Moderator: Ras
Code: Select all
Blocks Time Ratio Efficiency
28 453 20.66004415 0.73785872
14 750 12.47866667 0.891333333
7 1406 6.656472262 0.950924609
4 2406 3.889858687 0.972464672
1 9359 1 1
Average 0.910516267
If that is what you wrote, then I misunderstood. The correct comparison is always one cpu vs N cpus, where the cpus are identical, If that is what you did, you did it correctly. My reading suggested that you ran 41x faster on 100+ cores than on just one, and then you corrected that based on a clock speed difference...Daniel Shawul wrote:What are you missing? If I launch one cuda thread vs many threads that engage the whole device, i get something close to 100. That tells you the efficiency of your implementation compared to comparing to some random CPU.
That is allright. I post to get feed backs here. And I may not have been so clear with my setup and what I am trying to achieve with it.If that is what you wrote, then I misunderstood. The correct comparison is always one cpu vs N cpus, where the cpus are identical, If that is what you did, you did it correctly. My reading suggested that you ran 41x faster on 100+ cores than on just one, and then you corrected that based on a clock speed difference...
It appears to be my error... sorry...
Hello vincentdiep wrote:which type gpu are you using?
programming in CUDA or in OpenCL?
The famous 5 second rule of windows before it is hung...Daniel Shawul wrote:Hello vincentdiep wrote:which type gpu are you using?
programming in CUDA or in OpenCL?
My GPU is Quadro FX 3700 with 14 SM = 112 cores. The other details are in the result I posted somewhere in this thread. It is used for display as well so now and then I have to boot up when I get the blue screen..I use cuda for programming. For now I am just experimenting with simpler games. So there is nothing to report on chess yet except that it runs.
That is good. GPUs don't have SSE instructions so I suppose your speedup must have come from memory optimizations.Hey Daniel,
just want to mention that i doubled my move gen performance simply by using a vector datatype "long4" for the QuadBitboards.
Yes legality testing can be a killer. For my GGP engine Nebiyu I got a 100% speedup by allowing the king to be captured. If your attacks is slow you may want to try that.Running a simple loop with the starting position i get about 300Knps per SIMD Unit,
if i turn of legality check it is up to 1Mnps.
I don't have a full move generator but a random legal move generator. And I also don't store moves in global memory unlike you. Your engine is probably memory bound because of that. 300k nps is not that much when you divide it by the number of threads in a SIMD unit. When you decide to give up relying on registers (& shared mem), you should bear in mind that the kernel may not be faster than a single cpu eventually. Hashtables are typical examples that are better generated on cpu. Well chess probably is not suited for gpu computation because monte-carlo misses a lot of tactics, but my effort is concentrated on MCTS and games that are suitable for that approach. The MC part is really needed to harness the power of the gpu and hide global memory latency as well.How does your move generator perform?
You're buliding a chessprogram on a gpu as well Srdje?smatovic wrote:Hey Daniel,
just want to mention that i doubled my move gen performance simply by using a vector datatype "long4" for the QuadBitboards.
Running a simple loop with the starting position i get about 300Knps per SIMD Unit,
if i turn of legality check it is up to 1Mnps.
How does your move generator perform?
--
Srdja
You just shouldn't use monte carlo nor UCT for chessDaniel Shawul wrote:That is good. GPUs don't have SSE instructions so I suppose your speedup must have come from memory optimizations.Hey Daniel,
just want to mention that i doubled my move gen performance simply by using a vector datatype "long4" for the QuadBitboards.Yes legality testing can be a killer. For my GGP engine Nebiyu I got a 100% speedup by allowing the king to be captured. If your attacks is slow you may want to try that.Running a simple loop with the starting position i get about 300Knps per SIMD Unit,
if i turn of legality check it is up to 1Mnps.
I don't have a full move generator but a random legal move generator. And I also don't store moves in global memory unlike you. Your engine is probably memory bound because of that. 300k nps is not that much when you divide it by the number of threads in a SIMD unit. When you decide to give up relying on registers (& shared mem), you should bear in mind that the kernel may not be faster than a single cpu eventually. Hashtables are typical examples that are better generated on cpu. Well chess probably is not suited for gpu computation because monte-carlo misses a lot of tactics, but my effort is concentrated on MCTS and games that are suitable for that approach. The MC part is really needed to harness the power of the gpu and hide global memory latency as well.How does your move generator perform?
----
cheers
