AMD RYZEN 2950 Speed issue

[Joost Buijs]

Now I get the following speed. Is it normal? I somehow feel that the speed gain should be more than 8k.

Do you realize you have 16 physical cores? You can'r run 2x faster with hyperthreading. ~1.3x is normal.

With Intel hyperthreading you get ~1.3 x the performance. Although I never tried I have a strong feeling that with AMD hyperthreading you get nothing at all. Maybe I'm totally wrong.

I wouldn't say it's nothing.
I had ran some benchmarks with stockfish 10,all 64 mB on amd tr 1950x.
Best

Still difficult to tell if you don't compare with 32 threads and SMT off. 50% gain like the graph shows seems a bit exaggerated, maybe stockfish has a tendency to show a somewhat higher nps when you enable more threads. I would at least test it with several other engines. Maybe AMD SMT is really that good, but I doubt it. Unfortunately I can't test it myself because I only have Intel CPU.

[mwyoung]

Now I get the following speed. Is it normal? I somehow feel that the speed gain should be more than 8k.

Do you realize you have 16 physical cores? You can'r run 2x faster with hyperthreading. ~1.3x is normal.

With Intel hyperthreading you get ~1.3 x the performance. Although I never tried I have a strong feeling that with AMD hyperthreading you get nothing at all. Maybe I'm totally wrong.

I wouldn't say it's nothing.
I had ran some benchmarks with stockfish 10,all 64 mB on amd tr 1950x.
Best

Still difficult to tell if you don't compare with 32 threads and SMT off. 50% gain like the graph shows seems a bit exaggerated, maybe stockfish has a tendency to show a somewhat higher nps when you enable more threads. I would at least test it with several other engines. Maybe AMD SMT is really that good, but I doubt it. Unfortunately I can't test it myself because I only have Intel CPU.

Yes, Stockfish and any other chess program will always exaggerate NPS when given more threads.
That is why you should never measure NPS in this testing. You could gain NPS, but still be slower.

The correct why to measure in this testing is by using Time to depth.

And when measured correctly SMT works better then SMT off.

[Joost Buijs]

Now I get the following speed. Is it normal? I somehow feel that the speed gain should be more than 8k.

Do you realize you have 16 physical cores? You can'r run 2x faster with hyperthreading. ~1.3x is normal.

With Intel hyperthreading you get ~1.3 x the performance. Although I never tried I have a strong feeling that with AMD hyperthreading you get nothing at all. Maybe I'm totally wrong.

I wouldn't say it's nothing.
I had ran some benchmarks with stockfish 10,all 64 mB on amd tr 1950x.
Best

Still difficult to tell if you don't compare with 32 threads and SMT off. 50% gain like the graph shows seems a bit exaggerated, maybe stockfish has a tendency to show a somewhat higher nps when you enable more threads. I would at least test it with several other engines. Maybe AMD SMT is really that good, but I doubt it. Unfortunately I can't test it myself because I only have Intel CPU.

Yes, Stockfish and any other chess program will always exaggerate NPS when given more threads.
That is why you should never measure NPS in this testing. You could gain NPS, but still be slower.

The correct why to measure in this testing is by using Time to depth.

And when measured correctly SMT works better then SMT off.

Even 'Time to depth' doesn't mean a thing with a very selective program like Stockfish. NPS and 'Time to depth' both depend upon the shape of the search tree, and changing the number of threads will always alter the shape of the search tree. Chess programs are not very well suited to determine true hardware performance.

[viditchess]

P.S- When I run Stockfish at 16cores and LC0 with 2 cores. The temp goes up to 68-70. Is it normal? I use Kraken x62 liquid cooling.

I am running at 32 threads, with Lc0. My current temp is 63c.

When I run with only 16 threads, and Lc0. My temp runs 55c.

As I said before you are hitting the TJmax of 68c.

You need a better AIO.

Enermax liqtech tr4 ii 280 is designed for the 2950x with a full coverage water block. This is what you need to be running.

Without proper cooling you will never see the full performance of the 2950x. You will also need many computer case fans. I run 6 case fans, and + 2 for the AIO...

You must remember our systems will pull over 600 watts, you need proper cooling.

Wow! I regret going for the recommendations of my local tech guy. I tested the temperatures again. With 16 cores and lc0, the temp went up to 67. Even with 29threads and lc0, the temperature was around 67-68. I assume that the system is limiting its resources as it has hit a high core temperature.When idle the temp is around 35 degrees
Any precautions I should take in such scenarios?

[mwyoung]

P.S- When I run Stockfish at 16cores and LC0 with 2 cores. The temp goes up to 68-70. Is it normal? I use Kraken x62 liquid cooling.

I am running at 32 threads, with Lc0. My current temp is 63c.

When I run with only 16 threads, and Lc0. My temp runs 55c.

As I said before you are hitting the TJmax of 68c.

You need a better AIO.

Enermax liqtech tr4 ii 280 is designed for the 2950x with a full coverage water block. This is what you need to be running.

Without proper cooling you will never see the full performance of the 2950x. You will also need many computer case fans. I run 6 case fans, and + 2 for the AIO...

You must remember our systems will pull over 600 watts, you need proper cooling.

Wow! I regret going for the recommendations of my local tech guy. I tested the temperatures again. With 16 cores and lc0, the temp went up to 67. Even with 29threads and lc0, the temperature was around 67-68. I assume that the system is limiting its resources as it has hit a high core temperature.When idle the temp is around 35 degrees
Any precautions I should take in such scenarios?

No precaution, the chip is protecting itself. But is why you are running to less then full potential.

Your system is not broken, just not designed optimally.

You just need to be able to remove the heat from your system.

[Michael Sherwin]

I read all the replies and did not see the biggest issue of the 2950x addressed. On the 2950x 16 cores are directly connected to the memory bus. The other 16 cores are connected to the memory bus through the first 16 cores. This creates a memory bottleneck for the 2nd 16 cores. Chess programs are big data munchers. Magic bitboard engines are going to run the slowest on the 2950x. That is why Crafty gains virtually nothing from the 2nd 16 cores. SF is not magic and does better than Crafty. One thing you can try is decreasing the size of the SF hashtable.

[ankan]

I read all the replies and did not see the biggest issue of the 2950x addressed. On the 2950x 16 cores are directly connected to the memory bus. The other 16 cores are connected to the memory bus through the first 16 cores. This creates a memory bottleneck for the 2nd 16 cores. Chess programs are big data munchers. Magic bitboard engines are going to run the slowest on the 2950x. That is why Crafty gains virtually nothing from the 2nd 16 cores. SF is not magic and does better than Crafty. One thing you can try is decreasing the size of the SF hashtable.

I could be wrong, but I thought 2950x doesn't have that problem. It only has 16 cores (32 threads with HT). I think you are confusing it with 2990wx which is a 32 core CPU (64 threads with HT) but memory is connected only to 16 cores.

[Michael Sherwin]

I read all the replies and did not see the biggest issue of the 2950x addressed. On the 2950x 16 cores are directly connected to the memory bus. The other 16 cores are connected to the memory bus through the first 16 cores. This creates a memory bottleneck for the 2nd 16 cores. Chess programs are big data munchers. Magic bitboard engines are going to run the slowest on the 2950x. That is why Crafty gains virtually nothing from the 2nd 16 cores. SF is not magic and does better than Crafty. One thing you can try is decreasing the size of the SF hashtable.

FOR ENGINE AUTHORS
Something that I posted about before that might have a practical use case for SMP processors with memory bottlenecks is the following. I propose dedicating one thread just to the hashtable. All other threads would request hashtable data from the dedicated thread. That information can be requested at the end of MakeMove() so it will be waiting in the global structure when needed. The dedicated thread will poll the global structure for storage and retrieval request. This would eliminate any race conditions and would not suffer the performance hit of using locks. It would scale better up to the maximum number of threads the dedicated thread could service.

[Michael Sherwin]

I read all the replies and did not see the biggest issue of the 2950x addressed. On the 2950x 16 cores are directly connected to the memory bus. The other 16 cores are connected to the memory bus through the first 16 cores. This creates a memory bottleneck for the 2nd 16 cores. Chess programs are big data munchers. Magic bitboard engines are going to run the slowest on the 2950x. That is why Crafty gains virtually nothing from the 2nd 16 cores. SF is not magic and does better than Crafty. One thing you can try is decreasing the size of the SF hashtable.

I could be wrong, but I thought 2950x doesn't have that problem. It only has 16 cores (32 threads with HT). I think you are confusing it with 2990wx which is a 32 core CPU (64 threads with HT) but memory is connected only to 16 cores.

Yes, you are correct. My memory is failing me more often these days. I need to double check everything I guess.

[viditchess]

Hello!!

Unfortunately, I seem to face the same issue again. I have attached the screenshot for the reference. Somehow the virtual cores aren't being utilized.

Does anyone know a solution for it?