Post your 7-Zip v22.01 scores

[Det0x]

That sure is an interesting idea. Since the 5950X gets the best dies, you could test the limits of the architecture in terms of the highest ST/MT clocks for a specific core configuration. Then we might see what Zen 3 is realistically capable of with high-end cooling.

Zen3 6 cores / 12 threads "5600x" -> "~15.9 average gflops per core"

Compressing: 90.191 gflops
Decompressing: 100.974 gflops
Average: 95.582 flops


Zen3 8 cores / 16 threads "5800x" -> "~15.35 average gflops per core"

Compressing: 115.747 gflops
Decompressing: 129.894 gflops
Average: 122.821 flops


Zen3 12 cores / 24 threads "5900x" -> "~14.36 average gflops per core"

Compressing: 154.536 gflops
Decompressing: 190.556 gflops
Average: 172.396 flops


Zen3 16 cores / 32 threads 5950x -> "~13.53 average gflops per core"

Compressing: 182.202 gflops
Decompressing: 250.876 gflops
Average: 216.539 flops



Conclusion: as core counts go up, each core get less and less memory bandwidth resulting in lower throughput per core.

 

[P4-630]

Did a 3d re-run , sorry @QuietBob , last one.

119.2
123.7
121.8

 

[glnn_23]

12900k

164.627
168.830
166.729

 

[1100R]

12900k

164.627
168.830
166.729

View attachment 253747

Awesome DRAM frequency!

 

[InVasMani]

Zen3 6 cores / 12 threads "5600x" -> "~15.9 average gflops per core"

Compressing: 90.191 gflops
Decompressing: 100.974 gflops
Average: 95.582 flops
View attachment 253734

Zen3 8 cores / 16 threads "5800x" -> "~15.35 average gflops per core"

Compressing: 115.747 gflops
Decompressing: 129.894 gflops
Average: 122.821 flops
View attachment 253735

Zen3 12 cores / 24 threads "5900x" -> "~14.36 average gflops per core"

Compressing: 154.536 gflops
Decompressing: 190.556 gflops
Average: 172.396 flops
View attachment 253737

Zen3 16 cores / 32 threads 5950x -> "~13.53 average gflops per core"

Compressing: 182.202 gflops
Decompressing: 250.876 gflops
Average: 216.539 flops

View attachment 253738

Conclusion: as core counts go up, each core get less and less memory bandwidth resulting in lower throughput per core.

Disable SMT and test and compare for 16C/16T, 12C/12T, 8C/8T, and 6C/6T as well. I think 16C/16T and 12C/12T would be neat to compare how that stacks up against 8C/16T and 12C/24T.

 

[xrror]

yeay P5Kpro still lives!

Ryzen 1600 AE ("normal" OG Summit Ridge)

FX8310 - sadly mem controller speed is low because this motherboard has issues overclocking (gigabyte 990FXA-UD5)

last one - stock i7-2600 (it's stuck in a Dell Optiplex 990)

 

[QuietBob]

I moved most of the entries from the bottom to a new category While we could discuss what constitutes an "old-timer", I had to draw a line somewhere. This way we'll have more space for new submissions in the main chart.

 

[agent_x007]

@QuietBob Personally, I think everything without AVX2 support should be on separate spreadsheet/table [Haswell and Ryzen 1000 being oldest platforms thrown into main table]. Unless it's better to do it based on cores/threads [4 Core/8 Thread Ryzen and Skylake problem on "old-timers") or IMC (Integrated Memory Controller) presence inside CPU [Nehalem being slowest modern CPU] ?
There are many ways to call it too (like "Obselete" or "Retro table" are also an option.

 

[Det0x]

Disable SMT and test and compare for 16C/16T, 12C/12T, 8C/8T, and 6C/6T as well. I think 16C/16T and 12C/12T would be neat to compare how that stacks up against 8C/16T and 12C/24T.

Long story short, this benchmark loves threads/SMT

Zen3 6 cores / 6 threads -> "~11 average gflops per core" --> "~11 average gflops per thread"

• Compressing: 68.331 gflops
• Decompressing: 64.536 gflops
• Average: 66.434 flops

Zen3 8 cores / 8 threads -> "~10.9 average gflops per core" --> "~10.9 average gflops per thread"

• Compressing: 89.050 gflops
• Decompressing: 84.605 gflops
• Average: 87.139 flops

Zen3 6 cores / 12 threads "5600x" -> "~15.9 average gflops per core" --> "~7.9 average gflops per thread" = +43% scaling with SMT vs 6/6

• Compressing: 90.191 gflops
• Decompressing: 100.974 gflops
• Average: 95.582 flops

Zen3 12 cores / 12 threads -> "~10.2 average gflops per core" --> "~10.2 average gflops per thread"

• Compressing: 124.254 gflops
• Decompressing: 121.604 gflops
• Average: 122.929 flops

Zen3 8 cores / 16 threads "5800x" -> "~15.35 average gflops per core" --> "~7.7 average gflops per thread" = +40% scaling with SMT vs 8/8

• Compressing: 115.747 gflops
• Decompressing: 129.894 gflops
• Average: 122.821 flops

Zen3 16 cores / 16 threads -> "~10 average gflops per core" --> "~10 average gflops per thread"

• Compressing: 161.150 gflops
• Decompressing: 160.360 gflops
• Average: 160.755 flops

Zen3 12 cores / 24 threads "5900x" -> "~14.36 average gflops per core" --> "~7.1 average gflops per thread" = +40% scaling with SMT vs 12/12

• Compressing: 154.536 gflops
• Decompressing: 190.556 gflops
• Average: 172.396 flops

Zen3 16 cores / 32 threads 5950x -> "~13.53 average gflops per core" --> "~6.7 average gflops per thread" = +34% scaling with SMT vs 16/16 (must be bandwidth limited here)

• Compressing: 182.202 gflops
• Decompressing: 250.876 gflops
• Average: 216.539 flops

Pretty much ~5ghz clockspeed on every run

 

[InVasMani]

Seems that SMT has a very heavily impact on decompression though it's more muted for compression. A SMT chip that strays from conventional 1:2 ratio that's higher one more like 1:3 or 1:4 could would really make a lot of sense to speed up storage performance in regard to decompression. Luckily compression itself is far less concerning to the end user too from a performance narrative.

The PBO seems to like the 1T frequency a bit more on the 8C/8T and 12C/12T too, but for some reason not so much on the 6C/6T oddly while the 16C/16T is probably more thermal constrained. Perhaps the 6C/6T results were more a case of benchmark inconsistency I would've thought it would boost a little higher than the 8C/8T though seems to.

 

[AlwaysHope]

snip
...
Conclusion: as core counts go up, each core get less and less memory bandwidth resulting in lower throughput per core.

I have always thought that as soon as it went beyond dual cores & dual channel memory systems (either DDR or DDR2) but didn't have the proof to back it up. Yet here we are in 2022 & 7-zip benchmark shows that though

On another note; behold the POWER of Phenom II X6 1055T!
26.5
37.8
32.2

 

[QuietBob]

EDIT:
Added Top Ten chart for competitive benchers

I think everything without AVX2 support should be on separate spreadsheet/table [Haswell and Ryzen 1000 being oldest platforms thrown into main table].

That makes a lot of sense. Thanks for the tip, charts updated accordingly!

Long story short, this benchmark loves threads/SMT

Thanks for sharing your findings. It seems that the gains MT provides for compression are diminishing as the number of threads goes up. It's +32% for 6c/12t vs. 6c/6t but only +13% for 16c/32t vs. 16c/16t. However the MT benefit for decompression is more or less the same +57% throughout.

I have always thought that as soon as it went beyond dual cores

It's generally assumed that you need 1 GB of RAM per thread - at minimum - to take advantage of multiple threads, at least in professional software.

 

[Hardi]

well let's add another 5950X to the mix

163,2
238
200,6

 

[Canned Noodles]

I still have yet to play with the memory timings... for now I'm sticking to the stock XMP ones.

 

[Blaeza]

I'm back! With oc'd ram, lol.

61.3
75.0
68.1

 

[glnn_23]

Intel i7 3610QM

22.725
24.195
23.460

 

[freeagent]

Another run..

149.85
188.59
169.22

 

[Blaeza]

62.8
77.4
70.1
my final entry!

 

[68Olds]

I have another Retro sub. I'm surprised how close it is to my old i5 with its much faster DDR3 RAM.
Core 2 Quad Q9560
2x2GB DDR2 @ 5-5-5 533MHz
13.5, 16.5, 15.0

 

[jesdals]

Damn its hard to carve these scores down for some reson my decompresson went down

so close

and when decompression goes up compression goes down - sigh

 

[Canned Noodles]

I finally got around to fiddling with my timings! I had tried raising the FClk a bit, but it wouldn't boot, and I don't feel comfortable raising the voltage manually. I also tried out 3600MHz just to see what the results would be like with synchronized MClk and Fclk. That gave me an improvement, but I got a further improvement with altering the timings at 4000MHz. The previous results are 6 or 7 messages above this one.

 

[QuietBob]

Charts updated and it looks like we have a new winner! Congratulations @Canned Noodles
And the consolation prize goes to @QuietBob with the slowest 64-bit AMD desktop CPU

Sempron 2500+ from 2005 with 1400@1862 MHz clock and running single channel DDR1:

 

[izy]

Edit: I saw i had v21.07 but its same in v22.

 

[xrror]

1600AF @ 4Ghz - happened to get another 1600AF and... it's a lot better clocker than my other one! Still testing and tweaking but so far it's been so much nicer than my other AF!

 

[Canned Noodles]

wanted to see what results I would get with a dual Xeon system