Question Geekbench 6 released and calibrated against Core i7-12700

igor_kavinski · Feb 14, 2023

Geekbench

www.geekbench.com

Weird choice of baseline CPU and even weird is that the baseline score is 2500.

i7-12700 does hardly 2000 in GB5 with the fastest DDR5.

H433x0n · Dec 29, 2023

Downloaded Geekbench 6 today and this is the results I got.

The most interesting result is below:

System is not overclocked, I downclocked my ram from DDR5-6800 to DDR5-6400 to match what I was running with my old 13900K with same timings. This is the same system that I ran my 13900K in, so it's truly apples/apples. The 14900K has PL2=241W and IccMax=360A, so it's got less power than the 13900K (PL2=253W & IccMax=512A) had to work with.

igor_kavinski · Dec 29, 2023

H433x0n said:
System is not overclocked, I downclocked my ram from DDR5-6800 to DDR5-6400 to match what I was running with my old 13900K with same timings.

Pretty good. Interested to see how Det0x's and your system compare in GB 5.5

H433x0n · Dec 29, 2023

igor_kavinski said:
Pretty good. Interested to see how Det0x's and your system compare in GB 5.5

My system isn't overclocked and my memory timings are just default "XMP Tweaked" in the BIOS, so he'll definitely win the MT score. The ST score has more room to increase once I set my memory back to DDR5-6800.

igor_kavinski · Dec 29, 2023

H433x0n said:
The ST score has more room to increase once I set my memory back to DDR5-6800.

The timings would also increase a bit so I think the score shouldn't change much.

Det0x · Dec 29, 2023

igor_kavinski said:
Did you pay double for a known good sample or something?

I traded one of my 7950X3D's for this SP119 7950X with a fellow forum user from overclock.net, without any money involved.

Its a good sample but not anything super special.. (IMC could be better)

LLC 5, L3 idle temp ~20 degrees, max load ~60 temp (~185w PPT with 6000/2000 memory)
1200mv vcore set, ~1110mv get under load

Maximum stable 3x in a row Cinebench r23 allcore clocks:
7950x SP 115 (SP 120 + SP 110) = 5400/5300mhz -> 2233 PGE
7950x SP 117 (SP 121 + SP 112) = 5500/5200mhz (6200/2200) (insane CCD0, very bad CCD1)->BQ 2231 PGE
7950x SP 116 (SP 119 + SP 114) = 5450/5350mhz (6200/2100) ->2232 PGY
7950x SP 116 (SP 120 + SP 111) = 5425/5325mhz -> 2233PGE
7950x SP 115 (SP 119 + SP 110) = 5425/5275mhz (6400 maybe 6600 / 2200 -> 2240 SUY
7950x SP 114 (SP 121 + SP 108) = 5450/5250mhz with L0 cache error ->2241 SUY
7950x SP 116 (SP 119 + SP 112) = 5525/5350mhz (6600/2200) ->BQ 2239 SUY
7950x SP 117 (SP 119 + SP 115) = 5350/5275mhz (6600/2200) ->BQ 2333 SUE
7950x SP 119 (SP 120 + SP 118) = 5375/5425mhz (6600/2200) ->BQ 2245 SUY (delidded with direct die cooling)
7950x SP 116 (SP 119 + SP 112) = 5550/5200mhz (6600/2200) ->BQ 2252 SUY

Would actually say my 7800X3D and 7950X3D are better silicon quality (i'm the one with Norwegian flag)

H433x0n · Dec 29, 2023

igor_kavinski said:
The timings would also increase a bit so I think the score shouldn't change much.

You were right, the ST score didn't change at all. Surprisingly the MT score went up a tiny amount.

igor_kavinski · Dec 29, 2023

Possible to increase the MT score further with DDR5-8000: https://browser.geekbench.com/v6/cpu/3723040

Timur Born · Dec 31, 2023

I am not a big fan of Geekbench being so DDR5 memory overclock sensitive, because it doesn't reflect the realworld impact.

Doug S · Dec 31, 2023

Timur Born said:
I am not a big fan of Geekbench being so DDR5 memory overclock sensitive, because it doesn't reflect the realworld impact.

Depends on if that memory overclock kept the latency the same, or also reduced memory latency. If the latter then there is a realworld impact.

Timur Born · Dec 31, 2023

Most timings can be reduced without increasing frequency. Some need extra voltage to do so, which means that you can just increase frequency alongside then. That being said, I need to go quite out of my way to measure differences between higher frequency overclocks and my tuned 5600 CR1. At the same time I can use low voltages and be stable even at highest dimm temperatures (that are lower to begin with at low voltages). DDR5 is fast to begin with and tuned vs. untuned seems to have more of an impact than higher vs. lower frequencies, at least above 5600 MT.

My most memory intensive rendering application is Topaz Gigapixel AI, it uses 10 GB to upscale a *single* image to 25k+ pixels. Measurable difference between 5600 and 7000 MT: none. 7-Zip's compression is quite memory sensitive, between 5600 and 8200 MT you gain 15% compression performance. But compression is a niche application, DEcompression is much more common. Difference for DEcompression between 5600 and 8200 MT: none.

mikegg · Jan 3, 2024

H433x0n said:
Downloaded Geekbench 6 today and this is the results I got.

View attachment 91171

The most interesting result is below:

View attachment 91172

System is not overclocked, I downclocked my ram from DDR5-6800 to DDR5-6400 to match what I was running with my old 13900K with same timings. This is the same system that I ran my 13900K in, so it's truly apples/apples. The 14900K has PL2=241W and IccMax=360A, so it's got less power than the 13900K (PL2=253W & IccMax=512A) had to work with.

Can you post package power during the duration of the test?

Timur Born · Jan 3, 2024

GB6 doesn't hit close to power-limits, though, so that doesn't really matter. GB6's multicore tests are just too memory bandwidth dependent. As a result we see over 6% multi-core score improvements for only a 3.6% core ratio increase.

This is my undervolted 13900K at 60x single-core and 55x multi-core using 5600 MT CR1:

Side-note: very few parts of GB6 trigger the AVX offset, something to keep in mind when comparing numbers. I disabled the AVX offset for the above result for easier comparison, but usually use a -1 offset that affects single-core load and affects 2 out of 8 P cores for all-core load.

ikjadoon · Jan 16, 2024

Timur Born said:
I am not a big fan of Geekbench being so DDR5 memory overclock sensitive, because it doesn't reflect the realworld impact.

Timur Born said:
Most timings can be reduced without increasing frequency. Some need extra voltage to do so, which means that you can just increase frequency alongside then. That being said, I need to go quite out of my way to measure differences between higher frequency overclocks and my tuned 5600 CR1. At the same time I can use low voltages and be stable even at highest dimm temperatures (that are lower to begin with at low voltages). DDR5 is fast to begin with and tuned vs. untuned seems to have more of an impact than higher vs. lower frequencies, at least above 5600 MT.

My most memory intensive rendering application is Topaz Gigapixel AI, it uses 10 GB to upscale a *single* image to 25k+ pixels. Measurable difference between 5600 and 7000 MT: none. 7-Zip's compression is quite memory sensitive, between 5600 and 8200 MT you gain 15% compression performance. But compression is a niche application, DEcompression is much more common. Difference for DEcompression between 5600 and 8200 MT: none.

Is there data on Geekbench's memory scaling (latency, bandwidth)? Or a specific test you've seen? I'd be interested if there is some data, esp on a single system.

//

The test from H433x0n (6800 MT/s yields 3319 1T) to Igor's link (8000 MT/s yields 3509 1T) is +5.72%, though many variables might not be controlled (testing environment, temps for TVB, etc.). Assuming net latencies were similar, that's not a large increase.

And I'd definitely agree: tuning RAM probably can make a bigger difference than only clocks & CL out of the box.

//

For now, some tests 6800 MT/s CL34 → 8000 MT/s CL38 (Tom's Hardware). Cherry-picking some results where we see improvements, as some workloads like Office, Cinebench, etc. didn't budge or went the other direction (!).

That is, even if Topaz doesn't show a measurable difference, there should be some improvement that can get teased out in other "real-world" applications:

+4.89% in PhotoShop
+2.17% in Lightroom
+6.66% in Far Cry 6

I'm not saying these are representative; I just mean to share that it could be perhaps Topaz or methodology, as RAM speed should improve some "real-world" applications somewhat, too.

Representative of Tom's Hardware's full suite, there is some improvement to be had. I'd say margin of error, but it appears the slowest memories have the lowest scores in this geomean of some two dozen tests.

+2.4% in geomean of entire suite

^^ not ideal as this suite includes lots of synthetics, including a RAM bandwidth test (!)

Thus I'd say GB6 could be a little more memory sensitive than most applications, but not by a significant absolute sense (+2-3% more?).

"Memory sensitive" is vague there as I've not found any specific testing of latency-only vs bandwidth-only tests, but some combination of both.

igor_kavinski · Jan 16, 2024

Very anomalous result from the only 48GB kit. Guess the higher density memory chips can be quite advantageous in certain workloads.

igor_kavinski · Jan 16, 2024

48GB strikes again with AMD:

ikjadoon · Jan 16, 2024

igor_kavinski said:
View attachment 91988
Very anomalous result from the only 48GB kit. Guess the higher density memory chips can be quite advantageous in certain workloads.

That is interesting, a +9.7% improvement . I've learned UL Procyon's test is a pure export (i.e., render) test, so no edit-based benchmarks (like Puget does). I don't have this test and neither do many, it seems.

In general with Premiere Pro, some found faster exports with more RAM, others not so much. Premiere Pro seems to increase its RAM usage during export if you have more RAM, even if you weren't close to your physical limit, e.g,. 17GB usage on a 32GB system to 21GB usage on a 64GB system.

32GB to 64GB: ~9% improvement with a 4K, 100 Mbps source file: 687s on 32GB, 631s on 64GB
16GB to 32GB: ~12% improvement with a 4K export to 40 Mbps: 850s on 16GB, 726s on 32GB

Maybe based on the bitrates & total file sizes (unsure if source or destination is more important)?

Timur Born · Jan 21, 2024

Ain't these all XMP based comparisons? And concerning Geekbench testing, I only did my own tests of tuned 5600 CR1 vs. higher MT at equal to lower total timings (in nanoseconds). I did not save results, though, as I was only interested to see for myself how memory OC affects GB results, which it does. It's more than just 0.something % from what I remember.

My 5600 MT peak at about 94 gb/s read-rate. Of all the primary, secondary and tertiary memory timings the following are at their minimum and thus their latency could only be improved upon by using higher MT.

CR 1T, RDRD_DG 8T, WRWR_DG 8T

I am not sure if RDD_L/S is at a minimum at 4T or if lower values can be used on Intel, same goes for FAW at 16T.

All other timings can just be set to lower values for pure timing/latency improvements without needing to increase MT. BUT, some of those timings need higher voltages to stabilize at lower values, which in turn can then just as well be used to increase MT along the way (if voltage is already increased for timings anyway).

igor_kavinski · Jan 21, 2024

Timur Born said:
CR 1T

What kit do you have that can do CR 1T?

Timur Born · Jan 21, 2024

igor_kavinski said:
What kit do you have that can do CR 1T?

Hynix M-die (2x 16 GB) running 5600 MT at 1.20 V (was 1.15 V before one of the earlier BIOS updates, RDWR was 18 T before one of the later BIOS updates). I consider some of these timings "meme timings" meant to keep the angry "you are castrating your 13900K" crowd away from their torches and forks.

Timur Born · Jan 21, 2024

It's noteworthy that going above CR 1T doesn't work easily (or maybe at all) above 5600 MT even using considerably higher voltage. RP/RCD (coupled on Gigabyte) need 1.35 V for 12.0 ns while I am using 1.20 V for 12.5 ns here, but CR 1T doesn't even work at 5800 MT using 1.35 V.

So 5600 MT seems like a sweet spot. Also comes at the benefit of being basically JEDEC from the IMC's point of view, so 1.10 - 1.15 V VDD2 is enough for the timings shown in my screenshot. Since my IMC/board (6 layer PCB, GB Z790 Aero) refuses to run VT3 stable 6800 MT with similar timings (in ns) I suspect that my IMC isn't too great (and my 13900K is about average for core clocks).

Concerning Geekbench, I suspect it mostly likes memory bandwidth.

igor_kavinski · Mar 14, 2024

Comparison of some Ryzen top CPUs with the lone Core i9-14900KS result on Geekbench browser:

ASUS System Product Name vs ASUS System Product Name - Geekbench

Micro-Star International Co., Ltd. MS-7D67 vs ASUS System Product Name - Geekbench

ASRock X670E Steel Legend vs ASUS System Product Name - Geekbench

ASUS System Product Name vs ASUS System Product Name - Geekbench

Geekbench really needs to capture more differences in configuration, like the speed/latency of RAM so we can figure out why the same CPU is showing wildly different performance results in different subtests.

FlameTail · Mar 14, 2024

Geekbench should have 2 different types of MT tests.

gdansk · Apr 4, 2024

I ran Geekbench 6 on my 5900X server recently and noticed something. Geekbench 6 has the 5950X and 5900X at approximate parity in ST and MT.

AMD Ryzen 9 5900X Benchmarks - Geekbench

AMD Ryzen 9 5950X Benchmarks - Geekbench

Funny/Reality?

Hitman928 · Apr 4, 2024

gdansk said:
I ran Geekbench 6 on my 5900X server recently and noticed something. Geekbench 6 has the 5950X and 5900X at approximate parity in ST and MT.

AMD Ryzen 9 5900X Benchmarks - Geekbench

AMD Ryzen 9 5950X Benchmarks - Geekbench

Funny/Reality?

It makes sense given that they are very close in ST boost speeds and GB6 (overall score) doesn't really scale past like 6 cores.

gdansk · Apr 4, 2024

Hitman928 said:
It makes sense given that they are very close in ST boost speeds and GB6 (overall score) doesn't really scale past like 6 cores.

Yes, I guess so. It also shows the 5800X as faster than the 5900 (in MT). Well, I guess that's the limit of GB6. Can't use it for comparison of even small servers like this.

Question Geekbench 6 released and calibrated against Core i7-12700

Lifer

Senior member

Lifer

Senior member

Lifer

Golden Member

Senior member

Lifer

Senior member

Platinum Member

Senior member

Golden Member

Senior member

Member

Lifer

Lifer

Member

Senior member

Lifer

Senior member

Senior member

Lifer

Platinum Member

Platinum Member

Diamond Member

Platinum Member