Daniel Shawul wrote: ↑Mon Nov 19, 2018 6:01 pm
I wanted to buy a 2070 but they told me i need to buy a gaming desktop for it too. Silly me i thought i could just plug it in via thunderbolt or something.
I don't need the gaming desktop (minim 700$), so I am thinking to buy a cheaper desktop computer if it will work with the 2070. Maybe a used one from ebay will do or wait for black friday sell. But it just seems a lot of hassle to maintain and using the cloud may be a better option.
What do people that own GPUs think about this ?
you need two aspects, first, power supply, second the board you plug the GPU into needs to be with PCI express slots, usually the main board.
Third, cooling fans. I think some people have made banked GPU's on a separate board, you can look that up.
be careful the board with the PCI slots has them far enough apart (two hot adjacent GPUs not clever) if you want to run with multi GPU. Also some board layouts end with your big GPUs physically trying to occupy the same space as other PC components.
it might be worthwhile to use a system that can bank multiple gpus, and buy any gpus as they come available. Treat as a server.
Werewolf wrote: ↑Mon Nov 19, 2018 6:01 pm
I think power supplies are at their most efficient at about 50% load, as they reach maximum load the efficiency drops a bit - might be finding out what wattage you're consuming with a power meter.
But if your CPU is a quad, only one hard drive, a 500W supply should be OK I would have thought.
This varies from model to model, but on something with say an 80 PLUS certification it will be at most a few %.
Laskos wrote: ↑Mon Nov 19, 2018 3:00 pm
My power supply is not that strong (500W), hope it stays well.
I think power supplies are at their most efficient at about 50% load, as they reach maximum load the efficiency drops a bit - might be finding out what wattage you're consuming with a power meter.
But if your CPU is a quad, only one hard drive, a 500W supply should be OK I would have thought.
While efficiency is the highest at 50% load it is almost the same at 100% load and somewhat lower (and lowest overall) at 0% load.
Efficiency absolutely doesn't matter regarding if PSU can deliver its maximum power.
Daniel Shawul wrote: ↑Mon Nov 19, 2018 6:01 pm
I wanted to buy a 2070 but they told me i need to buy a gaming desktop for it too. Silly me i thought i could just plug it in via thunderbolt or something.
I don't need the gaming desktop (minim 700$), so I am thinking to buy a cheaper desktop computer if it will work with the 2070. Maybe a used one from ebay will do or wait for black friday sell. But it just seems a lot of hassle to maintain and using the cloud may be a better option.
What do people that own GPUs think about this ?
you need two aspects, first, power supply, second the board you plug the GPU into needs to be with PCI express slots, usually the main board.
Third, cooling fans. I think some people have made banked GPU's on a separate board, you can look that up.
be careful the board with the PCI slots has them far enough apart (two hot adjacent GPUs not clever) if you want to run with multi GPU. Also some board layouts end with your big GPUs physically trying to occupy the same space as other PC components.
it might be worthwhile to use a system that can bank multiple gpus, and buy any gpus as they come available. Treat as a server.
Thanks chris! I will take those into consideration.
Although it would be good to have a multi-gpu server, I am afraid it might drive the cost way up than i can afford.
I am happy if i can get one powerful gpu working for training non-stop.
I need this setup to run a server-client training setup just like leela's so that other could chip in with their gpus.
I need to have atleast one gpu for training on the server side so that i can upload nets faster than games are produced.
chrisw wrote: ↑Mon Nov 19, 2018 5:34 pm
Useful to know. Thanks.
Do you happen to know the ratio of MCTS nodes to positions sent to GPU?
It should be 1-1 in the algorithmic sense but the NNCache (sortof like a transposition table) muddies it.
Ah, I must have misread something. I thought LC0 was speculatively throwing (so to speak) positions at the GPU, storing the positions in a CPU-based cache, and then hoping that the cached positions would then later be used as and when the MCTS needed. And the MCTS wouldn't always need.
But suddenly, I thought, oh, is the NNcache in GPU memory? Or maybe that's just a random side issue.
Anyway, ratio positions sent to GPU / MCTS nodes is a (I thought) measure of the efficiency of the LC0 algorithm. 1:1 looks good. Unless I'm being totally random here.
whereagles wrote: ↑Mon Nov 19, 2018 4:41 pm
not a hardware specialist, but I think you may need to upgrade your power supply...
Don't scare the heck out of me, just running a full-blown match with SF_dev on all 4 i7 cores versus Lc0 v19 rc5 ID11261 on this thing, hope I don't fry something here.
Your i7 is 90W IIRC (and if not OCed) so you should be more then fine with 500W power supply.
If you plan OCed GPU or CPU than I would think of power supply upgrade.
Btw, if you try to draw more power from power supply than what it can provide, your supply voltage will drop on DC3.3V line so in the worst case your machine is gonna shut down (because MB will power down), you are not gonna fry anything .
I OC a bit my i7-4790 from 3.6 to 3.8 GHz, but as I run 4 threads (not full 8) on 4 cores, the power needed is always less than 90W. GPU is at stock speed, and I will not OC it, as it at best will be 5-10% faster with much higher consumption. The speed is so impressive compared to GTX 1060 (5-6 fold improvement), that I really don't care about those 5%.
Thanks, so if anything happens, I will just need another PSU, which is not a big deal.
Night long power testing (full CPU versus full GPU+some CPU) proved stable. Temperatures are fine too.
Daniel Shawul wrote: ↑Mon Nov 19, 2018 6:01 pm
I wanted to buy a 2070 but they told me i need to buy a gaming desktop for it too. Silly me i thought i could just plug it in to my laptop via thunderbolt or something.
I don't need the gaming desktop (minim 700$), so I am thinking to buy a cheaper desktop computer if it will work with the 2070. Maybe a used one from ebay will do or wait for black friday sell. But it just seems a lot of hassle to maintain and using the cloud may be a better option.
What do people that own GPUs think about this ?
I went with minimalist $500 new desktop 3 years ago or so, with powerful CPU for its time, i7-4790. The only good thing was the CPU, everything else was crap, including the motherboard, but it had all the slots I needed. Then upgraded, cooler for CPU, PSU, RAM, SSD, GTX 1060 and now RTX 2070 (maybe I missed some upgrades). I will sell now my GTX 1060 for $200 or so. It's pretty new and has a valid warranty. That crappy motherboard (forgot even the brand) seems to raise no problems. The HDD is still the original 500GB crap, but I have a 2TB external one. All in all, I am satisfied, but maybe in the next 6 months I will get a minimalist 8-core Ryzen system (would like one with a fast CPU), then I won't even need new products, just move from one case to another.
Andre van Ark wrote: ↑Mon Nov 19, 2018 6:53 pm
Nice upgrade, Kai excellent value for money, enjoy it!
Thanks, that was really something. Din't quite expected 5-6 fold speed increase. Also, not that far from what people get with significantly more expensive 2080. And now, Lc0 is the strongest engine on my PC, at least from a 3-mover opening suite (without books). Games at 1m + 1s:
Score of SF_dev vs lc0_v19_11261: 2 - 5 - 13 [0.425] 20
Elo difference: -52.51 +/- 91.20
Finished match
SF_dev is on all 4 cores.
With GTX 1060, in the same conditions, SF_dev was beating the hell out of Lc0 (something like +10 -1 =9).
.