Disabling SMT on 3950X vs 3800X

[Nathanpete]

I saw a build on PCPP a while ago where a guy bought a 3950X for his workstation and then disabled SMT. He claimed it had given him benefits. I don't remember what the benefits where but I assume it would be power consumption, frequency and probably single core performance. I had questioned his decision but didn't put a comment on his build. Clearly his workstation's task didn't benefit from SMT, but then I was wondering what are the specific things you lose and gain when you disable SMT? What is the difference in performance of a 16C/32T chip turned into 16C/16T vs a 8C/16T chip. From what I can tell, at least on Intel, the other thread uses silicon/resources within the core not being used by the other thread. For example, the FPU, if used by one thread can't be used by the accompanying thread. But that is on Intel, does each thread have a FPU and ALU in Zen 2, or do they compete just like they do in hyperthreading? (Then I remember AMD FX where there was 1 FPU for every 2 cores iirc)

 

Like do you lose access to some cache. I really wish I could read+understand the block diagrams for Zen 2 on wikichip, I even looked up the block diagrams for Coffee Lake refresh since some of the products already have half of the threads cut off, and it would prove the differences w/o testing. (9900K has 16MB Cache, and the 9700K, which doesn't have hyperthreading, had 12MB cache. Actually it says MiB, which is mebibyte, but doesn't everyone just say megabyte in this context anyway?)

Obviously Zen 2 is well known for their multi-threaded performance, so why would he choose to throw that away beyond the fact that his workload wasn't benefitting from SMT?

 

(This all assumes that we are locking the frequency at say, 4.1 GHz)

[SupaKomputa]

Disabling SMT decrease latency, somewhat beneficial for gaming or single cored apps, but for real multithreaded applications you will loose 30%.

[Nathanpete]

Just now, SupaKomputa said:

Disabling SMT decrease latency, somewhat beneficial for gaming or single cored apps, but for real multithreaded applications you will loose 30%.

I'm really just wondering about access to cache, the amount of available cache, access to resources like the ALU and FPU, and what sort of differences there are between Intel's hyper threading and AMD's SMT in Zen 2. Also curious about the differences between 9900K and 9700K, especially their cache differences and frequency differences. Although I believe the frequency differences is just about the 9900K's high bin.