Zen4 7950X Benchmarks

Mysticial · 2022-09-27, 06:50

Chip is running at stock. Memory is 4 x 16GB @ 4400 MT/s. So the memory is actually quite slow here.

Have results for both PBO on (up to 5.7 GHz) and PBO off (4.5 GHz) since I forgot to turn it on the first time.

Overall, the difference looks like a complete wash since it's memory bound. (which was expected)

Mysticial · 2022-09-27, 07:42

And the 1k benchmark: (with PBO enabled)

Code:

Timings for 1K FFT length (16 cores, 1 worker):  0.00 ms.  Throughput: 1040087.82 iter/sec.
Timings for 1K FFT length (16 cores, 2 workers):  0.00,  0.00 ms.  Throughput: 1644672.53 iter/sec.
Timings for 1K FFT length (16 cores, 4 workers):  0.00,  0.00,  0.00,  0.00 ms.  Throughput: 3938231.64 iter/sec.
Timings for 1K FFT length (16 cores, 16 workers):  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00 ms.  Throughput: 13322808.17 iter/sec.
Timings for 1K FFT length (16 cores hyperthreaded, 1 worker):  0.00 ms.  Throughput: 1037063.16 iter/sec.
Timings for 1K FFT length (16 cores hyperthreaded, 2 workers):  0.00,  0.00 ms.  Throughput: 1663182.71 iter/sec.
Timings for 1K FFT length (16 cores hyperthreaded, 4 workers):  0.00,  0.00,  0.00,  0.00 ms.  Throughput: 3388185.67 iter/sec.
Timings for 1K FFT length (16 cores hyperthreaded, 16 workers):  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00,  0.00 ms.  Throughput: 13692992.65 iter/sec.

mackerel · 2022-09-27, 07:44

I'd echo a request to benchmark at smaller FFTs. For my personal interests 64k-1M would be most interesting.

I'm part way through watching the replay of Ian Cutress' stream on the launch. One point I found interesting is he stated Infinity Fabric links are same bandwidth as for AM4, although they made it half as wide but twice as fast as apparently that allowed some power saving. This is particularly interesting since IF on AM4 was matched to DDR4 bandwidth to a single CCD (full bandwidth reads, half bandwidth on writes). As such I don't expect single CCD (6-8 core) models to be able to make use of the increased DDR5 bandwidth. Two CCD models can still make use of the higher bandwidth, although if they kept writes at half the read rate then it may not be possible to get full speed writes to ram even with two CCDs. I'll have to check other reviews and see if that was the case.

Ipmanchess · 2022-09-27, 07:58

Hi,

Are you interested to run some Chess-benchmarks on AMD R9 7950X!

It's very simple..just download Stockfish 14.1 pop ,avx2 ,bmi2 & avx512 version from my website:
https://ipmanchess.yolasite.com/amd-...-stockfish.php
And dubble click them one by one to run the bench like this in console,type:

bench 1024 32 26 default depth nnue

32 = cores/threads your cpu have (in case you want to use it on other systems)

You will see scrolling lines till it ends and looks like this: (numbers will be different)

===========================
Total time (ms) : 143344
Nodes searched : 6765552384
Nodes/second : 47198015

Takes a minute or so to finish the bench..and if possible to add the logs from these benches into a txt file or at least these 3 result lines as above from each bench you run! (When one bench is done ,close it before you start next one)

Nodes/second is the result i use to put in my list as ranking.
When you check my list then you know what info i would like to add to make the bench complete like others did.

Thanks in advance!
Ipman.

mackerel · 2022-09-27, 08:17

To partially answer myself on previous post, it seems like the IF links are more different than with AM4. Read bandwidths for single CCD Zen 4 are somewhat lower than two CCD, which is kind of expected. Write bandwidth is much higher. Definitely not the half speeds of AM4! If anything it has to be faster overall.

mackerel · 2022-09-27, 09:59

Using the benchmarks results here I've compared it to my 11700k and 7920X.

1k FFT, n core, n workers

0.86/core - 13.7M 16 cores 7950X
0.74/core - 5.9M 8 cores 11700k - peak power observed 195W
0.47/core - 5.6M 12 cores 7920X - this defaults to power 140W limit. I can't observe actual power

The Skylake-X result is much lower than expected. I don't have an explanation for this, but my guess is that at a 1k FFT size, it isn't big enough to load the extra execution resource. Ideally I'd like to see around 128k FFT as that should load the cores well while remaining in cache.

8064k n cores 1 worker - 1 task using all cores.

318 7950X - dual channel 4400
258 7920X - quad channel 3000 (2R per channel)
95 11700k - dual channel 2166 (4R per channel)

The 7950X is punching a good amount above the 7920X, presumably due to the cache.

8064k n cores n workers (1 task per core)

180 7950X
250 7920X
94 11700k

Ok, now we're definitely pushing it hard into ram bandwidth limiting, and the 7950X drops below the 7920X as expected. My Skylake-X system has 39% more potential bandwidth, and scores 36% higher, so this seems about right. Also 7950X vs 11700k is roughly 2x, as expected from the ram speed.

joblack · 2022-09-27, 10:49

Nice one. Could you please test the prime number times mentioned in

https://mersenneforum.org/showthread.php?t=13185

M344587487 · 2022-09-27, 11:35

Quote:

Originally Posted by joblack

Nice one. Could you please test the prime number times mentioned in

https://mersenneforum.org/showthread.php?t=13185

Data is good but it might be better to wait for someone to test with faster RAM as that's what'll typically be used, AMD officially supports 5200 and reviewers seem to be going for 6000 or even 6400. I can't even buy 4400 DDR5 from my normal retailer, the speeds on offer range from 4800 to 6400.

henryzz · 2022-09-27, 13:20

The really interesting cpus will be the ones with 3D cache when they come out. Rumors online seem to suggest early next year might be a timeframe for these(annoucement at CES which is Jan 5-8)

Mysticial · 2022-09-27, 15:01

Quote:

Originally Posted by M344587487

Data is good but it might be better to wait for someone to test with faster RAM as that's what'll typically be used, AMD officially supports 5200 and reviewers seem to be going for 6000 or even 6400. I can't even buy 4400 DDR5 from my normal retailer, the speeds on offer range from 4800 to 6400.

The memory is rated for 4800, but the mobo is running it at 4400. But as I mentioned in the other thread, I can't access the BIOS to tinker with anything.

AFAICT, 4400 is considered an overclock since 3600 is the rated speed for 2DPC. AMD sent me the system with 4 x 16GB pre-populated. And since I never touched Alder Lake with a 10ft pole, I don't have any other DDR5 to try.

mackerel · 2022-10-05, 13:14

From https://www.primegrid.com/forum_thre...ap=true#157312

Quote:

Hi my 7950 runs CPU ppt limit 140w =190w at wall.
Prime95 64-bit version 30.8, RdtscTiming=1
Timings for 240K FFT length (8 cores, 8 workers): 0.37, 0.37, 0.37, 0.37, 0.37, 0.37, 0.37, 0.37 ms. Throughput: 21517.66 iter/sec.
Timings for 240K FFT length (16 cores, 16 workers): 0.43, 0.42, 0.42, 0.43, 0.42, 0.42, 0.42, 0.42, 0.43, 0.43, 0.43, 0.44, 0.43, 0.43, 0.43, 0.43 ms. Throughput: 37418.08 iter/sec.

Without knowing the exact running clocks, I estimate from those results the effective AVX-512 IPC (for Prime95 and similar uses when not memory bound) of Zen 4 sits between Rocket Lake and Skylake-X. I was thinking of selling my Rocket Lake system for other reasons anyway, but I'll give Zen 4 some time to mature before considering it as a replacement.

2022-09-27, 07:44	#3
mackerel Feb 2016 UK 1BF₁₆ Posts	I'd echo a request to benchmark at smaller FFTs. For my personal interests 64k-1M would be most interesting. I'm part way through watching the replay of Ian Cutress' stream on the launch. One point I found interesting is he stated Infinity Fabric links are same bandwidth as for AM4, although they made it half as wide but twice as fast as apparently that allowed some power saving. This is particularly interesting since IF on AM4 was matched to DDR4 bandwidth to a single CCD (full bandwidth reads, half bandwidth on writes). As such I don't expect single CCD (6-8 core) models to be able to make use of the increased DDR5 bandwidth. Two CCD models can still make use of the higher bandwidth, although if they kept writes at half the read rate then it may not be possible to get full speed writes to ram even with two CCDs. I'll have to check other reviews and see if that was the case. Last fiddled with by mackerel on 2022-09-27 at 07:44

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Zen4's AVX512 Teardown	Mysticial	Hardware	21	2022-11-07 01:10
AVX512 and Zen4 pre-release speculations	Xyzzy	y-cruncher	12	2022-09-15 17:15
RPS benchmarks	pinhodecarlos	Riesel Prime Search	29	2014-12-07 07:13
Benchmarks for 24.12	Prime95	Software	60	2005-06-11 07:35
Benchmarks	Vandy	Hardware	6	2002-10-28 13:45

2022-09-27, 07:58	#4
Ipmanchess Sep 2022 1₈ Posts	Hi, Are you interested to run some Chess-benchmarks on AMD R9 7950X! It's very simple..just download Stockfish 14.1 pop ,avx2 ,bmi2 & avx512 version from my website: https://ipmanchess.yolasite.com/amd-...-stockfish.php And dubble click them one by one to run the bench like this in console,type: bench 1024 32 26 default depth nnue 32 = cores/threads your cpu have (in case you want to use it on other systems) You will see scrolling lines till it ends and looks like this: (numbers will be different) =========================== Total time (ms) : 143344 Nodes searched : 6765552384 Nodes/second : 47198015 Takes a minute or so to finish the bench..and if possible to add the logs from these benches into a txt file or at least these 3 result lines as above from each bench you run! (When one bench is done ,close it before you start next one) Nodes/second is the result i use to put in my list as ranking. When you check my list then you know what info i would like to add to make the bench complete like others did. Thanks in advance! Ipman.

2022-09-27, 08:17	#5
mackerel Feb 2016 UK 3·149 Posts	To partially answer myself on previous post, it seems like the IF links are more different than with AM4. Read bandwidths for single CCD Zen 4 are somewhat lower than two CCD, which is kind of expected. Write bandwidth is much higher. Definitely not the half speeds of AM4! If anything it has to be faster overall.

2022-09-27, 09:59	#6
mackerel Feb 2016 UK 3×149 Posts	Using the benchmarks results here I've compared it to my 11700k and 7920X. 1k FFT, n core, n workers 0.86/core - 13.7M 16 cores 7950X 0.74/core - 5.9M 8 cores 11700k - peak power observed 195W 0.47/core - 5.6M 12 cores 7920X - this defaults to power 140W limit. I can't observe actual power The Skylake-X result is much lower than expected. I don't have an explanation for this, but my guess is that at a 1k FFT size, it isn't big enough to load the extra execution resource. Ideally I'd like to see around 128k FFT as that should load the cores well while remaining in cache. 8064k n cores 1 worker - 1 task using all cores. 318 7950X - dual channel 4400 258 7920X - quad channel 3000 (2R per channel) 95 11700k - dual channel 2166 (4R per channel) The 7950X is punching a good amount above the 7920X, presumably due to the cache. 8064k n cores n workers (1 task per core) 180 7950X 250 7920X 94 11700k Ok, now we're definitely pushing it hard into ram bandwidth limiting, and the 7950X drops below the 7920X as expected. My Skylake-X system has 39% more potential bandwidth, and scores 36% higher, so this seems about right. Also 7950X vs 11700k is roughly 2x, as expected from the ram speed.

2022-09-27, 10:49	#7
joblack Oct 2008 n00bville 2⁵×23 Posts	Nice one. Could you please test the prime number times mentioned in https://mersenneforum.org/showthread.php?t=13185

2022-09-27, 13:20	#9
henryzz Just call me Henry "David" Sep 2007 Liverpool (GMT/BST) 37·163 Posts	The really interesting cpus will be the ones with 3D cache when they come out. Rumors online seem to suggest early next year might be a timeframe for these(annoucement at CES which is Jan 5-8)