those memory-consistency flags in C++.

The BIG advantage to Power9 is the massive L3 cache. Utterly, utterly huge. The 18-core Power9 chip has 90MB of L3 for example. With 8x RAM channels at only $1300. With 4xSMT, that's 72-threads.

Compare to Intel's Xeon Gold 6150 $3000 chip, which only has 24.75 MB of RAM and only 6x RAM channels. Intel only supports 2xSMT for 36-threads.