buying a new computer

[Zenmastur]

I started thinking about what would be an Idea machine. The idea I came up with is for AMD to design a new CPU that uses the defective chiplets ( i.e. chiplets with less than 6 usable cores on them but do have a working memory controller. Putting four of these on the chip would give the cpu 16-cores and 8 memory channels i.e. huge memory bandwidth. Then put four TPU's (newly designed) with DMA features on them. Put all chiplets on a Threadripper or Epyc compatible die. This would give you 16 cores plus four TPU's for use with NN. The TPU's would have direct access to large quantaties of high bandwidth main system memory. This would solve the limited memory problems with current GPU's. If you wanted to augment your TPU's they could design a card with 4 to 16 extra TPU's on them that use the PCIe Gen 4 X4, X8, and X16 slots. These would appeal to anyone using NN/AI such as researchers and data centers.

Regards,

Zenmastur

[Zenmastur]

If you can build an AB engine without PSQT, do it.

I don't know how to do that with good effect, but I do know how to keep the effects of the PSQT out of the opening. I think this is sufficient to have the desired effect.

I think to make a correct self play test to determine the Elo enhancement of the developing it needs different kind of start positions than it needs for making an opening book.
Using very short time control for keep consistency of an opening book will yield a weaker book mainly in the case of long time controlled games.

I don't agree with the last statement. There is no reason to build the book from scratch. You can start with the BEST book you can find. The book algorithm will then adjust the frequency of play for each of the lines in the book to better suit it's play and to increase it's playing strength. It will also extend the book when it finds better lines of play OR it encounters positions not already in the book.

A few years ago I did some analysis on the cost of computing the value of a position by searching the tree vice how much it would cost to store and then retrieve the pre-computed value. This analysis include the life cycle cost of the hardware and the cost of electricity used. I used 5-years (157,680,000 seconds) as the life of the hardware. It turns out that if you spend more than a fraction of a second searching the tree and evaluating the position, it actually costs less to retrieve it from a data base and it saves time in the process.

What this means for an opening book is that most/all positions that are searched for more than a faction of second should be stored in an opening data base. At least if they occur near the beginning of the game and meet other selection criteria. A huge fraction of them will be seen again and therefore it will save both time and money to store them. Since the book can only be extended, in most cases, by playing games and, on average, a game will add 1 divided by the branching factor position to the book data set per game it shouldn't be too hard to calculate the maximum book growth rate in MB per year.

The extended parts (i.e. leaf nodes) will be no weaker than the time control they are added from. If this depth of search isn't sufficient for the current time controls the book routine will exit to the search so no strength will be lost regardless. The lines the algorithm plays WILL get stronger with time. Even if all play is at short time controls.

So, I don't think your statements are justified.

Regards,

Zenmastur

[Dann Corbit]

My wife offered to take ANY interim system I build, including a Ryzen 3900X, as her new system. T

Every man should find such a wife as that.

[Zenmastur]

Well, you don't really disagree with the statement, either ("no sign so far").

I'd hope they'd get stronger in the endgame & middlegame before they get stronger in the opening.

Ok, I'll say it. I disagree! AND I hope NN engines get much stronger in the opening. In fact, I wouldn't mind it if they produced very large NN's that play nothing but the opening. My analysis is that AB engine if left to themselves play the opening mediocre at best. The primary reason for this is that the programmers have no clue and/or no motivation. So, any engines that play the opening better would be welcomed in my book!

Pun intended!

Regards,

Zenmastur

[Zenmastur]

My wife offered to take ANY interim system I build, including a Ryzen 3900X, as her new system. T

Every man should find such a wife as that.

I'm not complaining!

[zullil]

I seemed to have found the error in the first game but have't as yet looked at the second.

My claim: Stockfish played the error much earlier than it thinks. Several alternatives suggested by Stockfish would have also lost it the game (say, MultiPV shows 0.00 for 3 alternative moves, but one of them loses, how do you decide what to play? And if Stockfish is one move away from randomly losing the game, should it really show 0.00 on the previous move if we don't know it'll play the losing one in the next? Does 33% chance of losing the game in the next move look like 0.00?) This happens on a regular basis.

More claims that you are unwilling (and, I suspect, unable) to support. And, just to clarify, "would have also lost it the game" for you implies "at a time control of 12+2 on a certain chess site." You seem prone to over-generalization, perhaps without realizing it.

[Zenmastur]

I am going to try and build the pc myself -- never built one before but watching lots of youtube videos has given me confidence. Fingers crossed I don't toast one of the components since i seem to be highly charged.

The 3900x was out of stock yesterday but managed to buy one today on newegg. Based on pc part picker i need to spend atleast 1500$ before tax if I pair it with an RTX 2070 super. I am going for a cheap X570 motherboard but maybe i should go for a solid X470 without PCIe 4.0 since i probably will not have any use for the extra SSD speed. 16GB 3200 ram.

Daniel

Daniel, would you please post some performance data on your new machine when it's up and running. i.e stockfish NPS and Lc0 speeds on the 2070 super. I'm sure everyone here would like to see how both the CPU and GPU perform!

And if you need any help/info to complete the build let us know.

Best regards and thanks in advance,

Zenmastur

[Milos]

I started thinking about what would be an Idea machine. The idea I came up with is for AMD to design a new CPU that uses the defective chiplets ( i.e. chiplets with less than 6 usable cores on them but do have a working memory controller. Putting four of these on the chip would give the cpu 16-cores and 8 memory channels i.e. huge memory bandwidth. Then put four TPU's (newly designed) with DMA features on them. Put all chiplets on a Threadripper or Epyc compatible die. This would give you 16 cores plus four TPU's for use with NN. The TPU's would have direct access to large quantaties of high bandwidth main system memory. This would solve the limited memory problems with current GPU's. If you wanted to augment your TPU's they could design a card with 4 to 16 extra TPU's on them that use the PCIe Gen 4 X4, X8, and X16 slots. These would appeal to anyone using NN/AI such as researchers and data centers.

Regards,

Zenmastur

You obviously don't know much about chip design. Putting 8 chiplets where each is close or over 100W TDP on the same die is feat on par with putting man on Mars.

[Zenmastur]

I started thinking about what would be an Idea machine. The idea I came up with is for AMD to design a new CPU that uses the defective chiplets ( i.e. chiplets with less than 6 usable cores on them but do have a working memory controller. Putting four of these on the chip would give the cpu 16-cores and 8 memory channels i.e. huge memory bandwidth. Then put four TPU's (newly designed) with DMA features on them. Put all chiplets on a Threadripper or Epyc compatible die. This would give you 16 cores plus four TPU's for use with NN. The TPU's would have direct access to large quantaties of high bandwidth main system memory. This would solve the limited memory problems with current GPU's. If you wanted to augment your TPU's they could design a card with 4 to 16 extra TPU's on them that use the PCIe Gen 4 X4, X8, and X16 slots. These would appeal to anyone using NN/AI such as researchers and data centers.

Regards,

Zenmastur

You obviously don't know much about chip design. Putting 8 chiplets where each is close or over 100W TDP on the same die is feat on par with putting man on Mars.

You're funny!

Which chiplets do you think will be drawing 100W? It seems to me that the 8-core chiplets that AMD is using draw about 50W each at 100% thermal load with all 8 cores running. If half these cores are non-functional (i.e 4 cores per chiplet) why do you think they will be drawing "close or over 100W TDP"?

Even if they were drawing that much power It wouldn't be as hard as putting a man on Mars to simply open my browser and order a cryo-cooler for the CPU. You might have to have a custom cold plate built but I'm sure I could find someone that could do that for less than sending a man to Mars!

But thank you for your VERY well thought out response!

Regards,

Zenmastur

[jp]

Well, you don't really disagree with the statement, either ("no sign so far").

I'd hope they'd get stronger in the endgame & middlegame before they get stronger in the opening.

Ok, I'll say it. I disagree! AND I hope NN engines get much stronger in the opening. In fact, I wouldn't mind it if they produced very large NN's that play nothing but the opening. My analysis is that AB engine if left to themselves play the opening mediocre at best. The primary reason for this is that the programmers have no clue and/or no motivation. So, any engines that play the opening better would be welcomed in my book!

You disagree about the future. Do you disagree about the past and present?

The NN engines are not currently playing the opening better than the humans.

The problem for NN engines is that unless they play the endings better (than they do now), they may not be able to evaluate openings that depend on the resultant endings accurately. (If the openings depend on a lot of middlegame tactics, they won't be better than traditional engines either.)