schack wrote:In other words, the NPS / overhead tradeoff with Intel is barely worth it, but with the Ryzen chips, it's entirely so? Is that what these numbers would indicate?
Yes, definitely. Even on my i7 4790 Intel, SF and Komodo do seem to benefit a bit from hyperthreading. With Ryzen that will be pronounced.
Upper test of nps in SF and Komodo are almost certainly done with HT on. This almost certainly means i7-6900k will be still noticeably stronger than R7 1800X since its no-HT performance is better and for 8 core i7 no-HT is equal or better than HT on in terms of Elo.
Yes, but just by a bit. It can be reformulated as the following: in computer chess single thread performance is more important than the hyperthreading benefit. In any case, Ryzen 1800X on full charge (probably all 16 threads) is sensibly equal to i7-6900K on full charge (probably 8 or 10 threads), because as NPS goes, Ryzen seems even superior (in 16 versus 16 threads).
That is remarkable, because the price is less than half. Also, Stockfish and Komodo seem the applications best suited to Ryzen, where it equals i7-6900K. In 3D gaming the things are different, Ryzen can lose even to an i5, but the things are so inclined towards single thread there that a 4 core i7-7700K is better than a 10 core i7-6950X.
Laskos wrote:Naples is top of the line "enterprise chip" that is designed for the server market. My guess is it will cost at least $2,000. New Xeons are much more expensive than that core per core, I never dreamt of having a new 16-core dual Xeon, was just looking at second hand, but never managed to assemble one. I will go for Ryzen, these are great news.
I built a config that is very popular around here of dual e5-2670, and it costed me together with 32GB DDR3 and Mobo under 600$ (only Mobo was new, CPUs and memory used).
These Xeons are almost 5 years old now, but still 16 core machine (32T) with them gives 25Mnps running SF8 where i7-6950X on 10 cores (20T) gives 19Mnps and price difference is 4-5x.
That was the deal of the century. I was seeing your posts, but never managed to buy anything in this sense. My 4 core i7 is more expensive than that....
I saw it yesterday, they screwed the test completely. Seem to use 1 thread. If it is 1 thread, the result shows that Ryzen thread is faster than the thread of 5960X, which in itself would be very good, because multithreading is better with Ryzen.
I saw it yesterday, they screwed the test completely. Seem to use 1 thread. If it is 1 thread, the result shows that Ryzen thread is faster than the thread of 5960X, which in itself would be very good, because multithreading is better with Ryzen.
From the very professional French site www.hardware.fr I collected the relevant articles with Stockfish and Komodo:
I saw it yesterday, they screwed the test completely. Seem to use 1 thread. If it is 1 thread, the result shows that Ryzen thread is faster than the thread of 5960X, which in itself would be very good, because multithreading is better with Ryzen.
From the very professional French site www.hardware.fr I collected the relevant articles with Stockfish and Komodo:
I saw it yesterday, they screwed the test completely. Seem to use 1 thread. If it is 1 thread, the result shows that Ryzen thread is faster than the thread of 5960X, which in itself would be very good, because multithreading is better with Ryzen.
It is 1 thread. That benchmark just uses the "bench" command of Stockfish. The output displayed is the ms it took for Stockfish to complete that command.
I have an intel i7 7700 (not the K version), on my machine it took slightly less than 3 s to complete the "bench".
Just wondering if my thinking seems correct regarding Ryzen.
The single or dual engine (no pondering) total performance is very good with better HT performance for engines with strong parallel search (>4). However, for multi-engine tournaments/testing, wouldn't the variability with HT make it suspect for more than 7 engines at once?
Of course, price/performance would be another not insignificant factor in Ryzen's favor.
PS I was a huge AMD fan with Opteron and x64 (recall the Itanium IA-64 issues) and even have an early dual socket system that should still boot...
brianr wrote:Just wondering if my thinking seems correct regarding Ryzen.
The single or dual engine (no pondering) total performance is very good with better HT performance for engines with strong parallel search (>4). However, for multi-engine tournaments/testing, wouldn't the variability with HT make it suspect for more than 7 engines at once?
Of course, price/performance would be another not insignificant factor in Ryzen's favor.
PS I was a huge AMD fan with Opteron and x64 (recall the Itanium IA-64 issues) and even have an early dual socket system that should still boot...
If one is using only physical cores (say 8 parallel matches, ponder OFF), Intel 8 core i7-6900K is 10% better speed-wise per core than 8 core Ryzen 1800X. If one is using all logical cores (15 parallel matches, to leave one thread for OS) as I often do, then Ryzen is better speed-wise by 10% than Intel. Variability with HT seems to be a bit of a myth to me, after reading articles on how cache behaves with jumping threads, the important thing is that threads in Windows jump so often, that averaging over even 1 second, there is no difference between threads and hyperthreads for chess engines.