AMD Ryzen 9 7950X - Benchmarks

[Raphexon]

Long story short: I guess when it comes to multi-threaded engine analysis (lets say 5 min/move or overnight even), the improvements and MORE efficiency cores added (even at a slightly lower clock speed) results in a higher NPS for the new Intel 13th gen processors compared to the new (or last gen) AMD Zen 4 processors?

I remember the previous Zen 3 knocking the socks off the 12th gen Intel processors in NPS...not anymore.

No.

https://openbenchmarking.org/test/pts/stockfish

[j.t.]

Long story short: I guess when it comes to multi-threaded engine analysis (lets say 5 min/move or overnight even), the improvements and MORE efficiency cores added (even at a slightly lower clock speed) results in a higher NPS for the new Intel 13th gen processors compared to the new (or last gen) AMD Zen 4 processors?

I remember the previous Zen 3 knocking the socks off the 12th gen Intel processors in NPS...not anymore.

No.

https://openbenchmarking.org/test/pts/stockfish

I am not sure if this website reports realistic results, at least regarding the i9-13900K. It shows ~43 Mnps but when I compile with "make profile-build ARCH=x86-64-avx2", it reaches 74 Mnps (opposed to the 75 Mnps of the 7950X the website is saying) with "./stockfish bench 128 32 24 default depth", which is if I see correctly the approach that this website claims to use.

[CornfedForever]

Long story short: I guess when it comes to multi-threaded engine analysis (lets say 5 min/move or overnight even), the improvements and MORE efficiency cores added (even at a slightly lower clock speed) results in a higher NPS for the new Intel 13th gen processors compared to the new (or last gen) AMD Zen 4 processors?

I remember the previous Zen 3 knocking the socks off the 12th gen Intel processors in NPS...not anymore.

No.

https://openbenchmarking.org/test/pts/stockfish

I appreciate your input.
Right, I think we've all seen that. The numbers are...a bit confusing and appear to be (maybe?) all cores running 100%. That's not...exactly reality. For real your average chess enthusiasts purposes it probably would not matter which CPU you were to go with.


From the chart:

Processor -- Percentile-- NPS (avg)
AMD 5950x 16 core - 76th 54,002,548
AMD 7950x 16 core - 81st 75,489,026 +28.5% gain
Intel i9 12900k -70th 41,571,222
Intel i9 13900k - 70th 43,369,647 +4.25% gain

My understanding is that the extra e-cores + other improvements product as 45% or so multi-threaded gain. So...why does the intel #'s show so little gain?

Lets talk 'real world' for a moment as in how the average chess person would use their engines. That's normally for analysis and while doing other things with their computer.

Lets say I am running Chessbase and doing database work...maybe evaluating an opening or annotating a game while an engine(s) chews on a position for a few minutes mostly per move (or various positions overnight even...but rarely) - maybe using 4 cores a piece. Maybe I have a website or two open, following a football game...typing an email. I mean, Stockfish NNUE basically will offer nothing of importance after a very short period of time anyway...you get to a certain depth and probably just need to end its search.

Like I said...I am not sure one CPU is much different from another in such real world situations, but the lack of 'increase' in multi-threaded applications puzzles me a bit.

[Modern Times]

Yeah, in the real world it really doesn't matter I think. You'd be happy with either.

[CornfedForever]

Since they run so hot, I am curious about dropping the Zen 7950 and the I9 - 13900 to 105W ECO mode for...I guess it would be best used primarily longer analysis - and how much performance drop we would see regarding chess engines.

Obviously, this would save a lot of energy and heat on the CPU; but I wonder if there is a formula or way to measure the performance loss - even a rule of thumg? I am not sure how it would be measured...NPS I presume (?) Maybe something that would take 10 min at full throttle would need 12 min to accomplish the same amount of work in ECO mode?

[smatovic]

Since they run so hot, I am curious about dropping the Zen 7950 and the I9 - 13900 to 105W ECO mode for...I guess it would be best used primarily longer analysis - and how much performance drop we would see regarding chess engines.

Obviously, this would save a lot of energy and heat on the CPU; but I wonder if there is a formula or way to measure the performance loss - even a rule of thumg? I am not sure how it would be measured...NPS I presume (?) Maybe something that would take 10 min at full throttle would need 12 min to accomplish the same amount of work in ECO mode?

Let's say ~60% less power usage, ~20% less performance, ~20% less NPS, ~1.2x more compute time for your analysis, so yes, ~12 minutes, depends on the kind of analysis I guess (parallel speedup).

Hmm, "underclocking/undervolting" seems to be an option, more cores with less frequency - less power usage, or alike.

It's all about the efficiency sweet spot of a CPU. In today's benchmark races for marketing, both AMD and Intel configure their stock settings far away from the efficiency maximum. Means, you can reduce the power draw massively while only giving up relatively little performance. For example, the 7950X at 90W uses 62% less power than at 230W, but is only 18% slower.

It's not even underclocking or undervolting, just selecting profiles in the bios, or if it doesn't have that, using PPT (package power target) / EDC (electrical design current) / TDC (thermal design current) for which the profile is just a shorthand.

--
Srdja

[Vinvin]

Since they run so hot, I am curious about dropping the Zen 7950 and the I9 - 13900 to 105W ECO mode for...I guess it would be best used primarily longer analysis - and how much performance drop we would see regarding chess engines.

Obviously, this would save a lot of energy and heat on the CPU; but I wonder if there is a formula or way to measure the performance loss - even a rule of thumg? I am not sure how it would be measured...NPS I presume (?) Maybe something that would take 10 min at full throttle would need 12 min to accomplish the same amount of work in ECO mode?

Some benchmark on ECO mode :



https://www.overclock3d.net/reviews/cpu ... ead_this/4

https://www.anandtech.com/show/17585/am ... igh-end/20 -> Ryzen 9 7950X at 65 W (ECO Mode): Zen 4 has Superb Efficiency

[CornfedForever]

Since they run so hot, I am curious about dropping the Zen 7950 and the I9 - 13900 to 105W ECO mode for...I guess it would be best used primarily longer analysis - and how much performance drop we would see regarding chess engines.

Obviously, this would save a lot of energy and heat on the CPU; but I wonder if there is a formula or way to measure the performance loss - even a rule of thumb? I am not sure how it would be measured...NPS I presume (?) Maybe something that would take 10 min at full throttle would need 12 min to accomplish the same amount of work in ECO mode?

Let's say ~60% less power usage, ~20% less performance, ~20% less NPS, ~1.2x more compute time for your analysis, so yes, ~12 minutes, depends on the kind of analysis I guess (parallel speedup).

EXACTLY the kind of analysis I was looking for - THANK YOU SO MUCH!!
Sounds like a really good option for the top new AMD processors.

I guess that gets more complicated when it comes to the top Intel CPU given the mixture of 'P' and 'E' cores...I'll see if I can do some research on that.

[CornfedForever]

Some benchmark on ECO mode :



https://www.overclock3d.net/reviews/cpu ... ead_this/4

https://www.anandtech.com/show/17585/am ... igh-end/20 -> Ryzen 9 7950X at 65 W (ECO Mode): Zen 4 has Superb Efficiency

I checked the 2nd of these out again and at the bottom, the higher performance Eco code...does not seem to be an option as the Ryzen Master does not allow it. I believe you can adjust the BIOS every single time you want 105W Eco mode...but that would be a pain. Perhaps you can 'set a bios option' so it's not so painful. But, I am wondering if you can set...say just 4, 8 or whatever cores to 105W....and keep the others at full 170W...or if they ALL have to be set to the 'mid' power levels...

AMD, if you are reading this, a 105W Eco mode option for Ryzen Master would be a very useful option for Ryzen 9 7900X and 7950X users. Your current 65W mode leaves a lot of multi-threaded performance on the table, and while this mode is useful, there are a lot of users who will not want to be as aggressive with their power savings. We love Eco Mode, but we would like to see more balance here. That's it, we need a "Balanced" mode for Ryzen 9 series CPUs in Ryzen Master! Eco Mode 65W, Balanced Mode (105W), and Default Mode (170W). Please AMD, just make it happen!

[Modern Times]

I use Ryzen Master on my Zen3 5900X. It is incredibly easy, you don't actually need an ECO setting although it would be clearer. I change just one number - the PPT - down to 105W from 142W and that is it. Takes just seconds. No inconvenience at all. So on a 12 core workload, that stops the CPU from drawing lots of extra voltage for diminishing returns, and results in a quieter, cooler machine.

Zen 4 carried over Zen 3's propensity for pushing the voltage to increase performance but goes much further it seems. You can achieve an ECO mode to your liking by just by altering one number in Ryzen Master.