Performance Impact of EDC on Ryzen CPUs

[Mr Evil]

Seeing in HWiNFO64 that my CPU was hitting 100% of the EDC limit, I thought that clearly the clock speed is being held back, and so I would get better performance by raising the limit. Thus begins the journey.

Setting EDC higher than the default 140A didn't have the effect I expected, so I decided to run some tests. I tested with EDC from 75-200A in 25A increments, plus the default 140A. TDC and PPT were left at the default values (95A and 142W). The Curve Optimizer was set to -15 all cores. The test system consists of a 5950X in an Asus CH8DH, and 64GB of RAM @ 3600 CAS 18.

I used the CPU-Z benchmark, as it's quick and I'm lazy. Ten runs at each value of EDC looked like this...

I offset the two series on the x-axis slightly so the points don't overlap, which would make it hard to read. I added quadratic lines of best fit, though the drop on the left looks too sharp to be quadratic.

Counterintuitively, raising EDC actually reduced both single and multi-threaded performance. I hypothesize that this is because the highest current peaks are when the CPU is least efficient, so it gains only a little performance from that, while the increase in heat reduces clock speed long-term. If I can overcome my laziness I might test that.

 

[Mr Evil]

Should do a few more runs and plot it as a box plot
To my naked eye, 100-125A works out the best?

A box plot would be good. I would have to be quite bored to take enough measurements for one though. Yes, it looks like the best performance is somewhere at the lower end. I wonder if the drop below 100A is related to TDC being 95A?

On my 5950X I found that setting the EDC to 160A in PBO lowered temperatures without affecting performance.

Do you know what the default was? It's supposed to be 140A, but some motherboards default to having PBO enabled, which puts EDC at some very high value, above 160A.

 

[Mr Evil]

what cooler do you have?

It's a Thermalright Le Grand Macho RT. You can see how it performs here: https://www.overclockers.co.uk/forums/threads/ducting-case-fan-to-cool-cpu.18916895/

i think your curve looks something like this:

I'm going to have to collect more data. *sigh*

 

[Mr Evil]

Single thread going down at higher EDC must be because it runs right after multithread (and now hotter).

That's a really good point. I should be able to verify that easily by setting the multithread benchmark to 1 thread.

 

[Mr Evil]

To try to pinpoint the peak of the curve, to see if it's the shape that tamzzy thinks, I ran more benchmarks at 80, 90, 110 and 120A.

From the multi-thread score, it looks like it's at 100A. The single-thread scores are noisier today than they were yesterday. There are a lot more outliers <690 points, and they occured randomly throughout testing, not towards the end as they would if it was caused by heat.

I then tested alec's theory by setting the multi-thread benchmark to 1 thread. Since the multi and single-thread scores are the same then, I plotted both sets together. If the theory is correct, the curve should be flatter, and the peak higher.

Noisy again (both the highest and lowest scores so far are on this chart), but it looks consistent with the theory.

Another prediction from that theory would be that the single-thread scores would be lower than the 1T-multi-thread scores, as the CPU will be hotter on the second run, even if it's less than when the multi-thread benchmark is using all cores. I plotted the diff between the two scores as a box plot this time, because.

It's definitely biased towards a negative difference, which again is consistent with the theory.

It would be possible to test this further by repeating the whole set of tests with the CPU fan set to different fixed speeds to vary the CPU temperature. At higher speeds, the multi-thread score should go up, and the optimum EDC should go up too. That would be a lot of work though.