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.
AVX512 is not used by ARM CPUs, thats why they removed it.
Score is ruined by these:
Our concern is that most users, including those considering high-end workstations, don't know whether their workloads scale well.
We've seen several workstation workloads that do not scale well (where an 8-core system performs similarly to a 16-core system) and some that have negative scaling (where an 8-core CPU outperforms a 64-core CPU).
If users understand their workloads scale well, the workloads that scale well (e.g., Clang, Asset Compression, and Ray Tracer) can be used as a reference. Users can focus on those individual workload scores rather than the overall multi-core score.
Our assertion is that very few applications can take advantage of HEDT systems. Even developer workloads like compilation aren't guaranteed to take full advantage of HEDT systems (e.g., RISC-V toolchain compilation performance).
Our other assertion is that users, in general, do not know if their applications will scale. We've heard numerous reports of instances where users go off benchmark numbers and buy a HEDT system only to find that it's slower than their previous system.
Users who purchase these systems should understand that HEDT systems come with compromises (e.g., ThreadRipper systems have a non-uniform memory topology). Having a benchmark that reports best-case scaling does a disservice to users (and the industry in general). Instead, having a benchmark that uses a mix of workloads to measure average scaling (as Geekbench 6 does) gives users a better understanding of the trade-offs that come with HEDT systems.
Users who are sophisticated enough to understand these trade-offs and who are aware that their application will scale on HEDT systems can either use individual workloads in Geekbench 6 to measure performance or use other benchmarks entirely (even Geekbench 5). Adding a second multi-core score to Geekbench would make things more, not less, confusing for our users.
Agreed. If anything, the new MT test should make Intel look worse and AMD look better. Intel's strategy of "stuff as many little cores as possible to win at Cinebench" is much less effective when the cores actually have to work together on the same task, which is how most applications are designed.I applaud Poole for their new approach on estimating multicore performance. Nobody is running the same task replicated across every physical core. Measuring multi-core benchmark in this naive way essentially makes every task embarrassingly parallel and massively overestimates the real-world performance impact of many-core designs in the desktop/workstation space. GB6 seems to have a good mix of tasks which tend to be embarrassingly parallel (like ray tracing) as well as tasks that have complex dependencies (like compiling code). This is a step in the right direction.
It would be nice of the GB website showed multicore scaling for each of the task, this would allow one to evaluate the advantages and disadvantages of various CPU designs (like Intel's reliance on many small throughput cores vs. AMDs scalable symmetric cores) for cooperative task solving.
I actually agree. I was complaining before already that the correlation of the MT score to the ST score is way too strong now.That's not because of all the E-cores, but rather primarily the 8 P-cores yielding strong lightly threaded performance. If someone has 8+0 vs 8+8 vs 8+16 numbers, that would probably help illustrate it.
It's not. I was previously complaining that the new MT score moves the bottleneck from the chip to the workload. And if this MT test suite is (as I interpreted the quote) indeed targeted at benching the particular difference between hybrid and non-hybrid designs, the bottleneck on hybrid designs possibly moves onward to the scheduler.I guess, but how's that an advantage for hybrid designs? If anything, it makes the benchmark harder than blindly giving each thread an identical, isolated task.
That's exactly what Geekbench is aiming for. They even said "client applications" which is another way of saying "consumer level applications". If you want to test a 128c system, you need something else.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 ?
Seeing as how my 96 core Genoa was beat by an Alderlake something CPU mul multicore, I am not even going to bother trying it, due to this comment "and multi-core scores to be raised by up to 10%.".Geekbench 6.1 has been released: https://www.xda-developers.com/geekbench-6-1-released/
- Clang 16
- SVE/AVX512-FP16
- Fixed-point math
- Improved multi-core performance (I guess this one will appease some... or not )
Looks like they aren't going to use AVX-512 anymore.
Are these numbers (3027/21421) for a stock or overclocked CPU?For my 7950X