buying a new computer

[corres]

The huge memory usage of Lc0 is a returning theme on the sites of LC0.
But this is the first established answer for it.
Why, Dear "crem"?
Is it so great secret?
I can not believe the developers of Leela can not make modifiable the usage of RAM with an UCI parameter. Maybe it would decrease the Leela`s Elo in some measure but this effect would be decreased with a timed and structured Hash table as it is in AB engines.
It is not an elegant solving for this issue to let the Leela freezing down when it consumes the full of system RAM.

There already is a parameter, although there is some question about how accurate it is.

--ramlimit-mb=0..100000000
Maximum memory usage for the engine, in megabytes. The estimation is very rough,
and can be off by a lot. For example, multiple visits to a terminal node counted
several times, and the estimation assumes that all positions have 30 possible
moves. When set to 0, no RAM limit is enforced.
[UCI: RamLimitMb DEFAULT: 0 MIN: 0 MAX: 100000000]

The question is now it is working or not. Another question how behave Leela when she reaches the limit..
I also note the 250 bytes / node usage from RAM and 350 bytes / NN Cache are new information for me and I am afraid not only for me.

[Ovyron]

The question is now it is working or not.

Doesn't seem to work on Shredder GUI, using the up arrow of the box to increase it makes it go to -1... negatives values reset to 0. Setting the value to 1 manually doesn't have any effect (still going up to +360MB used), except Leela apparently gives up and aborts analysis after a while (in this case, Depth 9 / 3148 nodes) (This is CPU Leela running at... 0 kn/s...)

[supersharp77]

I think to buy a new computer and I wonder what is the elo of stockfish and lc0 for different prices of hardware.

Is there a table for comparison that give price and elo estimate for both?

Problem Solved....enjoy!!

Budget PC A:https://www.youtube.com/watch?v=7hd9q1xapwI


Budget PC2: :https://www.youtube.com/watch?v=-E2JZKk53bs

[Zenmastur]

Apparently an AMD TR4 (SP3r2) socketed CPU with 16 cores and 32 threads was benchmarked by someone on a public benchmarking site. From the data it's concluded that early Threadripper silicon single core performance is on par with ZEN 2 Ryzen chips and multi-core performance is about 11% faster than Threadripper 2950X and 35% faster than Ryzen 9 3900X. It's speculated that they may also be releasing a new chip set for the chip as an upgrade to the X399 chipset to better support PCIe gen 4.

This would make a good machine for NN because of the huge number of PCIe lanes available.

For those wanting a cheaper system that don't need more memory or as many PCIe lanes Ryzen 9 3950X (if it performs as well as this benchmark suggests) will be the way to go.

But first we'll have to wait for the vendors to stop price gouging on the AMD parts. I've seen the Ryzen 9 3900X priced as high as $800 US or $300 above MSRP!

Regards,

Zen

[chessware]

Hi guys, I just registered and this is my first post here, although I'm reading this useful forum for a long time. A bit excited now to talk to such a great experts, whose knowledge I can hardly enrich, so please let me just ask a, probably stupid, question:

Is ECC RAM bad idea for chess engine like Stockfish on desktop with Ryzen 3000 cpu ?
To be more precise:
1. What would be the difference in nps between ECC 2400MTs CL17 and "non-ECC 3200MTs CL14"
2. What is the expected "spoil" in time on the strength of a chess engine(Stockfish) with 16Gb hash table, in a long run (lets say a week), due to occasional bit flips in non-ECC memory, on a computer which reboots once a month. Or even if it reboots often, it would stack more and more entropy on the HDD in time, which may cause some direct and indirect impact on the strength of the engine.

I've red very thoroughly a great discussion about ram speed/strength dependency here: http://www.talkchess.com/forum3/viewtopic.php?t=63886

I get about 5% increase in NPS by going from 2133 to 2666 or 2800. But these figures will vary depending on how well your system is tuned and what the latency of the ram is. In some cases it's actually better to run a lower clock frequency on the ram if this allows the ram to be run at lower latencies.

but not a word for ECC, hence my supposition about the stupidity of the question. On the other hand I see computer chess championships like TCEC and CCC go with ECC memory - why should they slow down the things unnecessarily ?


Regards.

[Zenmastur]

Hi guys, I just registered and this is my first post here, although I'm reading this useful forum for a long time. A bit excited now to talk to such a great experts, whose knowledge I can hardly enrich, so please let me just ask a, probably stupid, question:

Is ECC RAM bad idea for chess engine like Stockfish on desktop with Ryzen 3000 cpu ?
To be more precise:
1. What would be the difference in nps between ECC 2400MTs CL17 and "non-ECC 3200MTs CL14"
2. What is the expected "spoil" in time on the strength of a chess engine(Stockfish) with 16Gb hash table, in a long run (lets say a week), due to occasional bit flips in non-ECC memory, on a computer which reboots once a month. Or even if it reboots often, it would stack more and more entropy on the HDD in time, which may cause some direct and indirect impact on the strength of the engine.

I've red very thoroughly a great discussion about ram speed/strength dependency here: http://www.talkchess.com/forum3/viewtopic.php?t=63886

I get about 5% increase in NPS by going from 2133 to 2666 or 2800. But these figures will vary depending on how well your system is tuned and what the latency of the ram is. In some cases it's actually better to run a lower clock frequency on the ram if this allows the ram to be run at lower latencies.

but not a word for ECC, hence my supposition about the stupidity of the question. On the other hand I see computer chess championships like TCEC and CCC go with ECC memory - why should they slow down the things unnecessarily ?


Regards.

Answer # 1: About 7% -10% depending on various "other" factors.

Answer # 2: > than a year

In short, I wouldn't worry much about "spoiling" your calculations. Stockfish has a hash collision rate greater than 500 times a second at 38M nps. These are much more likely to "spoil" your calculations than an occasional bit flip in ram.

Regards,

Zenmastur

[Zenmastur]

I see that a couple of benchmarks for Ryzen 9 3900X's have been posted at Ipman chess. All core 4.15 Ghz popcnt 24 threads at 41.4 Mnps and with large pages 47.4 Mnps. A direct scale to the Ryzen 9 3950X would yield 55.2 Mnps and 63.2 Mnps with large pages.

Still not much info on third generation Threadripper but with twice the memory bandwidth and reduced latency due to interleaving the four channel memory a 16 core version should perform as well or slightly better than the Ryzen 9 3950X.

I wish they would hurry up! I'm getting impatient. It would help if they released pricing info and some stats on top of the line Threadrippers. If Ryzen 9 3900X is any indication a 32-core gen 3 threadripper should be north of 100Mnps. At least if they released some info you could plan ahead for which CPU to buy and which peripherals will fit in the budget etc.

Regards,

Zenmastur

[Dann Corbit]

The commercial chips come out tomorrow (at least the announcement does)

[Zenmastur]

The commercial chips come out tomorrow (at least the announcement does)

It's none too soon. I spent most of the last two days trouble-shooting my computer. It has some kind of hardware issue that I can't pin down. I got it working again. I was trying to install new HD's and NVMe drives. It quickly turned into a nightmare of cassacading problems. I would no sooner fix one before another one would rear it ugly head. I did finally manage to get them installed but I'm still not happy because it's still partially crippled. I had to reduce the OC of the CPU and memory to make it stable.

So, I'm definitely ready for a new machine!

Regards,

Zenmastur

[Zenmastur]

Leaked benchmarks on a 32-core AMD non-epyc cpu shows about a 35% increase in compute power. Assuming this is a benchmark for the 3rd gen Threadripper 3990WX it should be hitting around 106M nps. The base clock increased from 3.0ghz to 3.6ghz and about a 15% IPC increase yields a theoretical 38% increase in performance. They may have fixed the memory problem as well as the memory score for this CPU increased substantially over the Threadripper 2990WX. Without more data it's not possible to know what is causing the large increase in memory performance.

With enhanced memory performance the 3990WX would seem to be the chip of choice for chess since you get both high core count and high clock speeds. If the price is about the same as the 2990WX the price to performance ratio will be a little less than for a Ryzen 9 3900X on a new build and it'll be much better than the price performance ratio of a new Epyc build "IF" you don't need the extra PCI lanes and memory channels that EPYC provides.

Regards,

Zenmastur