Which Cinebench version for CPU reviews?

[W1zzard]

While we'll definitely be using R23 for Alder Lake reviews.

I just wondered, if you had the choice, which version would you prefer to see in our CPU reviews going forward?

 

[Melvis]

Which ever one is the fairest and most accurate for both companies.

 

[W1zzard]

Which ever one is the fairest and most accurate for both companies.

I have no idea. I do know that R23 has increased the runtime duration so that Intel CPUs will drop into their lower power limit state. And I think R15 workload was small enough to fit into Ryzen L3 cache.

 

[Ibizadr]

R20 have a low time for benchmark and I think its pretty good and fast.

 

[Aquinus]

R20 have a low time for benchmark and I think its pretty good and fast.

I guess that's good if you're goal is to see how CPUs compare with bursty loads and not sustained loads.

 

[ratirt]

I have no idea. I do know that R23 has increased the runtime duration so that Intel CPUs will drop into their lower power limit state. And I think R15 workload was small enough to fit into Ryzen L3 cache.

Maybe there should be a divide in burst loads and sustained loads? The point with Intel is a good one. During a sustained full load Intel will have to drop the clocks to finish and thus get lower score (R23 10 loops to finish for example) in burst load to test the highest performance in certain scenario is different (one run, finish as fast as possible and get the score, and since it wont hit the thermal wall it will be higher). Although, I'm sure that is not your goal to make more benchmarks but rather focus on one and present the data.
Some sort of distinguished approach.

 

[Ibizadr]

I guess that's good if you're goal is to see how CPUs compare with bursty loads and not sustained loads.

Yeap but the main question here is to review cpus and not to see if they were stable at certain settings

 

[W1zzard]

Although, I'm sure that is not your goal to make more benchmarks but rather focus on one and present the data.

I have no plans to add more synthetics. Will only run one Cinebench version (at 1T and nT), and focus on real application benchmarks otherwise

Added a "Don't Care" option that shows me you saw this thread, thought about it and don't care

 

[TheoneandonlyMrK]

Yeap but the main question here is to review cpus and not to see if they were stable at certain settings

Some of us Do want to run some stuff on them 24/7, we don't all use them the same clearly but why should only gamer's get relevant benches?!.

 

[Ibizadr]

Some of us Do want to run some stuff on them 24/7, we don't all use them the same clearly but why should only gamer's get relevant benches?!.

I think it's not about gamer benches cause all benches it's just a number, position in relation with other cpus. That's the main question, and if you want to bench for example ryzen 5000 with a certain ram it's more faster to bench. In r23 you need to do 10min(default I don't know if it's possible to change time or by passes) if you need to do 3 times to got avg score it's 30min for ST and 30min for MT than dismount cooler cpu put all back again to other bench consume more time. I don't know if r23 is more accurate than r20, but most of last leaks off Alder Lake was r20 I think.

 

[TheoneandonlyMrK]

I think it's not about gamer benches cause all benches it's just a number, position in relation with other cpus. That's the main question, and if you want to bench for example ryzen 5000 with a certain ram it's more faster to bench. In r23 you need to do 10min(default I don't know if it's possible to change time or by passes) if you need to do 3 times to got avg score it's 30min for ST and 30min for MT than dismount cooler cpu put all back again to other bench consume more time. I don't know if r23 is more accurate than r20, but most of last leaks off Alder Lake was r20 I think.

Yeah and most leak's of Ryzen were on r15.

I don't care what was tested before, those results on disparate systems with unknown configurations at various times and with various OS update levels are not relevant to this round of testing.

Any test of CPU performance IMHO needs to apply to the whole time I will use a CPU not just the first two minutes at pl2 or the performance while the cooling system saturates and settles.

 

[mrthanhnguyen]

Hi Wizzard, for gaming benchmark, can u do with the built in benchmark ( if included) so we can compare it easier?

 

[chrcoluk]

While we'll definitely be using R23 for Alder Lake reviews.

I just wondered, if you had the choice, which version would you prefer to see in our CPU reviews going forward?

The same as older reviews so you can show comparison over multiple generations of products.

 

[cvaldes]

It makes more sense to use a test that reflects modern world usage and workloads with today's hardware. Workloads larger than L3 cache? Happens all the time. Longer duration workloads? Absolutely.

There's little sense in using an ancient test with modern hardware in 2021.

Test duration is important. All modern CPUs have impressive boost frequencies for short bursts of work. Real world performance is better captured by a sustained workload. I think the 10 minute duration of R23 is appropriate.

The main point of Cinebench is to assess the fitness of hardware for Cinema 4D's rendering performance, mostly on higher-end PCs and workstations. Do you know anyone who has 3D rendering projects that take two minutes? I don't. Do you know anyone who frequently renders 3D scenes on smartphones, tablets or low-end computers like a $200 Chromebook? I don't. These are sustained workloads of complex projects for hours, sometimes days.

Anyhow, that's the way I see it. I know someone here will disagree.

 

[TheoneandonlyMrK]

The same as older reviews so you can show comparison over multiple generations of products.

Windows 10-11 happened though, as did driver and OS updates.

They're not comparable in any way or case in reality.

 

[docnorth]

CB 23 results for 12900K just arrived on Videocardz. Very good results with 150w power limit, only small gain beyond that. 217,8w maximum consumption with power limit set to 241w.

 

[Arctucas]

Maybe there should be a divide in burst loads and sustained loads? The point with Intel is a good one. During a sustained full load Intel will have to drop the clocks to finish and thus get lower score (R23 10 loops to finish for example) in burst load to test the highest performance in certain scenario is different (one run, finish as fast as possible and get the score, and since it wont hit the thermal wall it will be higher). Although, I'm sure that is not your goal to make more benchmarks but rather focus on one and present the data.
Some sort of distinguished approach.

So, a one-pass test will score higher than a ten minute test?

 

[TheoneandonlyMrK]

So, a one-pass test will score higher than a ten minute test?

With modern multi-core's it doesn't really matter, the fast burst characteristics are actually captured by the single-threaded bench adequately, because the load gets passed about around the core to keep speeds up and temperature spread about, IMHO.so with that and a long multi-thread bench, it's sorted.

Sorry, I went and quoted the wrong guy again, Yes would answer your question though.
As heat builds, speeds and power use drop as the parts are made to do.

 

[chrcoluk]

Windows 10-11 happened though, as did driver and OS updates.

They're not comparable in any way or case in reality.

I think as far back as haswell was done on 10, the new tests can be done on 10 as well (in my opinion should be as thats dominant OS).

If vendors are asking for 11 to be used then its not an independent review.

I would do on 10 and on both R15 and R23 so a phased move to R23 can be done whilst also allowing people to compare to their older current chip.

What reviewers may forget is that the useful info is new gen vs older gens not new gen skus against each other.

If im considering an upgrade I need to know performance gain vs my current cpu. So not 12900k vs 12700k, not 12900k vs 5950, but 12900k vs 9900k, on same software.

I know there has to be a cutoff point but I think 3-4 generations is a reasonable minimum to have comparable data.

 

[ratirt]

So, a one-pass test will score higher than a ten minute test?

With Intel that's the case. The CPU will have to drop to the lower power state due to continues stress so the score for MT for instance will be lower than if you had run it just one time.
I think it affects all processors but Intel has the bursts clocks which can only be sustained for a short time (short intervals) or be skipped completely if thermals are too high which obviously will be the case if you run a loop for 10-15 min. Heck it may not even turbo boost to the max due to that.

 

[mama]

R23 may consider temps better.

 

[TheoneandonlyMrK]

I think as far back as haswell was done on 10, the new tests can be done on 10 as well (in my opinion should be as thats dominant OS).

If vendors are asking for 11 to be used then its not an independent review.

I would do on 10 and on both R15 and R23 so a phased move to R23 can be done whilst also allowing people to compare to their older current chip.

What reviewers may forget is that the useful info is new gen vs older gens not new gen skus against each other.

If im considering an upgrade I need to know performance gain vs my current cpu. So not 12900k vs 12700k, not 12900k vs 5950, but 12900k vs 9900k, on same software.

I know there has to be a cutoff point but I think 3-4 generations is a reasonable minimum to have comparable data.

Ok but, the tests need re-doing, they're still not comparable.
And I would imagine in a CPU review it won't be against just it's own range, I expect a few different reviewers to span a reasonable range of CPU.

 

[Melvis]

I have no idea. I do know that R23 has increased the runtime duration so that Intel CPUs will drop into their lower power limit state. And I think R15 workload was small enough to fit into Ryzen L3 cache.

\
Well I think you answered your own question here really.

Since you will be using the same cooler across all platforms then go with R23.

 

[Arctucas]

With Intel that's the case. The CPU will have to drop to the lower power state due to continues stress so the score for MT for instance will be lower than if you had run it just one time.
I think it affects all processors but Intel has the bursts clocks which can only be sustained for a short time (short intervals) or be skipped completely if thermals are too high which obviously will be the case if you run a loop for 10-15 min. Heck it may not even turbo boost to the max due to that.

I presume this is only with default settings, i.e. not adjusting PL1 or PL2 in BIOS?

 

[freeagent]

I would run R23. You don’t have to let it run for 10 minutes..

Edit:

I use the one that is bundled with BenchMate if that helps