How do you know if your cpu is equal to your gpu?

[jp]

additional CPU cores will do little for most users but a fast GPU will make a big difference in games and VR for desktops. 35% of total U.S. population plays games on a PC https://www.pcgamesn.com/pc-gamers-vs-c ... rs-numbers. Some may play browser games only, so let's say 25%. But an average household in the U.S. is 2.6 people, so the percentage of PCs being used for games is high. There are many fewer users who have any use for more than 4 cores.

There's a huge difference between "plays games" and "is a gamer". Maybe we should start a poll thread to ask who here considers themselves a gamer. (What would the poll options be?) I don't think I know any true gamers.

Those who have no use for more than 4 cores and who don't like unused cores on their computers will be forced to go for a laptop, by the looks of what we've seen, because the budget desktop we looked at has a 6-core CPU.

[jp]

I'd rather teach them about how to make better use of what they already have, and about the diminishing results of chess (double your hardware speed and get only 1 depth more in the same time, that barely makes a difference), and the draws one can get with 10 year old hardware against Leela on fastest GPUs.

Computer chess users want faster hardware because of bad education by their parents on a time computers weren't even mainstream, so teaching their kids (or grandkids) now "you need better hardware" instead of "you can do the same with slower hardware" is a disservice to them.

This is why the scaling curve (performance versus computation time) is the most important thing. The "bad education" you speak of (probably not the fault of their parents) is perhaps why computer chess users are distracted by completely meaningless ratios, instead of ratios that at least are vaguely related to scaling issues.

[mmt]

There's a huge difference between "plays games" and "is a gamer". Maybe we should start a poll thread to ask who here considers themselves a gamer. (What would the poll options be?) I don't think I know any true gamers.

This study also says that the average for an "active gamer" (anybody who played in June 2018) is 12 hours per week.

[jp]

Maybe that's what the poll options could be:
Play computer games on average for >20hrs/wk, 15-20hrs/wk, 10-15hrs/wk, ..., 0-5hrs/wk, rarely, never.

e.g. random mainstream budget desktop: 9th Gen Intel® Core™ i5 9400 (6-Core, 9MB Cache, up to 4.1GHz with Intel® Turbo Boost Technology); Intel UHD Graphics 630 with shared graphics memory.

For example https://www.dell.com/en-us/shop/deals/x ... vmstcs002s.

Yes, this is fine, so in this example we have: 9th Gen Intel® Core™ i5 9400 (6-Core, 9MB Cache, up to 4.1GHz with Intel® Turbo Boost Technology); NVIDIA® GeForce® GT 1030 2GB GDDR5.

GT 1030 is supposedly +146% effective speed vs. UHD 630.
RTX 2080 is supposedly +1,859% effective speed vs. UHD 630.

And the benchmark link in the other thread says Leela on RTX 2080 gets 31,723 nps (on 20x256 NN).
So the UHD 630 in the budget desktop has an estimated Leela speed of 1620 nps.
The GT 1030 on the better (but non-gaming) desktop has an estimated Leela speed of 3980 nps.

Still need to find out what SF speed (nps) the i5 9400 gets.

[jp]

Still need to find out what SF speed (nps) the i5 9400 gets.

Found this:



I don't know what difference the "F" makes, but it shows SF9 on Core i5-9400F gets 13,601,719 nps.


So our estimated "fair" ratio of (nps Stockfish)/(nps Leela) according to "balanced" mass-market PCs is:

~8400 (based on budget non-gaming PC);
~3400 (based on better non-gaming PC).


Example of expensive gaming desktops: "up to" AMD Ryzen™ 9 3950X (16-Core, 64MB L3 Cache, Max Boost Clock of 4.7GHz); "up to" NVIDIA® GeForce RTX™ 2080 Ti 11GB GDDR6 (OC Ready).

Okay, the plot above only goes to Ryzen 7, so might need to downgrade a bit.

[jp]

Example of expensive gaming desktop: "up to" AMD Ryzen™ 9 3950X (16-Core, 64MB L3 Cache, Max Boost Clock of 4.7GHz); "up to" NVIDIA® GeForce RTX™ 2080 Ti 11GB GDDR6 (OC Ready).

Okay, the plot above only goes to Ryzen 7

But the benchmark link shows RTX 2080Ti gives Leela speed of 50456 nps, which is a Leela speedup of only a factor of 1.59. Even going to the Ryzen 7 from i5 9400F is a Stockfish speedup of 2.02.

So the estimated "fair" ratio is actually bigger (better for SF) based on high-end gaming machines than it is based on non-gaming machines.

So we're stuck with a "fair" ratio of (nps Stockfish)/(nps Leela) according to "balanced" consumer PCs (based on crem's suggestion) of >~3400.

(But computer-chess users have been conditioned to a ratio of only ~875, which gives Leela ~4-10 times as much computing power.)

[jp]

Ah, but then GPUs might use more energy over time, so the energy cost would eventually catch up to anything

Good point but the energy cost will only make a difference over months of 24h/day use. A 2080 ti uses around 280W (https://www.tomshardware.com/reviews/nv ... 05-10.html).

Energy consumption is another suggestion many people have made.

RTX 2080Ti gives Leela speed of 50456 nps at a cost of 280W.
Ryzen 7 3800X gives SF9 speed of 27285311 nps at a cost of 91W (https://www.tomshardware.com/reviews/am ... 226-3.html).
We'd like to use Ryzen 9 3950X, but we don't have SF speeds for that.

So our estimated "fair" ratio of (nps Stockfish)/(nps Leela) according to equal power consumption is:

~1660.

(The ratio users have been conditioned to, i.e. ~875, gives Leela ~2 times as much computing power.)


And I'd say the estimated "fair" ratio according to cost (money), which mmt was talking about, overlaps with the estimated "fair" ratio according to "balanced" mass-market PCs above, because I'd say the cost considered should be additional cost for people who'd own a computer anyway, and the "balanced" PCs are zero additional CPU or GPU costs.

[jp]

So our estimated "fair" ratio of (nps Stockfish)/(nps Leela) according to equal power consumption is:
~1660.

On second thoughts, it seems I should have added to the GPU energy cost the energy cost of some CPU because it still needs a CPU. It looks like most CPUs are at least 65W (even those you would not want to use with the RTX 2080Ti), so

the revised estimated "fair" ratio of (nps Stockfish)/(nps Leela) according to equal power consumption is:

~2050.

[Ovyron]

It seems the fairest ratio to use is Leela and Stockfish at $0/extra cost (with the hardware you already have.)

[jp]

... fairest ratio to use ... the hardware you already have.

That's also reasonable (or at least not meaningless), including people using 10- or 15-year-old computers.

It seems to me that'll probably either overlap with the ratio from "balanced" PCs (above), e.g.

(nps Stockfish)/(nps Leela) ~8400 (based on budget non-gaming PC),

or else it'll be SF vs. CPU-Leela on the same CPU.

For that, posters on this forum say Leela gets 30-40 nps (on a 20x256 NN). Let's assume (possibly generously to Leela) that this is on a low-end CPU like i5-9400F, which has SF9 speed 13601719 nps.

So then the estimated "fair" ratio of (nps Stockfish)/(nps Leela) according to "same CPU" is:

~450,000.

(The ratio users have been conditioned to, i.e. ~875, gives Leela >500 times as much computing power. You, Ovyron, will be left with the job of placating outraged Leela fans.)