Question Geekbench 6 released and calibrated against Core i7-12700

[igor_kavinski]

Geekbench 6 - Geekbench Blog

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.

 

[igor_kavinski]

9700X vs. "average" 7700X: https://browser.geekbench.com/v6/cpu/compare/6853850?baseline=6832932

File compression, Clang, Text Processing and Raytracer are all good improvements, especially if the 7700X sees a price drop.

 

[CouncilorIrissa]

9700X vs. "average" 7700X: https://browser.geekbench.com/v6/cpu/compare/6853850?baseline=6832932

File compression, Clang, Text Processing and Raytracer are all good improvements, especially if the 7700X sees a price drop.

It's actually the worst 9700X run, there are two better runs.

ASUS System Product Name vs Micro-Star International Co., Ltd. MS-7D75 - Geekbench

ASUS System Product Name vs Micro-Star International Co., Ltd. MS-7D75 - Geekbench

Regression in the Navigation run seems to be a reoccuring theme.

 

[igor_kavinski]

It's actually the worst 9700X run

That's even better. Even with the worst 9700X, you get reasonable gains!

 

[igor_kavinski]

With crappy DDR5-4800, still good enough! https://browser.geekbench.com/v6/cpu/compare/6853850?baseline=6832932

Imagine what it could do with DDR5-6000 CL28.

 

[Hail The Brain Slug]

Comparison with souped up machine of @Hail The Brain Slug: https://browser.geekbench.com/v6/cpu/compare/6852850?baseline=6830619

Comparison with an "average" 7950X: https://browser.geekbench.com/v6/cpu/compare/6852850?baseline=6844850

It's... not really souped up. The CPU is fully stock on a normal sized air cooler. FCLK and RAM is tuned nicely, but it seems like we are seeing tons of examples of Zen 5 chips running 6400/2133 so the only real advantage is the somewhat extra tuned memory timings.

Is there a different ram/fclk config that you think would better represent a typical 7950X?

I never trust any provided benchmarks like this as representing the true stock performance of a part because they always seem to score worse than when I have the same hardware configured completely stock, running on a fresh OS install without a bunch of garbage services/background applications running.

 

[igor_kavinski]

Is there a different ram/fclk config that you think would better represent a typical 7950X?

I think a 7950X running EXPO @ DDR5-6000 CL30 or CL32 would better represent the vast majority of such machines in the wild.

A lot of users/system builders may not want to tweak anything. Just set the RAM to EXPO in UEFI and that's it. Why mess with stability and spend days troubleshooting it or have the customer call you with an unhappy tone?

 

[Hail The Brain Slug]

I think a 7950X running EXPO @ DDR5-6000 CL30 or CL32 would better represent the vast majority of such machines in the wild.

Single or dual rank (2x16 or 2x32)?

Gigabyte's EXPO improvement settings off or on? (High bandwidth, fast timings settings)?

 

[igor_kavinski]

Single or dual rank (2x16 or 2x32)?

Gigabyte's EXPO improvement settings off or on? (High bandwidth, fast timings settings)?

I'm not sure about that. Whatever is the default setting for EXPO. 2x16GB is by far the most common, don't you think?

 

[Hail The Brain Slug]

I'm not sure about that. Whatever is the default setting for EXPO. 2x16GB is by far the most common, don't you think?

The settings seem to be reliable and stable in my experience, but I don't think any other MFG have something quite like it. I'll just enable EXPO and nothing else.

2x16 is probably much more common, yeah

 

[igor_kavinski]

The settings seem to be reliable and stable in my experience

Gigabyte is the DDR5 guru.

 

[igor_kavinski]

Linking this post coz it is relevant: http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...ranite-ridge-ryzen-9000.2607350/post-41248641

 

[Hail The Brain Slug]

After some Google'ing and some Linux perf play, I found out that geekbench_x86_64 still uses AVX2 instructions.

Beware with Geekbench v6 results! | Theldus's blog

AVX-512 instructions are used in both cases, while the number of AVX2 instructions is reduced, but they are still in use:

perf stat -e cycles:u,instructions:u,fp_arith_inst_retired.512b_packed_single:u,fp_arith_inst_retired.256b_packed_single:u,fp_arith_inst_retired.512b_packed_double:u,fp_arith_inst_reti\
red.256b_packed_double:u  ./geekbench_avx2
     2082633539364      cycles:u                                                      (66.67%)
     2478076989817      instructions:u            #    1.19  insn per cycle           (83.35%)
       13135540644      fp_arith_inst_retired.512b_packed_single:u                                     (83.35%)
       42443442793      fp_arith_inst_retired.256b_packed_single:u                                     (83.32%)
                 0      fp_arith_inst_retired.512b_packed_double:u                                     (83.35%)
          27397319      fp_arith_inst_retired.256b_packed_double:u                                     (83.33%)

     335.663318006 seconds time elapsed

     505.778856000 seconds user
      21.272345000 seconds sys

perf stat -e cycles:u,instructions:u,fp_arith_inst_retired.512b_packed_single:u,fp_arith_inst_retired.256b_packed_single:u,fp_arith_inst_retired.512b_packed_double:u,fp_arith_inst_reti\
red.256b_packed_double:u  ./geekbench_x86_64
     2180095192498      cycles:u                                                      (66.69%)
     2748777186937      instructions:u            #    1.26  insn per cycle           (83.35%)
       13099820608      fp_arith_inst_retired.512b_packed_single:u                                     (83.30%)
       17024206599      fp_arith_inst_retired.256b_packed_single:u                                     (83.33%)
                 0      fp_arith_inst_retired.512b_packed_double:u                                     (83.36%)
                 0      fp_arith_inst_retired.256b_packed_double:u                                     (83.33%)

     338.838301492 seconds time elapsed

     521.407118000 seconds user
      22.125630000 seconds sys

I will let @Hail The Brain Slug post his own results, both with AVX-512 disabled and enabled at BIOS level.

I'll get to this eventually, I swear!

 

[Nothingness]

I'll get to this eventually, I swear!

Take your time, there's no hurry 😀

 

[Hail The Brain Slug]

Take your time, there's no hurry 😀

Those AVX instructions don't appear to exist on my AM5/7950X system, so I am running the only AVX512 instruction that perf lists and the matching AVX256 instruction.

perf stat -e cycles:u,instructions:u,fp_ops_retired_by_width.pack_256_uops_retired:u,fp_ops_retired_by_width.pack_512_uops_retired:u ./geekbench_avx2

AVX 512 enabled results:

Performance counter stats for './geekbench_avx2':

1459209690095 cycles:u
2201820943555 instructions:u # 1.51 insn per cycle
311381405008 fp_ops_retired_by_width.pack_256_uops_retired:u
25566637170 fp_ops_retired_by_width.pack_512_uops_retired:u

250.079178916 seconds time elapsed

285.817969000 seconds user
24.687398000 seconds sys

AVX512 disabled results:

coming soon

Edit: I can't get AVX512 disabled results. The system reboots at a random point during MT testing without fail. I messed with some UEFI settings to see if it was some kind of CPU or RAM instability issue but it is not. I'm running a live boot so there are no persistent logs after the reboot. When I have more time later I'll have to do a full install so I can take a look at the logs and see what is going on.

I did stop a run before it rebooted and there are still fp_ops_retired_by_width.pack_512_uops_retired:u instructions counted. Perhaps it is not the correct instruction to check for, but there were no other 512 instructions referenced under the fp list on my system.

The AVX512 disable is definitely doing something otherwise it wouldn't be rebooting before it can complete every time I try to run it. (Runs perfectly fine with AVX512 enabled).

 

[mmaenpaa]

Which subtest (if any) of GB6 is using SSE4.1 or SSE4.2 extensions?

I am building a new workstation for my customer (for Mastercam). I have asked supplier if software uses some form of AVX, but no answer so far.

I assume they are supporting only these, because Intel I9-14900 has only these:
Intel® SSE4.1, Intel® SSE4.2, Intel® AVX2

I am building it with AMD anyway but I am wondering if I need to wait for Zen5.

New workstation will be 7950X or 9950X based.

 

[Nothingness]

Which subtest (if any) of GB6 is using SSE4.1 or SSE4.2 extensions?

I am building a new workstation for my customer (for Mastercam). I have asked supplier if software uses some form of AVX, but no answer so far.

I assume they are supporting only these, because Intel I9-14900 has only these:
Intel® SSE4.1, Intel® SSE4.2, Intel® AVX2

I am building it with AMD anyway but I am wondering if I need to wait for Zen5.

New workstation will be 7950X or 9950X based.

Geekbench6 technical information is here: https://www.geekbench.com/doc/geekbench6-benchmark-internals.pdf

No mention of SSE4 it seems, though it might be used in Embree library (but in that case, I guess that lib will use the more recent extensions rather than the old SSE4).

 

[igor_kavinski]

I am building it with AMD anyway but I am wondering if I need to wait for Zen5.

What's ten more days? 9950X would be better, especially if they update Mastercam with a revision to squeeze more performance out of the new CPU.

 

[igor_kavinski]

The only explanation would be that ARL is dispatched 5-6 threads in as much cores, since it doesnt have SMT, while Zen 5 is dispatched the 5-6 threads in 4 cores with two cores using SMT, but whatever the reason this expose blatlantly GB 6 as being a doubtfull bench, either badly designed or worse, tricked.

Det0x 7950X: https://browser.geekbench.com/v6/cpu/4005773
Det0x 9950X ES: https://browser.geekbench.com/v6/cpu/7040992

My conclusion?

Yeah, maybe something is wrong with the Gigabyte 9950X since it is showing lower scaling.

HOWEVER,

Notice that 9950X ES scaling is far less than 7950X in Object detection.

While in Photo filter, the scaling is improved.

It just comes back to the same conclusion many others have drawn so far: Something in 9950X is gimping integer workloads badly and it could probably be bad design or some overlooked shortcoming.

Another one is that ARL scales better in MT workloads, at least in GB6.

Previous post:



The 9950X is running @ DDR5-7200 while ARL is at 6400 MT/s.

This one is against a 9950X @ 6200 MT/s: https://browser.geekbench.com/v6/cpu/compare/7425534?baseline=7465273



So a tuned 9950X isn't that far behind ARL in ST INT tests and obliterates it in MT INT.

Arrow Lake gonna need higher speed RAM, at least 8400 MT/s.

 

[Nothingness]

View attachment 106116

Det0x 7950X: https://browser.geekbench.com/v6/cpu/4005773
Det0x 9950X ES: https://browser.geekbench.com/v6/cpu/7040992

My conclusion?

Yeah, maybe something is wrong with the Gigabyte 9950X since it is showing lower scaling.

HOWEVER,

Notice that 9950X ES scaling is far less than 7950X in Object detection.

While in Photo filter, the scaling is improved.

It just comes back to the same conclusion many others have drawn so far: Something in 9950X is gimping integer workloads badly and it could probably be bad design or some overlooked shortcoming.

Another one is that ARL scales better in MT workloads, at least in GB6.

Previous post:

View attachment 106117

The 9950X is running @ DDR5-7200 while ARL is at 6400 MT/s.

This one is against a 9950X @ 6200 MT/s: https://browser.geekbench.com/v6/cpu/compare/7425534?baseline=7465273

View attachment 106118

So a tuned 9950X isn't that far behind ARL in ST INT tests and obliterates it in MT INT.

Arrow Lake gonna need higher speed RAM, at least 8400 MT/s.

It would be helpful if you added a column showing MT vs ST speedup for both CPU 😀

 

[igor_kavinski]

It would be helpful if you added a column showing MT vs ST speedup for both CPU 😀

Most excellent idea!



Comparison of the speedup. First graph is INT and second is FP test speedup.

 

[Abwx]

Most excellent idea!

View attachment 106120

Comparison of the speedup. First graph is INT and second is FP test speedup.

View attachment 106122
View attachment 106123

For the 285K the geomean average speed up from ST to MT is 6.387 for INT and 7.24 for FP, so that s tailored for 8 P cores/8 threads, and the few tests that use more than those 8 P cores are leveled by the geomean average, Text Processing FI has a 1.18 speed up, wich is just ridiculous but not so much if the point is to level the geomean.

That s why the MT scores are more and more compressed as you increase the core count as to yield a minimal improvement for the added marginal core.

 

[igor_kavinski]

Assuming AMD plans to pit the Ryzen 160 PRO against Lunar Lake: https://browser.geekbench.com/v6/cpu/compare/7483669?baseline=6787697

Loses in ST but looks good in MT with decisive wins for LNL in only two MT tests.