When Deeper Blue vs. Kasparov happened Don believed we would all be processing on PCs at the same strength as the Deeper Blue hardware. As I recall DB was chugging along at 200,000,000 nodes per second, I don't believe a typical one CPU PC can match that yet.
So when will I have a PC that will match or better the DB hardware ??
Sean
The comparison is more complex than that. The DB processors were not as efficient with respect to move ordering as today's software programs are. The 1997 deep blue didn't have hash tables in the hardware. The processors were capable of using one, but the multi-ported memory was never designed. So while it could peak at 1 billion nodes per second, and average 200M nodes per second, a program today running at 200M would be significantly better. And we can actually get there with a parallel architecture. I run at 20M on a dual-cpu quad-core box (8 cores total). There are 16 x 4-way boxes already out and available...
How much money would you be looking at to build a beast that could sustain 200M nodes per sec. ?
I would think quite a bit.
Can you grab some of that Bail Out funding?
P.S. Check next door, they have your country pushing daisies
Our departmental chair is interested in acquiring some sort of "interesting machine" such as an 8-way quad (or better) core box we are looking around to see what is available price-wise...
On a dual quad-core crafty runs around 16-20M nodes per second. to reach 200M would take something with around 80 cores. 16 6-core processors would do the trick today. In a year, 8x8 will be real close.
Sean Evans wrote:As I recall DB was chugging along at 200,000,000 nodes per second
Most of them garbage. Rybka showed that it's more important to look at the right nodes than to look at as much nodes as possible (Even though you have to multiply Rybka's nodes by about 17 to get a more reliable figure.)
Take out all the evaluation from some program and the node count will for sure skyrocket, but the program with evaluation will also for sure beat it badly.
You do realize that Rybka is just a fast searcher that hides the true NPS, right? It is not a "super smart" program at all...
Smart enough to cream everything else at present though!
And if Rybka isn't that smart, what does that imply about the others?
that his search is significantly better, nothing more.
Sean Evans wrote:As I recall DB was chugging along at 200,000,000 nodes per second
Most of them garbage. Rybka showed that it's more important to look at the right nodes than to look at as much nodes as possible (Even though you have to multiply Rybka's nodes by about 17 to get a more reliable figure.)
Take out all the evaluation from some program and the node count will for sure skyrocket, but the program with evaluation will also for sure beat it badly.
You do realize that Rybka is just a fast searcher that hides the true NPS, right? It is not a "super smart" program at all...
I disagree.
It is clearly also a "super smart" program.
Disagree all you want. Strelka is _not_ a "super-smart" program. And we know where it came from...
Rybka may search more nodes per second than other programs but this is certainly not the only explaination for rybka's superiority.
The poster also explained that he does not trust rybka's node count
but even if you use realistic number for rybka it is clear that rybka beat other program with equal nodes per second.
Rybka may be 2 times faster or 3 times faster than other programs in nodes per second but certainly not more than it.
It can explain maybe 100 elo difference but the difference between rybka and second best program is clearly more than 100 elo.
Two things make me think that we are much closer to reaching 200 M nodes on a PC than people think. New processor architectures seem much much more efficient and the cores per socket are increasing as well. Also the Ghz speed is increasing ... so add all those things together and you have a steady incremental gain in hardware performance. I for one am looking forward to seeing what a 16 core (32 thread) Nehalem system can do.
M ANSARI wrote:Two things make me think that we are much closer to reaching 200 M nodes on a PC than people think. New processor architectures seem much much more efficient and the cores per socket are increasing as well. Also the Ghz speed is increasing ... so add all those things together and you have a steady incremental gain in hardware performance. I for one am looking forward to seeing what a 16 core (32 thread) Nehalem system can do.
the hyper-threading is not going to help good programs that have been optimized for memory accesses. It actually doesn't hurt, but it won't run any faster. But factor in the parallel search overhead for the second thread and suddenly you are giving up 30% for each hyper-threaded process, and the net effect is a slower search...
