Looking for tools to evaluate IPC on certain Intel processors? Software recommendations are available.
Looking for tools to evaluate IPC on certain Intel processors? Software recommendations are available.
Early last year I conducted performance comparisons between Zen 2 and Skylake, and you can follow the links for additional details on Zen, Zen+ and Skylake-X as well. Recently I purchased some Alder Lake components and plan to repeat similar evaluations across different Intel platforms: Broadwell Skylake-X (without AVX-512 usage, which is typical for Skylake, Alder Lake, Coffee Lake and Comet Lake), Rocket Lake Alder Lake (P-cores only). I’m curious about AMD options. My only current AMD processor is a Zen 3 APU in my laptop, but I lack the necessary access to run comprehensive tests there. It’s important to clarify that I’m assessing the CPU architecture rather than individual models. This distinction matters, and I’ll address any constraints that could affect results at the architectural level. The main concern is selecting appropriate testing tools. Here are some ideas that come to mind: Windows-based software, free and focused on CPU performance—no games required. Prime95 for FFT calculations, Y-cruncher for benchmarking, possibly Pi benchmarks around 25M or 1B. I’m aiming to minimize RAM-related issues. Cinebench targets should be at least R15, R23, or better; R20 seems unnecessary beyond R23. I’m uncertain if testing older CPUs (pre-AVX) adds meaningful data since they’re not AVX-optimized. Blender 3DPM and Cinebench are on my list, but I need more recommendations. The current suggestions feel limited. Aida64 feels too artificial and hard to relate to real-world performance. Updating to the latest versions is challenging for me. CPU-Z offers detailed metrics but doesn’t cover everything.
I’m open to contributing 8600G clock cycles if needed. For a CPU benchmark, using 7-zip compression and decompression seems like a solid choice, though it does involve some memory usage (not as extreme as Y-Cruncher, but still significant enough to note). Adding code compilation tests—such as Firefox, Chromium, or Linux Kernel benchmarks—could be helpful too. V-Ray is also free and worth considering, even though I haven’t tried it myself. If your system doesn’t use AVX-512, that’s probably not a concern. R20 and R23 are very similar except for scoring methods and AVX-512 support; picking one should suffice.
Can we turn off certain cores and adjust the clock speed? Yes, we can discuss this later after planning a clearer test approach. To minimize cache and memory access effects, I once restricted lower CPU cores at lower frequencies. Now I’m more willing to increase frequency and risk memory constraints. I’ve been using 7-Zip for years but haven’t tested its performance yet. I’ll check it later. It seems this might focus more on caches rather than overall speed. If true, it wouldn’t be ideal. I believe Y-cruncher runs efficiently at smaller sizes with a footprint under 200MB, which is about the same as 1B (~5GB). I’m worried about too much variability in the environment affecting results. Thanks, looks like there’s potential. I’ll dig deeper. R15 predates AVX; according to my knowledge, R20 and R23 support AVX but not AVX-512.
I fully manage which cores are turned off and their respective speeds, confirmed by testing it as a triple-core locked at 2GHz. I can also toggle hyperthreading if needed. Based on my trials, performance grows proportionally with core clock, much like Y-Cruncher, though memory usage impacts scaling more noticeably compared to Y-Cruncher 1B, which relies more on timing data. According to what’s known, R20 does support AVX-512, but that feature was dropped in R23 because its advantages were minimal.
It'll be worth checking HT scaling as well, though it will definitely increase testing time. I'll finalize the setup once Alder Lake is ready, as that should offer the best performance and let me observe how RAM impacts it. Zen 4 seems likely to match or exceed it, which might reduce its usefulness for this benchmark since we're more interested in execution speed than memory capacity. I'll experiment with lower compute levels to minimize RAM effects. Curious. The source I found only mentions R23, so I'll treat R20 as unconfirmed. Perhaps testing on Skylake-X systems—both enabled and disabled AVX-512—would help. Also, the later note about 2024's AVX-512 usage being irrelevant might bring us back to older benchmarks.
You're referring to the performance gaps between different benchmarks. The difference in scores between 6000 CL30 and 5200 JEDEC was roughly 2%—around 80800 versus 79000. It's noticeable but not extreme. On Raptor Lake (13700k), the improvement from 7200 CL34 to 4800 CL40 is about 6%, which is more apparent, though still likely within acceptable limits.
I overlooked that point. It’s not as severe as it seemed. With less CPU power, the gap should become smaller.
Need to experiment a bit with it. Usually I treat Prime95 and Linpack together since they both struggle with FP64. It mentions support for AVX-512, but I wonder if it actually uses it. The MKL version is from 2018, and Skylake-X came out in 2017, which might be enough.