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.
So you are saying this benchmark is ONLY for use on desktop PC's ? not even high end ones that CAN use up to 32, or even 64 cores ?
So, anybody that has need of more cores than 4 or 6 and uses a desktop is stupid if they want 32 cores ? Like the people in the DC forum that buy old 14 and 22 core Xeons ? They have a desktop, but want and need more cores.Sure it can be used for high-end systems too but it benchmarks typical desktop-use cases. For such a use 32 or 64 cores won't bring any perfomance uplift. And Geekbench should show it so stupid users won't go and buy high-end workstation type cpu's for desktop usage. https://www.geekbench.com
So, anybody that has need of more cores than 4 or 6 and uses a desktop is stupid if they want 32 cores ? Like the people in the DC forum that buy old 14 and 22 core Xeons ? They have a desktop, but want and need more cores.
The bottom line is that you are validating that this benchmark is worthless for more than 8 cores. Thanks for that opinion. Most of us would like something that shows real MT performance.
You sound like an ad for geekbench, even linked it. If you are an employee, thats fine, but several of us do not agree that its a good benchmark, myself for one.It benchmarks typical desktop-use cases. There's not much additional performance to be found on those use-cases with additional cores and Geekbench6 shows it pretty well.
If someone has workload that scales to unlimited cores he should use benchmark with single-thread workload copied to all cores. Like Geekbench did before - but that totally failed to show system's multithreaded performance on typical desktop-loads. And as pretty much everything at desktop is nowadays MT MT score correlates better to real desktop-performance than pure ST - if done with real desktop-type applications. GB6 is a big step to right direction for desktop-benchmarking.
You know who has exactly EIGHT performance cores? Intel.For most of desktop use cases cores beyond 8 are totally worthless.
You know who has exactly EIGHT performance cores? Intel.
You know who has more than that? AMD.
Isn't that suspicious???
And them yanking GB5 downloads so abruptly from their site. Like someone forced them to...
AMD's consumer-oriented chips also have 8 cores, including their mobile ones. You could even argue that this scaling hurts Intel more, comparing 2+8 or 6+8 vs AMD's 6/8c offerings.You know who has exactly EIGHT performance cores? Intel.
You know who has more than that? AMD.
Isn't that suspicious???
And them yanking GB5 downloads so abruptly from their site. Like someone forced them to...
So why do you think a benchmark that has been explicitly targeting client systems from day 1 should ignore how those workloads behave in favor of server ones?You sound like an ad for geekbench, even linked it. If you are an employee, thats fine, but several of us do not agree that its a good benchmark, myself for one.
Not that this could be done.... but what I personally would like is something that truly shows the total performance of a chip. One case would be that Zen 4 or P-cores are fully loaded as well as E-cores if present. The E-cores would add less to the overall score, as we know they are weak compared to Zen4/P-cores, but the total "horsepower" of a chip would be represented. higher end chips for desktop, as well as true HEDT and workstation would also work.AMD's consumer-oriented chips also have 8 cores, including their mobile ones. You could even argue that this scaling hurts Intel more, comparing 2+8 or 6+8 vs AMD's 6/8c offerings.
And what is the alternative you propose? That benchmarks should ignore how real software behaves just because you don't like the implications for your brand of choice?
Here is an example.... Games... The fastest chips are the 13900k and the 7950x most of the time on highly CPU dependent games. This benchmark has shown to be almost worthless with more than 8 cores, let alone 32.So why do you think a benchmark that has been explicitly targeting client systems from day 1 should ignore how those workloads behave in favor of server ones?
Not that this could be done.... but what I personally would like is something that truly shows the total performance of a chip. One case would be that Zen 4 or P-cores are fully loaded as well as E-cores if present. The E-cores would add less to the overall score, as we know they are weak compared to Zen4/P-cores, but the total "horsepower" of a chip would be represented. higher end chips for desktop, as well as true HEDT and workstation would also work.
And essentially none of that performance advantage is thanks to the extra cores vs the lower tier offerings. It's primarily the extra single core boost speed, and in the case of the 13900k, more L3 cache - both things that Geekbench MT should represent. And as a reminder, Geekbench doesn't have a hard limit on the number of threads. It's just that most of its workloads don't really benefit much beyond a dozen or so.The fastest chips are the 13900k and the 7950x most of the time on highly CPU dependent games.
As I said above, GAMES, a highly used desktop function have proven that more than 8 cores can be helpful. So for a 7950x and a 13900k, this benchmark is totally misleading.GB6 shows total found horsepower for those workloads it represents. More cores won't bring much additional performance to them - making fastest cores faster should increase performance pretty directly to both ST and MT. Many people don't understand that for most workloads increasing thread-count will not bring any benefit beyond their thread scaling limit - which is typically somewhere 4-8 threads for desktop loads.
As I said above, GAMES, a highly used desktop function have proven that more than 8 cores can be helpful. So for a 7950x and a 13900k, this benchmark is totally misleading.
HUH? Intel has many more cores than 8. Intel has 50% cores on a desktop platform than AMD! Why do you think that Intel would be happy with a benchmark, which plateaus after 8 cores, and that cannot show a difference between e.g. 13700K and 13900K?You know who has exactly EIGHT performance cores? Intel.
You know who has more than that? AMD.
Maybe would would make people happy are two MT score results. One for the 'embarrassingly parallel' workloads similar to GB5's MT tests and another for the 'thread cooperation' workloads that more accurately represent MT performance on most applications. Then you can compare results based on the type of workload you care about.
This depends on the workload. E-cores are quite strong for Integer workload. An E-core at 43x *exactly* matches the performance and wattage of a P core at 40x in 7-Zip benchmark. That's not too much of a difference between the two, especially considering that 16 E cores use the same space as 6 P cores....The E-cores would add less to the overall score, as we know they are weak compared to Zen4/P-cores...
Popular benchmarks like Geekbench should be forward looking. They should show what the future could look like with multicore optimized software. GB6 is now giving the impression to consumers that cores don't matter anymore.and further progress in this area should be halted.And what is the alternative you propose? That benchmarks should ignore how real software behaves just because you don't like the implications for your brand of choice?
I agree with this and think that it really needs to be done given the current hardware and software landscape. I'm OK with the kind of about face done with GB6, it just feels like a bit of a shock due to how different the MT works compared to GB5. As many others have mentioned, no benchmark is perfect/complete and any test needs to be looked at through the lens of what and how it is measuring. However, I think if the people behind GB want to have their benchmark used by a wide variety of systems from phones to workstations (as they have indicated), then I think they need to provide some kind of way of adjusting to the ever increasing use of cores in desktops and workstations.
It is true that the vast majority of programs used by desktop users are scaling much higher than 6 or so cores at most, but there is still a not insignificant portion that can use more (transcoding, streaming, multi tasking, compiling) and workstation users are even more likely to use many cores. I'm not sure the best labels to use for each category but something like single threaded, light desktop, and heavy workstation categories or something like that. Then people can look at each category and see how things stack up according to their use cases.
What evidence is there that there will be a substantial shift in client software to make dramatically better use of more threads? Progress over the last decade has generally been quite slow and incremental.They should show what the future could look like with multicore optimized software.
Geekbench 6 does still have extremely parallel workloads, like compiling and ray tracing. They're simply no longer the sole component of the score. And the reality is, they're right. It doesn't make sense to buy 16+ cores for consumer workloads, and that seems unlikely to change soon.GB6 says, there will never be optimized software. Don't waste money. Stick to 8 cores and stop daydreaming!
Golang and Rust make it easier to write parallel and concurrent programming code. Rust may even supplant C/C++ in the Linux kernel in the future. Just because progress in the recent past has been slow, it's no indication that things will remain that way in the future too.What evidence is there that there will be a substantial shift in client software to make dramatically better use of more threads? Progress over the last decade has generally been quite slow and incremental.
I'm certainly skeptical of the claim that newer languages, of all things, will provide substantially better parallelism. But even if you accept that as a possibility, what makes more sense for a benchmark today? To bet on an as-yet unsubstantiated hope, or to use current patterns and historical trends?Golang and Rust make it easier to write parallel and concurrent programming code. Rust may even supplant C/C++ in the Linux kernel in the future. Just because progress in the recent past has been slow, it's no indication that things will remain that way in the future too.
The cloud. Microsoft is trying to move everybody toSo... Which runs MS Office faster: current gen i5 of EPYC? Or games?