CPU Cache speed question on the x299

[Addiecool]

Hi

In my recent build as soon as I finished installing the system I ran some benchmarks. I saved a screen shot for only CPU-Z 1.80. To my surprise I got a score of 7100.
I had set the following parameters on the first stage and the system passed all stability tests.

4.7Ghz @ offset of 0.012v
3.2Ghz Cache @ 1.16v
DRAM at 3600Mhz @1.35v (16-16-16-32)
Rest everything I set to Auto

While testing a few things, I set the cache to 3 Ghz down from 3.2Ghz @1.155v. Though the system worked perfectly still and made no noticeable difference, the CPUZ score fell by 1200 points to 5700-5800. I did not know at this time that this was due to the what reason.

Baffled much, I tried everything I knew to fix this (thanks to my OCD :-) )

And suddenly I realized setting the cache to 3200Mhz upped the scores back to 7100 and a little better on other benchmarks.

So my question is, can cache speed make such a big difference in speeds? Is this possible? or a bug? I am sorta baffled still.

System spcs

Intel 7900x OCd to 4700ghz @ 1.238v
Asus Rampage Extreme VI
64GB GSkill DDR4 Ocd to 3600mhz
Samsung EVO 960 256GB on the PCH NVME slot
Kingston 240 GB NVME on the DIMM.2 Slot
2 x Samsung 850 1TB SATA SSD's
2 x GTX 1080TI SLI
Corsair 1500Axi PSU
Corsair Obsidian 900D
Enermax pump
2 x 360 RAD's and 1 x 240 RAD. Single loop
Heatkiller IV CPU Waterblock
EKWB 1080ti waterblocks

Screen shots below. The pic with the higher score is at 3200mhz cache.





Thanks a ton

[Moloch]

While testing a few things, I set the cache to 3 Ghz down from 3.2Ghz @1.155v. Though the system worked perfectly still and made no noticeable difference, the CPUZ score fell by 1200 points to 5700-5800. I did not know at this time that this was due to the what reason.

Baffled much, I tried everything I knew to fix this (thanks to my OCD :-) )

And suddenly I realized setting the cache to 3200Mhz upped the scores back to 7100 and a little better on other benchmarks.

So my question is, can cache speed make such a big difference in speeds? Is this possible? or a bug? I am sorta baffled still.

That seems like a huge difference from 3000 to 3200, but I'm no expert on cache.

My personal experience was having ~85GB/s RAM bandwidth with default 2400 cache speed. When I bumped it to 3200, I got 105GB/s... 25% increase in RAM bandwidth from a 33% increase in cache speed

So yeah, I'd guess that is somewhat normal, though I wouldn't expect to see a 20% change in benchmark with only a 5% change in cache speed. That seems like too much... that's similar to what I got moving from 2400 to 3200 (though I did not test to see where the gains were made, it could have been 90% in the 3000-3200 range)

[Raja@ASUS]

Real world gains are often very small. However, it's worth taking what's on the table, without needing to increase voltages excessively.

[Addiecool]

That seems like a huge difference from 3000 to 3200, but I'm no expert on cache.

My personal experience was having ~85GB/s RAM bandwidth with default 2400 cache speed. When I bumped it to 3200, I got 105GB/s... 25% increase in RAM bandwidth from a 33% increase in cache speed

So yeah, I'd guess that is somewhat normal, though I wouldn't expect to see a 20% change in benchmark with only a 5% change in cache speed. That seems like too much... that's similar to what I got moving from 2400 to 3200 (though I did not test to see where the gains were made, it could have been 90% in the 3000-3200 range)

yup. Major increase seems in the higher cache speeds. Also improves gaming performance in some games.*

[Addiecool]

Real world gains are often very small. However, it's worth taking what's on the table, without needing to increase voltages excessively.

agree. I also find it exciting to keep discovering new things about this platform*

[jab383]

Raja hit the key. Yes, cache speed in relation to memory speed can improve or reduce benchmark scores. The point of diminishing returns is where the cache can handle all the memory data traffic - after that higher cache clock makes very little difference. Where it hurts with the current crop of Skylake-X is that higher cache clocks and the required voltage burn a lot of power and heat. As you raise cache clock, watch CPU package temperature.

[Addiecool]

Raja hit the key. Yes, cache speed in relation to memory speed can improve or reduce benchmark scores. The point of diminishing returns is where the cache can handle all the memory data traffic - after that higher cache clock makes very little difference. Where it hurts with the current crop of Skylake-X is that higher cache clocks and the required voltage burn a lot of power and heat. As you raise cache clock, watch CPU package temperature.

Found the sweet spot at 3.2ghz with 1.16v. Very little or no temp increase. System is rock solid

[ExcessiveGBH]

yup. Major increase seems in the higher cache speeds. Also improves gaming performance in some games.*

Went from default 2.4GHz to 3.2GHz and had a 10% performance in 3 games I tested, GTA5, Tomb Raider 2013 and Rise of Tomb Raider DX12.

Here is GTA5 benchmark @4K

z.

[FaaR]

Hi

Hi!

4.7Ghz @ offset of 0.012v
3.2Ghz Cache @ 1.16v
DRAM at 3600Mhz @1.35v (16-16-16-32)
Rest everything I set to Auto

Interesting... Where, in what software do you set these parameters? I took a quick peek in UEFI, and there was no direct cache speed setting that I could find, only an offset setting, which I did not dare touch.

Also, do you run all cores at 4.7GHz? I don't watercool, so there's like no way I could pull that off. OCing used to be so simple back in the day, now there's millions of options and settings, and many different turbo modes.

Like, my CPU (7900X) is IIRC 3.3GHz base frequency on all cores, but it runs AVX loads at 3.5GHz all cores (Prime95 max burn test), and apparently non-AVX loads at 4GHz on all cores (I used Folding@Home for some stress testing), so I'm a bit confuzzled here! Then there's turbo 2.0 and 3.0 on top of that and how they work and on how many cores they apply to is even sketchier... I was thinking maybe I could get away with bumping all the various turbo levels like two bins without having to raise volts any and watch CPU temp skyrocket, maybe three for the turbo boost Max 3.0 to 4.7 or 4.8GHz for what, 4 cores? Not sure how to accomplish that though or if it is even possible to be so specific.

Thanks for any help and input!

[RTitan]

Another sample point for all on my X299 Prime Deluxe. I am running 4.8 (Silicon Lottery 4.7 purchase) using 1.3v. Running 4000Mhz memory. I'm on water using Koolance's best. Still cannot contain heat!

Addiecool's sample is the only one I could find in regards to CPU Cache voltage. That is where I started - thanks. I'm at 3.1 at 1.185v (manual setting). I have a 25% increase in performance via CPU-Z - 7285. Linx w/AVX is only 5.5% increase - @ 275 GFlops. I can only run Linx w/AVX approximately 10 passes before it scares me - not real work load for my work anyway. Still no throttling or critical tempsI I have AVX down 300 Mhz as per Silicon Lottery specs. So really I'm running 4.5 on AVX.

Would love to see more samples and examples of cache adaptive or offset mode. I have no idea where to start there. And I cannot find what the defualt cache voltage is.