Y-Cruncher Tips and Tricks
I really like this benchmark as it tests outside of your normal skill set and requires a bit more consideration when benching for a global position on hwbot.org
https://hwbot.org/rules?referenceId=5273
This benchmark relies heavily on AVX instructions, cache speed, memory timing and speed. This benchmark teaches you to balance stability and to relax settings to get better results.
If you can master this benchmark, it will improve your other scores in other benchmark, and it will help you identify other problems quicker. To know between memory volts and bad timing settings from various error messages in the bench.
As I mentioned AVX and AVX-512 speeds, this is by far the most important setting to learn and master. Learning the AVX offsets with your core multiplier is key. Finding the highest possible AVX speed and starting with CPU multi 1:1 with AVX (offset 0) is brutal, it seems I get better scores with offset 1, offset 2 my scores are a bit slower, offset 3 is slower, and so on...
You need to find your AVX offsets first, keep cpu multi on auto and volts on auto, even keep your memory at SPD or XMP for this to rule out issues. Lets us my XEON for example, I have two AVX offsets that should track together but they give individual settings. If I do not set these the same my machine will not post, so I keep my AVX and AVX-512 offset the same.
I use Hwinfo to see my AVX ratio, it is 37x by default "fused" at boot to the CPU multi, so if your boot multi is 43x on CPU then your offset is 6, each time you lower the offset it increases the AVX ratio by 1. When I say fused at boot, if you go into turbo vcore or another PLL adjust program for your mobo and change the cpu multi it will change the avx multi as well, up or down.
You will find that keeping your CPU close to 1:1 with AVX produces the best scores. Using PLL increasing the CPU multi will scale nicely.
Cache speed, find you max cache speed or you may already know it seems ok. This bench scales a little with cache speed. You may need to scale cache back 1x for 2.5B and 10B
Memory timing is very important but slightly relaxed timing might produce better results and the same timings for 25M might not pass the 2.5B and 10B test.
The benchmark length is pretty quick for 25M and you can run with fastest timing, highest AVX, memory and cache speeds. The 25M should be good with your 3D profiles since it takes .25s run.
Leave your vcore on auto or you are going to have a bad time. You may see huge improvements in score by reducing CPU multi and keeping CPU form throttling on longer tests.
You will see huge loads with Y-Crunch, you will find these new AVX settings work well with other benchmarks, you will find a new level of stability, you will also learn how to decipher what crash is what component and adjust correctly quicker.
I use PLL program to find best settings and set in bios for boot, this produces the best results. Manually adjusting bclk scales, find your motherboards maximum on this bench.
All memory slots full also produces a better result
HT disabled produces a better result, even with the 10B plus this also helps keep CPU load far lower with it disabled in bios.
Using AIDA64 memory read test helps dial in memory
Shutting down Explorer produces a better result
First run can "bonus" rarely.
https://hwbot.org/benchmark/y-cruncher_-_pi-1b/
