Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

AMD Ryzen Threadripper 3000 32-core CPU is more bad news for Intel

zdnet.com

Post reply on HN
Print view View original

101–110 of 165 posts

Re: AMD Ryzen Threadripper 3000 32-core CPU is more bad news for Intel

#101

Post by JohnJamesRambo »

eyegor wrote: post #92
zelon88 wrote: post #87

Earlier quoted context omitted.

I think most server operators look at overall performance. Once you start buying hardware specifically for one purpose you're cornering yourself. Besides, who spends $15,000 on a mid-high end server to run single threaded applications anyway?

I happen to know of several companies doing physics problems that scale poorly across cores that spend far north of that, usually building out small clusters. Then you run 100s of independent simulations since each individual one doesn't really scale.

What are the physics problems?

Re: AMD Ryzen Threadripper 3000 32-core CPU is more bad news for Intel

#102

Post by robocat »

DuskStar wrote: post #89
gameswithgo wrote: post #49

Earlier quoted context omitted.

the clock rate has nothing to do with a bad implementation, all the computation just makes a lot of heat. the performance increase is still massive, often more than 2x avx, which also throttles btw

The clock rate issue isn't the fact that it downclocks a core when moving to AVX512, it's that it downclocked all the other cores on the processor at the same time.

From what I have read, AVX512 only affects the one core (downclocking License level L1 or L2), it is older CPUs with AVX2 that affected all cores.

Independently thermal throttling can occur which would affect all cores, although presumably the CPU is generating heat per numeric operation, so AVX512 is neutral versus other instructions per numeric operation.

Re: AMD Ryzen Threadripper 3000 32-core CPU is more bad news for Intel

#103

Post by qwerty456127 »

kllrnohj wrote: post #58
qwerty456127 wrote: post #52

I'd rather choose really big caches per core than really big number of cores.

Zen2 is 16MB L3 per ccx across the entire stack (hence why even the Ryzen 5 3600 is 32MB of L3 - it has 2 CCXs). There's no reason to assume this won't be true on Threadripper 3000. So a 32-core Threadripper 3000 would have at least 128MB of L3.

L2 cache is what matters. Zen2 has just 512 KiB per core. My old Pentium-M had 2MB. I wish modern CPUs had at least this much.

Re: AMD Ryzen Threadripper 3000 32-core CPU is more bad news for Intel

#104

Post by undersuit »

kllrnohj wrote: post #58
qwerty456127 wrote: post #52

I'd rather choose really big caches per core than really big number of cores.

Zen2 is 16MB L3 per ccx across the entire stack (hence why even the Ryzen 5 3600 is 32MB of L3 - it has 2 CCXs). There's no reason to assume this won't be true on Threadripper 3000. So a 32-core Threadripper 3000 would have at least 128MB of L3.

Top end EPYC has 256MB of L3 cache, I still have a Debian 8 Netboot usb drive of the same capacity.

Re: AMD Ryzen Threadripper 3000 32-core CPU is more bad news for Intel

#106

Post by gameswithgo »

eyegor wrote: post #92
zelon88 wrote: post #87

Earlier quoted context omitted.

I think most server operators look at overall performance. Once you start buying hardware specifically for one purpose you're cornering yourself. Besides, who spends $15,000 on a mid-high end server to run single threaded applications anyway?

I happen to know of several companies doing physics problems that scale poorly across cores that spend far north of that, usually building out small clusters. Then you run 100s of independent simulations since each individual one doesn't really scale.

simulations that don't scale, that do scale after all.

Re: AMD Ryzen Threadripper 3000 32-core CPU is more bad news for Intel

#107

Post by gameswithgo »

DuskStar wrote: post #89
gameswithgo wrote: post #49

Earlier quoted context omitted.

the clock rate has nothing to do with a bad implementation, all the computation just makes a lot of heat. the performance increase is still massive, often more than 2x avx, which also throttles btw

The clock rate issue isn't the fact that it downclocks a core when moving to AVX512, it's that it downclocked all the other cores on the processor at the same time.

i understand why downclocks are an issue, and i understand that on some intel cpus the while chip downclocks with certain instructions. i was commenting on the supposed reason the down clocks happen, and mentioning that performance is spectacular despite then (assuming you schedule your workload appropriately).

Re: AMD Ryzen Threadripper 3000 32-core CPU is more bad news for Intel

#108

Post by greglindahl »

gameswithgo wrote: post #48
vj44 wrote: post #24

Earlier quoted context omitted.

I work in this space... and let's just say that MKL is definitely NOT well optimized for AMD's chips. You'll be lucky to get 10-20% efficiency. Nevermind openblas.

does this remain true on the zen2 cpus which finally do avx properly?

Intel is famous for checking for 'IntelInside' instead of cpu feature bits, and taking a generic and slow code path if it's not IntelInside.

Re: AMD Ryzen Threadripper 3000 32-core CPU is more bad news for Intel

#109

Post by gameswithgo »

robocat wrote: post #102
DuskStar wrote: post #89

Earlier quoted context omitted.

The clock rate issue isn't the fact that it downclocks a core when moving to AVX512, it's that it downclocked all the other cores on the processor at the same time.

From what I have read, AVX512 only affects the one core (downclocking License level L1 or L2), it is older CPUs with AVX2 that affected all cores. Independently thermal throttling can occur which would affect all cores, although presumably the CPU is generating heat per numeric operation, so AVX512 is neutral versus other instructions per numeric operation.

on intel cpus the license levels basically are discrete thermal throttles. vs and which doesn’t do that, just constantly monitors thermals and adjusts clock.

intels method makes benchmarking simpler! but may leave performance on the table.

Re: AMD Ryzen Threadripper 3000 32-core CPU is more bad news for Intel

#110

Post by eyegor »

gameswithgo wrote: post #106
eyegor wrote: post #92

Earlier quoted context omitted.

I happen to know of several companies doing physics problems that scale poorly across cores that spend far north of that, usually building out small clusters. Then you run 100s of independent simulations since each individual one doesn't really scale.

simulations that don't scale, that do scale after all.

If you spawn 100 independent instances, it's not really the problem itself scaling. The point is that given a single set of operating conditions you won't see any meaningful gains going from 2 to 100 cores. Using idle resources for other simulations doesn't make the problem itself scale.
Post reply on HN