Testing ARROW lake - power limited

[BoggledBeagle]

This is a sweet range, not a spot. But I agree that some voltage tuning had to take place.

Speaking abou sweet ranges, you can see that from cca 160W performance rises just slowly.



Points per watt go from 110 to 570, that is 460 width.
If you wanted to avoid the worst third, you would start from 110+153, 263 corresponds to 120W.

So if you employ intuitive graphical approach or formulate some rule - e.g. avoid the worst third of the perf/watt range, you end up with the upper limit between 120 and 160W.

I do not believe that it is reasonable to choke the CPU so much that it does less than half of the maximal output.

 

[evernessince]

From testing both a 13900k and a 14900k I don't think this is remotely true. You don't have much room for undervolting in either one of them, which means they are already running at the lowest voltage possible. Compared to 12900k, which I could easily drop from the 1.27 stock voltage to 1.06.

You almost certainly have bad silicon quality with your 13900K and 14900K but not with your 12900K. Derbauer tested voltage tuning on high end Intel CPUs like the 13900K and 14900K and found that there was a wide variance chip to chip but on average there was a good amount of headroom so long as you don't get unlucky.

Another example of a chip with high amount of variance was a piece Derbauer did on the Ryzen 1600 (it was awhile ago, not sure if it was the AF, X, or non-X model) and he found that there was a significant difference in the power consumption (20 - 25w). That's notable given the chip maxes at 142w.

So if you employ intuitive graphical approach or formulate some rule - e.g. avoid the worst third of the perf/watt range, you end up with the upper limit between 120 and 160W.

I do not believe that it is reasonable to choke the CPU so much that it does less than half of the maximal output.

Yes, it's a bit extreme to limit it that much. That's where UV would come in, it might enable you to achieve higher performance while having greater perf per watt

This is a sweet range, not a spot.

I'm just calling it by it's official name but I've no qualms in regards to whichever term you prefer.

 

[BoggledBeagle]

Here is a comparison of the three CPUs:



From this video:

 

[evernessince]

Here is a comparison of the three CPUs:

View attachment 370393

From this video:

The figures in that graph seem off. Other review outlets have the 285K roughly equal to the 14900k in the same R23 bench: https://www.notebookcheck.net/Intel...-Ryzen-9000-in-power-efficiency.901799.0.html

The difference between the 9950X and 285K seems accurate but there should not be a huge gap between 14900K and the recently released AMD and Intel CPUs.

These processors are very close in CB 2024 Multi as well:

I don't know what to make of this, perhaps it's the result of performance variance as a result of windows issues. That's been a PITA for reviewers with this new CPU.

 

[JustBenching]

You almost certainly have bad silicon quality with your 13900K and 14900K but not with your 12900K. Derbauer tested voltage tuning on high end Intel CPUs like the 13900K and 14900K and found that there was a wide variance chip to chip but on average there was a good amount of headroom so long as you don't get unlucky.

Another example of a chip with high amount of variance was a piece Derbauer did on the Ryzen 1600 (it was awhile ago, not sure if it was the AF, X, or non-X model) and he found that there was a significant difference in the power consumption (20 - 25w). That's notable given the chip maxes at 142w.

Well sure but if there is such a huge variance what's the point of undervolting products to compare? FYI most 12900k were like mine (I don't have a particular good chip, SP 93 is ~average) with tons of UV headroom. Most 13/14900k i've seen in my ocing community were all crap. Only the KS parts were actually decent undervolters.

Here is a comparison of the three CPUs:

View attachment 370393

From this video:

Look at that 50-100w performance, that's why I asked you to test 65 before. This looks insane if it's actually accurate.

 

[BoggledBeagle]

Look at that 50-100w performance, that's why I asked you to test 65 before. This looks insane if it's actually accurate.

Yes, the Arrowlake improvement looks good in that range, too bad that there are no real data points on the curves similar to mine, we cannot judge how accurate the curves are.

 

[phanbuey]

Im just waiting for the obligatory screenshot where this gets rekt by 70W M4.

 

[JustBenching]

Im just waiting for the obligatory screenshot where this gets rekt by 70W M4.

In MT workloads, it ain't happening. Even though the m4 pro will be much bigger than the 285k, I don't think it's winning in anything MT.

 

[phanbuey]

In MT workloads, it ain't happening. Even though the m4 pro will be much bigger than the 285k, I don't think it's winning in anything MT.

M4 Max Makes Short Work Of 24-Core Intel Ultra 9 285K, Outperforming It By 17 Percent In Single-Core And Nearly 16 Percent In Multi-Core Performance

IDK - these leaks are pretty nasty.

 

[JustBenching]

M4 Max Makes Short Work Of 24-Core Intel Ultra 9 285K, Outperforming It By 17 Percent In Single-Core And Nearly 16 Percent In Multi-Core Performance

View attachment 370410

IDK - these leaks are pretty nasty.

This is geekbench man. The MT portion of geekbench is useless, it doesn't scale past 12-14 THREADS (not cores). You can compare the score of a 9700x vs a 9950x for example.

The m3 max for example is matching my 12900k in Geekbench 6 MT, but then you try cinebench and yeah, nope. There is no way in hell the m4 max is getting anywhere near 40k+CBR23. I'd be amazed if it went past 30k to be fair.