buying a new computer

[Dann Corbit]

...
My next machine will definitely be a modern one. I might get a 3950x, with a couple 2080 cards, but more likely the next threadripper or one of the commercial systems.

Yea, keep in mind that AMD Ryzen based systems have only 2 memory channels, compared to 4 for Threadripper, and 8 for Epyc server systems.

--
Srdja

Even with the threadripper, you can see the highest end model tailing off when the thread count gets really high (not just SMP loss, it tails off drastically). If they don't do something to correct it, I am guessing the next threadripper won't be the superstar for chess that we dream of.

[smatovic]

Someone ever considered a refurbished workstation? I got a HP Z620, I7, 2.8-3.5Ghz, 20 cores, 40 threads, 16Gb ram, 500Gb SSD, 2 x 1 TB harddisk for under 1000 euro with 2 year guarantee.

...unless you have your own power plant and running 24/7, I guess it pays off to
invest in CPUs/GPUs based on smaller fabrication processes that consume less
watts.

***edit***
resp. have more instruction throughput per watt

--
Srdja

[smatovic]

...
My next machine will definitely be a modern one. I might get a 3950x, with a couple 2080 cards, but more likely the next threadripper or one of the commercial systems.

Yea, keep in mind that AMD Ryzen based systems have only 2 memory channels, compared to 4 for Threadripper, and 8 for Epyc server systems.

--
Srdja

Even with the threadripper, you can see the highest end model tailing off when the thread count gets really high (not just SMP loss, it tails off drastically). If they don't do something to correct it, I am guessing the next threadripper won't be the superstar for chess that we dream of.

I am not into the details, but afaik AMD crippled something with the older 32 core systems and memory,
time will tell how the new ones perform.

--
Srdja

[Dann Corbit]

...
My next machine will definitely be a modern one. I might get a 3950x, with a couple 2080 cards, but more likely the next threadripper or one of the commercial systems.

Yea, keep in mind that AMD Ryzen based systems have only 2 memory channels, compared to 4 for Threadripper, and 8 for Epyc server systems.

--
Srdja

Even with the threadripper, you can see the highest end model tailing off when the thread count gets really high (not just SMP loss, it tails off drastically). If they don't do something to correct it, I am guessing the next threadripper won't be the superstar for chess that we dream of.

I am not into the details, but afaik AMD crippled something with the older 32 core systems and memory,
time will tell how the new ones perform.

--
Srdja

I guess it is nothing but bus contention,
58,856,709 32 cores
63,113,198 36 cores
77,576,847 64 cores
edit: should be threads, not cores
No such behavior on Epyc

Of course, they might have done something so that the Threadripper was not simply a cheaper, faster Epyc at high thread count

[smatovic]

...
My next machine will definitely be a modern one. I might get a 3950x, with a couple 2080 cards, but more likely the next threadripper or one of the commercial systems.

Ah, another note, the Ryzens seem to have only 16 PCIe lanes for GPUs,
considering PCIe 4.0 and two GPUs this could be enough for LC0,
but the Threadrippers and Epycs have more lanes...

--
Srdja

[Dann Corbit]

...
My next machine will definitely be a modern one. I might get a 3950x, with a couple 2080 cards, but more likely the next threadripper or one of the commercial systems.

Ah, another note, the Ryzens seem to have only 16 PCIe lanes for GPUs,
considering PCIe 4.0 and two GPUs this could be enough for LC0,
but the Threadrippers and Epycs have more lanes...

--
Srdja

That would be nice. I have boatloads of GPUs, and I use them for lots of things besides chess

[corres]

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?

Hi Uri!

Based on the multicore intetger score of 77115. You can use NPS as a proxy for Elo - for the same engine , they are correlated to each other - just keep mind at these levels , it's not clear if doubling of nps will be worth 70 Elo like like in the past , but it should be close.

Stockfish should bench at 60,000 nps at a minimum with this processor using all 16 cores It could be higher, we shall wait and see. If you hyper thread , it could be even more. Just to clarify, "should" also has a meaning that maybe it will not. I'm not totally familiar with AMD processors, but if this was an Intel processor with Integer score of 77K, it would provide 60,000 nps at a minimum.

From looking at the performance delta of the second gen Threadripper 2920 compared to the Ryzen 9 3900X and the performance of the Threadripper 1950X and 2950X on chess bench marks It's likely that the Ryzen 9 3950X will perform even better than your estimate.

Intel is dead this year. AMD is the way to go this year if you want to upgrade. All the Ryzen 3000 can not be touched. Should be a great year for AMD.

Important to note. You need to buy 3700 speed memory. For best performace with the Ryzen 3000. 3700 ram speed will sync 1 to 1 with the new ryzen 3000 infinity fabric.

From a practical point of view this statement is wrong. There is very little performance delta between DDR4 3200 CL 14 and DDR4 3733 CL17. AB chess programs are not known for their memory bandwidth requirements. If that were the case then the Threadripper 2990wx would totally suck at SF benchmarks since it has slightly less than half the per core memory bandwidth of the 2950X. And yet it's still reaches 77.5M nps vice 41M nps for the 2950X. Therefore for SF and other AB chess engines latency matters more as long as the memory is not clocked “TOO” slow. The latency difference between DDR4 3200 CL14 and DDR4 3733 CL17 is a mere 3% due to the infinity fabric. The price for 32GB of each is $239.99 for the DDR4 3200 and $429.99 for the DDR4 3733. So, you can get by with DDR4 3200 with little impact if you are on a budget. Even if you are building a system from scratch the price and performance differences doesn't justify the extra cost of the faster memory unless you have an unlimited budget or you have other requirements to meet.

This:

CPU: AMD Ryzen 9 3900X 12 Core --- nominal $500
Should give with Stockfish some 35 million NPS in the openings

GPU: NVidia RTX 2070 Super --- nominal $500
Should give with Leela T40 (depending on the net) 25-35k NPS in the openings (after say one minute search with a large cache)

32 GB DDR4 RAM, Leela with its MCTS is especially RAM hungry.

All in all --- probably below $2000 for the whole system.

Elo wise, Lc0 with late T40 nets should be somewhat stronger than Stockfish on full CPU at longer time controls. At fast time controls, a bit weaker.

EDIT: This system is more than half of what we have today as TCEC hardware (2 TCEC machines combined). The results at LTC will look similar to TCEC results at LTC.

EDIT2: Don't forget the cooling: large case, many fans, a good CPU cooler. I am overclocking both my CPU and GPU even with this heatwave which hit here 37C, by almost 10% above nominal speeds, having these little things.

16 core 3950X will come only in September, and is not a very good deal. It will cost $750 or 50% more than the 12 core 3900X ($500). Has a lower base clock and at best a similar stable clock (getting rid of Boost for consistency) to 3900X. More important, the L2 + L3 cache sizes are almost the same in 12 core 3900X and in 16 core 3950X, which looks not very good for Stockfish on 16 core 3950X. All in all, I expect about no more than 25-30% higher Stockfish NPS for 16 core vs 12 core (additionally keep in mind not perfect scaling too), at 50% higher price.

If you want really cheap and still strong (about 70% in speed compared to my earlier proposal), under $1400 for all the system you can get the 8-core Ryzen 7 3700X for $330 and NVidia RTX 2060 (not Super, Super is more expensive) for $350.

Not sure what you thinking when you posted this but, if you are building a system from scratch, like the system you proposed then using a Ryzen 9 3950X instead of a Ryzen 9 3900X raises the cost of the system from $2000 to $2250 or 12.5%. In the mean time performance has increased by substantially more than 12.5% therefore this is a MUCH BETTER option than the Ryzen 9 3900X. Furthermore, having a faster system will allow you to use it longer before upgrading. e.g. a four year upgrade instead of 3 years (for example). Therefore the cost per month of use (excluding utilities) would be $55.56 per month for the Ryzen 9 3900X vice $46.88 per month for the much faster Ryzen 9 3950X.

Thanks for the all the replies.

Note that I do not understand much about hardware.
I read basically the following information when I do not know exactly what is the meaning of everything.

1)CPU: AMD Ryzen 9 3900X 12 Core or AMD 16 Core Ryzen 9(not sure if 16 cores are faster and how much faster)
2)GPU: NVidia RTX 2070 Super
3)32 GB DDR4 RAM
4)a good CPU cooler
5)large case
6)many fans,
7)PCIe 4.0 M.2 drive.
8)new AMD motherboards. 5.0 GB/s

Don't forget a BIG power supply. Those GPU cards drink electricity with a slurp and a burp. I would say a KW is minimum,
The motherboards that support PCie 4.0 cost a lot. Figure 700 dollars.
Since you play correspondence chess you might consider 64 or even 128 GB of RAM. I read that the 3600 speed ram performs about as well and is cheaper

One of the best points about Ryzen is power usage and heat generation. Both the Ryzen 9 3900X and the Ryzen 9 3950X are rated at 105 watts at there boost clock speed. As a note Intel rates their CPU's power usage at their base clock speed. This makes a HUGE difference. The AMD CPU's come with an adequate cooler. Unless you live in a hot climate or are overclocking there seems little point in buying a more expensive cooler for the AMD cpu's. Secondly, a 250-300 watt GPU doesn't need a 1000W power supply. A 650-700 watt is plenty unless you are doing something special like dual, triple, quad GPU's or massive over-clocks.

You don't need a large case or lots of fans in most cases. I ran my 5820K all core over-clocked to 4.8Ghz in a small case with just a top and rear fan 24/7 365 since the CPU was introduced and I live in the desert where temperatures regularly exceed 45 degrees C (109 degrees F). I recently slowed it down a bit due to stability issues as a result of not cleaning the case out for several years ( I normally do this every spring). But it's time for an upgrade as 20M-24M nps is now too slow. I'm looking at both the 3900X, 3950X and 7nm Threadripper if it's released soon enough.

For me the much touted X570 motherboard seems like over kill and brings little to the table for AB chess engines. B450, X370 and X470 motherboards yield identical CPU performance as the X570 at half the cost (in some case less than $75 US). So unless you need PCI gen 4 for some reason, spend your money on a faster CPU like the Ryzen 9 3950X.

One Last note: Do check your motherboard compatibility before buying! Not all MB's will handle the new chips, but most will, so check the manufactures web site before buying.

Regards,

Forrest

As I see you are a machine builder expert.
I am not a great expert but I agree you in every points.

[Laskos]

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

One thing: 32 GB of RAM. You are doing many experiments and maybe don't need long runs like when one is analyzing, but keep in mind that on your GPU and CPU, something like Leela MCTS on infinite mode will fill up the free part of 16 GB RAM (say 12-13 GB) in less than 15-20 minutes. That's an issue, unlike that with the traditional engines. Also, if your OS cache is on HDD, after such runs, other applications will almost freeze.

[Laskos]

Not sure about what you say. With this reasoning, Uri would better buy 3000 series 32 core Threadripper with 2080ti Super.

Punctually: 12 core 3900X and 16 core 3950X.

Base clock is significantly (300MHz) higher with 12 core 3900X. I am skeptical the clocks can be equalized when the Boost is stabilized. I guess a gap of 200MHz will remain, or 5% in speed.
L2 + L3 cache size per core is significantly higher on 12 core 3900X compared to 16 core 3950X. IPCs are the same.

All in all, single core performance might be about 10% better with 12 core CPU, and that affects Leela speed, for example (and many other applications). And multi-core performance of 16 core CPU with Stockfish (NPS) is probably no more than 25% higher than that of 12 core CPU.
$750 for 3950X is pricey to my taste, actually a bit overpriced. I am replying to Uri from "a poor guy perspective" thinking of some limited budget. And spending $250+ more for a dubious all-in-all advantage of having 20% faster Stockfish, but 5% slower Leela and 10% slower single core applications is an issue for me.

Well,

First, the Threadripper 2990WX has the disadvantage of haveing only the same memory band width as the 2950X because the two extra chiplets don't have memory controllers on them. Therefore all memory access by cores on those chips must be satisfied by the infinity fabric between chips and eventually uses up bandwidth that would be available to the chip satisfying the request. So not only does each core have half the bandwidth of the 2950X half of the memory accesses will also incurs the infinity fabric latency. Even though the 2990WX has terrible memory bandwidth per core comparatively speaking, it still does relatively well for AB chess engines because they don't require a great deal of bandwidth to start with. It would be much faster for AB chess engines if all chiplets had a memory controller on them. I haven't tried to analyze the new Threadrippers yet because I have no data on them. So, determining if they would be a good buy for this application will have to wait to see what thy do with them. The could "fix" the problem with the 2990WX in the 3990WX in which case I'm pretty sure it will be the better buy depending on cost of course. We will see when the time comes.

Therefore a direct comparison between the older generation crippled 32 core cpu and the new generation non-cripped cpu is flawed.

Yes, I speculated on the issue that 3000 series 32 core Threadripper is rumored to fix the issue, and then, not having a clear budget, might be a "best buy". But such a system will be in $5000 range, if coupled with other powerful hardware components. There are "best buys" and "hard to get even better buys", and now that Uri wrote about $2500 budget, this one is ruled out.

Second, the chiplets are pre-binned by AMD before placing them into a cpu. AMD has as much as stated that the best binning chiplets will go on the higher priced SKU's such as the Ryzen 9 3950X. This means that the coolest running and fastest chiplets will be found on the 3950X. This is why thermal throttling will be less than one might expect. By adjusting the power target you can virtually guarantee that the cores will be boosted to their thermal limit. Since AMD sets the thermal limit based on boost clock and NOT base clock, arbitrarily discounting the boost clock would be a mistake. I'm not saying that you will be running all cores at the boost clock with out over-clocking. But I am saying it will not be averaging just the base clock unless it's in an adverse thermal environment. From what I've seen, with reasonably good cooling, I expect average clocks on all cores while running an AB engine to be higher than the base clock of the 3900X and it could be significantly higher. This is based on running power virus software on all cores (this draws significantly more power than ANY chess engine) and still maintaining average clocks above base clock speed.

I'm going to wait and see before I buy, but that's what I'm expecting.

I Think you are wrong on how much you discount the 16 core based on heat throttling. But time will tell. Theoretically, if you keep the CPU cool and supply enough power (e.g. increase the power target 50%) all cores can maintain boost clock indefinitely. Since the boost clock is greater on the 3950X by virtue of chiplet binning I think it will be a far superior CPU for running chess engine.

Regards,

Forrest

There is much speculation in all this. Are you sure that "the coolest running and fastest chiplets will be found on the 3950X"? Are you sure that 3950X with same reasonable cooling can match 3900X on average clocks on all running cores? Nobody knows. Also, L2 + L3 cache per core in 3900X is about 30% higher than that per core in 3950X (that is not a speculation), and for a cache intensive software like Stockfish, it does matter. All in all, I would bet that per core performance with SF on all cores will be higher in 3900X versus 3950X, having similar cooling and temperatures. And the reasons to spend $250 more on 3950X compared to 3900X are debatable ($150 seems to be a fairer price addition). A good chunk of applications, Leela included, depends more on 1-2 core performance than that on all cores. Yes, one can go back and boost those very fast 1-2 cores on 3950X, keeping the other on base clock, but it's not what we are usually doing when mixing up full CPU and 1-2 core performances. Let's see, I am speculating a bit too.

[Milos]

Someone ever considered a refurbished workstation? I got a HP Z620, I7, 2.8-3.5Ghz, 20 cores, 40 threads, 16Gb ram, 500Gb SSD, 2 x 1 TB harddisk for under 1000 euro with 2 year guarantee.

...unless you have your own power plant and running 24/7, I guess it pays off to
invest in CPUs/GPUs based on smaller fabrication processes that consume less
watts.

***edit***
resp. have more instruction throughput per watt

--
Srdja

"Own power plant" for 100 extra W, seriously??? You can run it for a full month non-stop to accumulate the same amount of energy as one charge of Tesla Model 3 car.