RE: [computer-go] MCTS, 19x19, hitting a wall? moore's law limits
Lots of simpler cores is possible, but only for running specialized code that doesnt need much memory or memory bandwidth. If I have thousands of cores with small caches the total bandwidth to off-chip memory will be way too high, and performance will be limited by external memory throughput. Look at Tilera http://www.tilera.com/products/TILEPro64.php 64 cores on a chip in the same technology as Intel used to get two cores on a chip. But local memory is small, so it's no good for general computing. Someone might try it for computer go though. David > -Original Message- > From: computer-go-boun...@computer-go.org [mailto:computer-go- > boun...@computer-go.org] On Behalf Of Mark Boon > Sent: Friday, June 12, 2009 1:59 PM > To: computer-go > Subject: Re: [computer-go] MCTS, 19x19, hitting a wall? moore's law limits > > 2009/6/10 David Fotland : > > I think we will get another 64x to 256 x density then it will stop, for > > single chips. We should eventually get desktop machines with thousands > of > > cores, but probably never with millions of cores. There really are > limits > > built into physics L > > > > How about the cores becoming much smaller and simpler? > > Intel's CPUs are approaching a billion transistors on a chip. But you > can probably make a very decent and fast CPU with just a million > transistors. Maybe double that number to give each a bit of cache > memory. If you can see computers with thousands of cores, does that > already assume they'll be simpler? Or could we have a few (hundred) > heavy-duty CPUs like today's for multi-purpose use and a card with a > million simpler CPUs on them next to it for tasks suitable for > parallel processing? A hybrid system if you will. > > Just thinking out loud, I'm obviously a layman when it comes to > semiconductors. > > Mark > ___ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MCTS, 19x19, hitting a wall? moore's law limits
The advertisement claims that an OS can be installed on each core including linux. So I don't know why you say it's no good for general computing. I think you mean each core is far less powerful than a core 2 duo core? I have not looked at the overview paper yet, but it seems like there is no point having 64 on a chip if each is severely crippled up, unless of course there is something they can do much better than a general purpose CPU. Would you estimate that applications that are pure computation and not memory bandwidth would run significantly faster on one of these? I fear we are going to get to the point where we cannot utilize more processors very effectively without taking huge compromises. We will always be able to utilize more to a degree, but we will get to the point where doubling the number of cores only adds 10% to the speed of the computation.We will change our algorithms and adapt to this, but we will always be working around the problems that require some degree of serialization. Tree search can never be fully parallel and remain as efficient as a serial algorithm. - Don On Sat, Jun 13, 2009 at 4:21 AM, David Fotland wrote: > Lots of simpler cores is possible, but only for running specialized code > that doesn’t need much memory or memory bandwidth. If I have thousands of > cores with small caches the total bandwidth to off-chip memory will be way > too high, and performance will be limited by external memory throughput. > > Look at Tilera http://www.tilera.com/products/TILEPro64.php 64 cores on a > chip in the same technology as Intel used to get two cores on a chip. But > local memory is small, so it's no good for general computing. Someone > might > try it for computer go though. > > David > > > -Original Message- > > From: computer-go-boun...@computer-go.org [mailto:computer-go- > > boun...@computer-go.org] On Behalf Of Mark Boon > > Sent: Friday, June 12, 2009 1:59 PM > > To: computer-go > > Subject: Re: [computer-go] MCTS, 19x19, hitting a wall? moore's law > limits > > > > 2009/6/10 David Fotland : > > > I think we will get another 64x to 256 x density then it will stop, for > > > single chips. We should eventually get desktop machines with thousands > > of > > > cores, but probably never with millions of cores. There really are > > limits > > > built into physics L > > > > > > > How about the cores becoming much smaller and simpler? > > > > Intel's CPUs are approaching a billion transistors on a chip. But you > > can probably make a very decent and fast CPU with just a million > > transistors. Maybe double that number to give each a bit of cache > > memory. If you can see computers with thousands of cores, does that > > already assume they'll be simpler? Or could we have a few (hundred) > > heavy-duty CPUs like today's for multi-purpose use and a card with a > > million simpler CPUs on them next to it for tasks suitable for > > parallel processing? A hybrid system if you will. > > > > Just thinking out loud, I'm obviously a layman when it comes to > > semiconductors. > > > > Mark > > ___ > > computer-go mailing list > > computer-go@computer-go.org > > http://www.computer-go.org/mailman/listinfo/computer-go/ > > ___ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ > ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MCTS, 19x19, hitting a wall? moore's law limits
Cool, but what is the cost? The website doesn't even have a word about how to obtain thier product. Insane, IMO. http://www.google.com/products?q=tilera returns ONE (useless) hit. Or is this chip not yet released? David Fotland wrote: Lots of simpler cores is possible, but only for running specialized code that doesn’t need much memory or memory bandwidth. If I have thousands of cores with small caches the total bandwidth to off-chip memory will be way too high, and performance will be limited by external memory throughput. Look at Tilera http://www.tilera.com/products/TILEPro64.php 64 cores on a chip in the same technology as Intel used to get two cores on a chip. But local memory is small, so it's no good for general computing. Someone might try it for computer go though. David -Original Message- From: computer-go-boun...@computer-go.org [mailto:computer-go- boun...@computer-go.org] On Behalf Of Mark Boon Sent: Friday, June 12, 2009 1:59 PM To: computer-go Subject: Re: [computer-go] MCTS, 19x19, hitting a wall? moore's law limits 2009/6/10 David Fotland : I think we will get another 64x to 256 x density then it will stop, for single chips. We should eventually get desktop machines with thousands of cores, but probably never with millions of cores. There really are limits built into physics L How about the cores becoming much smaller and simpler? Intel's CPUs are approaching a billion transistors on a chip. But you can probably make a very decent and fast CPU with just a million transistors. Maybe double that number to give each a bit of cache memory. If you can see computers with thousands of cores, does that already assume they'll be simpler? Or could we have a few (hundred) heavy-duty CPUs like today's for multi-purpose use and a card with a million simpler CPUs on them next to it for tasks suitable for parallel processing? A hybrid system if you will. Just thinking out loud, I'm obviously a layman when it comes to semiconductors. Mark ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] MCTS, 19x19, hitting a wall? moore's law limits
Tilera works well for computations that do not require a lot of memory, for example video transformations, or packet processing. It probably won't work so well for computations that require all processors to have access to off-chip memory, since the memory bandwidth per processor is much lower than general purpose machines. David From: computer-go-boun...@computer-go.org [mailto:computer-go-boun...@computer-go.org] On Behalf Of Don Dailey Sent: Saturday, June 13, 2009 11:15 AM To: computer-go Subject: Re: [computer-go] MCTS, 19x19, hitting a wall? moore's law limits The advertisement claims that an OS can be installed on each core including linux. So I don't know why you say it's no good for general computing. I think you mean each core is far less powerful than a core 2 duo core? I have not looked at the overview paper yet, but it seems like there is no point having 64 on a chip if each is severely crippled up, unless of course there is something they can do much better than a general purpose CPU. Would you estimate that applications that are pure computation and not memory bandwidth would run significantly faster on one of these? I fear we are going to get to the point where we cannot utilize more processors very effectively without taking huge compromises. We will always be able to utilize more to a degree, but we will get to the point where doubling the number of cores only adds 10% to the speed of the computation.We will change our algorithms and adapt to this, but we will always be working around the problems that require some degree of serialization. Tree search can never be fully parallel and remain as efficient as a serial algorithm. - Don On Sat, Jun 13, 2009 at 4:21 AM, David Fotland wrote: Lots of simpler cores is possible, but only for running specialized code that doesn't need much memory or memory bandwidth. If I have thousands of cores with small caches the total bandwidth to off-chip memory will be way too high, and performance will be limited by external memory throughput. Look at Tilera http://www.tilera.com/products/TILEPro64.php 64 cores on a chip in the same technology as Intel used to get two cores on a chip. But local memory is small, so it's no good for general computing. Someone might try it for computer go though. David > -Original Message- > From: computer-go-boun...@computer-go.org [mailto:computer-go- > boun...@computer-go.org] On Behalf Of Mark Boon > Sent: Friday, June 12, 2009 1:59 PM > To: computer-go > Subject: Re: [computer-go] MCTS, 19x19, hitting a wall? moore's law limits > > 2009/6/10 David Fotland : > > I think we will get another 64x to 256 x density then it will stop, for > > single chips. We should eventually get desktop machines with thousands > of > > cores, but probably never with millions of cores. There really are > limits > > built into physics L > > > > How about the cores becoming much smaller and simpler? > > Intel's CPUs are approaching a billion transistors on a chip. But you > can probably make a very decent and fast CPU with just a million > transistors. Maybe double that number to give each a bit of cache > memory. If you can see computers with thousands of cores, does that > already assume they'll be simpler? Or could we have a few (hundred) > heavy-duty CPUs like today's for multi-purpose use and a card with a > million simpler CPUs on them next to it for tasks suitable for > parallel processing? A hybrid system if you will. > > Just thinking out loud, I'm obviously a layman when it comes to > semiconductors. > > Mark > ___ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] Suggesting more Standard opponents on CGOS
While working Pebbles up the ladder, there are times when Pebbles was squarely in the middle of a huge gap. For example, there was a time when the opponents were Aya (~2300) and AverageLib (~700). That is pretty extreme. But even now the next higher rated opponent is +200 and the next lower rated opponent is -300. (Lingo, come back! We need you!) The problem caused by large gaps is that all games have a predictable result, so it is hard to evaluate differences. Can the new CGOS has a more scalable set of standard opponents? E.g., every 200 rating points from 800 through 2400 should be feasible (technically) using available versions of Mogo or Fuego. I understand that expense is an issue, but only the highest of these would be expensive. And perhaps the community can share. Technical details: you can only have one fixed point, which is FatMan at 1800 by definition. The other standard players hit their targets by adapting effort levels. For example, if a player has a current rating > 2000, then reduce effort by a few percent, and otherwise increase effort by a few percent. Best, Brian ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Suggesting more Standard opponents on CGOS
It would please me to have all the gaps filled like this, but it's a matter of getting them to play on the server. In a related issue, I plan to build a weak player into the server. This would be a player who's only function would be to provide an opponent so that no players has to wait out a round. It would much stronger than random, but not strong. It would be nice if standard players over the whole spectrum could be built into the server, but the server is a very low powered virtual box and I don't want to stress it. Another alternative is to get a serious machine, such as a quad or octal capable of supporting several standard players to be run on the side. Many programs are scalable, so it would really be cool if the server itself could have a small set of "ready to go" players available which it could pick and choose as needed to fill gaps when the appropriately rated players are not playing. It would be ideal if such a dedicated machine and the CGOS server itself were sitting in a data center somewhere performing these functions.But that is too expensive to consider. >From time to time we get volunteers to run the anchor player, and indeed we depend on this. But since people do this on a voluntary basis we cannot expect complete reliability and we cannot enforce the existence 24/7 of an anchor player. So that's another issue. - Don On Sat, Jun 13, 2009 at 3:54 PM, Brian Sheppard wrote: > While working Pebbles up the ladder, there are times when Pebbles > was squarely in the middle of a huge gap. For example, there was a > time when the opponents were Aya (~2300) and AverageLib (~700). That > is pretty extreme. But even now the next higher rated opponent is > +200 and the next lower rated opponent is -300. (Lingo, come back! > We need you!) > > The problem caused by large gaps is that all games have a predictable > result, so it is hard to evaluate differences. > > Can the new CGOS has a more scalable set of standard opponents? E.g., > every 200 rating points from 800 through 2400 should be feasible > (technically) using available versions of Mogo or Fuego. I understand > that expense is an issue, but only the highest of these would be expensive. > And perhaps the community can share. > > Technical details: you can only have one fixed point, which is FatMan at > 1800 by definition. The other standard players hit their targets by > adapting effort levels. For example, if a player has a current rating > > 2000, > then reduce effort by a few percent, and otherwise increase effort by a > few percent. > > Best, > Brian > > ___ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ > ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MCTS, 19x19, hitting a wall? moore's law limits
Standard approach to the cpu is getting to the limit. New technologies could emerge. The future is hard to predict. One approach has not yet being published in Go, that is the special purpose cpu. One probably should start getting familiar with terms of digtal design, such as barrel shifter and priority encoder. When an algorthm start to tapper off wth increasing number of playout, it could mean that the hgher order terms begn to dominate. DL ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
