Can someone explain to me how such a benchmark can possibly exist?
Lets go over the basic:
>7700k doing well
ST bound test, this is obvious
>1800X doing well
Now it being ST bound doesn't seem to be so sure..
>6900k doing like dogshit
What the fuck? The 1800X and 6900k are regularly neck and neck
>6900k DOING AS BAD AS BULLDOZER
WHAT IN THE FUCK?!
So, it's not a ST bench, it's not MT bench, FUCKING BULLDOZER IS EQUAL TO A BROADWELL-E THAT HAS OVER 60% HIGHER IPC AND 8 REAL CORES.
This kind of thing shouldn't exist, a bench where the 7700k, 1800x doing well but the 6900k gaping like a retard should not fucking logically EXIST.
>>59926597
The most I can think of is that this benchmark scales only up to 5 cores but the Broadwell has compatibility issues, example the branch predictor is completely fucking up its job, or the caches are constantly being trashed, a CPU utilization test would go rather well with this bench.
Anyhow, this seems nothing more than a optimization problem, but it's awfully weird the Ryzen is the one not having problems with the optimization but the monolithic 6900k design that's been used by Intel for well over 30 years. While Ryzen looks more like a two CPUs to dumb applications
>>59926765
Not enough people testing the 6900K becuz its too fucking expensive
it's cache/memory bound.
>>59926765
Believe me when I tell you that Intel in the last 6 years has spent an enormous amount of money to get their branch predictor to not stall at these edge cases, as said by one of their engineers: "getting 95% working is easy, getting the last 5% is hard"
So color me surprised when I still see stuff like this existing, when they actually took a lot of effort into making it sure it does not exist.
Especially since this seems to be part of the SPEC suite.
And the SPEC suite is pretty much the defacto server benchmark
>>59926597
Can I use this for shitposting? BDW-E doing worse than Bulldozer will get some pretty fun reactions.
>>59926808
How does the 6900k have any problem with its caches or memory? They're both bigger than on Ryzen and the 7700k
>>59926831
I branch predicted Intel would ride their garbage arch off a cliff.
>>59926957
It might be a housefire but that will change, give it time.
Even Intel is allowed mistakes
>>59926918
http://7-cpu.com/cpu/Broadwell.html
http://www.7-cpu.com/cpu/Zen.html
l3 cache access latency is almost ~20 cycles faster on zen.
>>59927322
Now that's pretty interesting.
>>59926765
>The most I can think of is that this benchmark scales only up to 5 cores but the Broadwell has compatibility issues, example the branch predictor is completely fucking up its job, or the caches are constantly being trashed, a CPU utilization test would go rather well with this bench.
I would guess branch predictor is fucking up worse, but the 9590's branch predictor is godawful and caches are even worse (higher latency at lower bandwidth) than the 6900K.
>Anyhow, this seems nothing more than a optimization problem, but it's awfully weird the Ryzen is the one not having problems with the optimization but the monolithic 6900k design that's been used by Intel for well over 30 years.
>While Ryzen looks more like a two CPUs to dumb applications
Two 4c8t cpus with a branch predictor that's phenomenally ahead of the competition. There's a slightly-greater-than-1% uptick in performance when doing repetitious instructions.
It might also have something to do with Ryzen's very large L2 and very very large L3.
>>59926597
a possibility:
- poorly threaded software with a typical working set size/access pattern that benefits from largish L2 (512 kB vs. 256 kB) but not from really big but slightly slower L3 (8 MB vs. 20 MB)
In this scenario, the high clocked 7700k wins over everything else, but the R7s do respectably just by having fewer L2 cache misses, and the 6900k's huge L3 can't quite come into play.