Nvidia RTX 30xx

[iamaibeta]

I think I found my solution. It's a bit frankenstein but should work. I'll swap the motherboard on my Intel machine to a MB with 3xPCIex16 slots and a thunderbolt 3 port. I'll then put the 3090 on top and a 3080 at the bottom and attach an eGPU enclosure with 3080 to the thunderbolt 3 port. In the future, if the waterblock 3080/3090s become available, I could just drop two of those in as well.

Now the problem is even those motherboards seem to be sold out.

[RobertJasiek]

Isn't eGPU/Thunderbolt too slow for mining?

Do you search for preassembled waterblocks or add-ons, which you would have to assemble?

Which mainboard do you consider?

[iamaibeta]

Thunderbolt does not slow down mining at all as far as I could tell, at least with the ehtereum mining algorithms Nicehash has been running since like November last year. Mining in an eGPU is still slightly less efficient because the eGPU unit itself draws more power, but it's really not much, like 50w overhead, maybe. At any rate, mining is just a bonus, a subsidy for the setup, not the primary purpose of the rig.

I am looking at something like the Gigabyte Z490. 3 PCIe x16 slots spaced for triple slot height cards and comes with two thunderbolt 3 ports. It's not ideal because I can only configure the GPUs to run in 8x/8x or 8x/4x/4x (if I put 3 on the MB) but for KataGo and mining it should be a non-issue. Alternatively, I could go with a X299 (such as Asus Prime X200-Deluxe II) but those are like $100 more and for my purposes don't really add to the performance.

[iamaibeta]

Hooked up two PSUs, tried running 1x3090 and 3x3080 on an X299 board with risers. I also plugged in an old 2060 super and 5700xt in eGPU units as well. I just wanted to make sure the setup could run stably if I could feed it enough power.

With all the other stuff (monitors, another computer, and that sort of thing) I also have plugged in, breaker tripped and my dreams of running 4xRTX 3080/3090 at home went dark for good.

I'm just going to stick with 3 GPUs and water cool the two on top when Asus EKWB cards become available.

[RobertJasiek]

PC with RTX 3090 shall have 850W, an RTX 3080 at peak might use 550W so

850W + 3 * 550W = 2500W and boom! Even without adding extra cards.

850W + 2 * 550W = 1950W might work with a 2000W PSU.

What W are your PSUs?

One 2000W PSU should fit into one electric socket in most European countries but I have read that US sockets might be tighter or at least not accept two huge PSUs in one socket.

So you have spent $1200 on the 4th GPU but I guess you earn from selling it:)

***

RTX 3060 in Germany: unavailable online (of course, but who wants that card anyway).
RTX 3060 TI €1469.
RTX 3070 €1199 ~ 1749.
RTX 3080 €2399 (yesterday €1799, sometimes €2700).
RTX 3090 €1940 ~ 3183,25 (including shipping, surprise!).

***

In restrospect, I could have bought an RTX 3080 from Zotac or PNY in October for a reasonable price but I do not want loud mining card designs. Asus TUF has never been available here. MSI GXT only at scalping prices. MSI Suprim were offered €100 ~ 200 above MSRP but I cannot know whether they were actually available or just waiting list / prepayment offers like the one I tried at MSRP. Silent PSUs are extinct. In conclusion, I have never had a chance to build a silent RTX 3080 PC in Germany. My best chance appears to have been RTX 3080 from Gigabyte for roughly €150 above MSRP in late autumn but I dislike their kind of fan noise.

Apparently, I have to wait for RTX 4000 and prey that those cards will be available together with acceptable PSUs at reasonable prices. For the next generation, I might even consider upper mid range RTX 4060 TI or 4070 cards, which could have similar performance as RTX 3080 and, if TDPs should be about 200W, the best could operate at about 31dB. Well, one can dream...

[iamaibeta]

I had 2x 1600w PSUs connected. The 3080s varied (FE, Zotac, and MSI) but none touched 500watts and typically hovered around 300-350 with occasional spikes over 400 even with Nicehash running. The 3090 (Asus Strix) was 350-400 with rare spikes over 500 but never over 550.

Note the spikes are really rare and I suspect that nvidia may eventually patch out the spikes with drivers. Alternatively, I could probably limit power draw more aggressively with minimal loss of performance.

The 3080s are going back to where they belong. I salvaged them from multiple PCs I had lying around (like the VR machine in living room). I actually still need 2x Asus EKWB water blocked 3080/3090s to finish the build. Maybe I'll get the EVGA ones but I am quite taken by EKWB's water block for my favorite O11D XL case.

If not for EKWB's distroplate looking so good, I'd still not go for water cooling. A motherboard with 2x thunderbolt ports and PCIe slots far enough to comfortably air cool 2x3080/3090 would be enough. I can attach two more cards via thunderbolt.

[RobertJasiek]

On 2021-06-03 15:00 in Germany at launch of RTX 3080 TI and maybe LHR models, I have tried but nothing has been available. Paperlaunch again!

[thirdfogie]

Robert, Scan Computers in the UK are also claiming to launch the 3080 TI on 3 June 2021.
I'm not in the market, so I haven't tried to order one.

https://www.scan.co.uk/

[RobertJasiek]

I have noticed Scan Computers long ago and seems to be a reasonable retailer for UK customers. It also ships to the EU but since exit of the UK from the European Market on 2021-01-01 I do not buy electronics etc. from the UK any more because German / EU law (6 months return and full refund in case of defect, 2 years if one also proves initial cause) does not apply any longer etc. Scan Computers offers 14 days return voluntarily but this is not 6 months.

The UK wanted Brexit and such are some consequences. I have heard of British who do not even dare to order clothes from the EU any more because the customs formalities esp. of returns are much too difficult.

I have heard one can easily buy GPUs in Argentina but international orders of electronics is just way too risky and slow, even if one pays the import tax etc.

[Uberdude]

I got my computer from Scan. Happy with it. I didn't want Brexit.

[Mike Novack]

This discussion makes me think of an interesting question. We have go playing software capable of playing stronger than the strongest humans but do we have this yet at equivalent power? For example, a reasonable estimate of power for the human player is about 100w. What happens if we were to apply that limit to the computer?

[RobertJasiek]

100W is about a notebook with RTX 3060 Laptop, which is roughly faster than 1/10 RTX 3080 desktop. So I think currently 100W is slightly weaker than AlphaGo strength, maybe stronger than human but not necessarily always stronger (in typical, practical positions) yet. In the next GPU generation, we can reassess, that is, ---if--- it will be available for purchase.

[thirdfogie]

I got my computer from Scan. Happy with it. I didn't want Brexit.

Me too, in both parts. I guess no-one cares except us.

[lightvector]

KataGo should be easily superhuman at 100W. It might be superhuman at merely several hundred playouts per move, which I'd guess you might reasonably get on a good mobile device (e.g. ~10W) at medium-paced time controls. I wonder if it might start to give pros a strong challenge even at several tens of playouts (e.g. https://online-go.com/player/902691/), because the *raw neural network alone* (1 playout) was capable of holding 8d on KGS many months ago.

I think people often hold AI to much higher standards than they hold human players, without even realizing they're doing so. A bot run at low playouts can be terrible at analysis, make many embarrassing mistakes that even kyu and dan players can see in retrospect, be obviously and frequently self-inconsistent in the scores and evaluations it gives... and yet be capable of winning more than 50% of even games against strong professional players. And I expect that the minimum number of playouts to be superhuman, whatever it is, will result in a bot that is precisely all of these things.

The amount of computation to produce mostly-self-coherent, seemingly-oracle-level analysis (i.e. what people have come to expect out of AI), is multiple orders of magnitude larger the amount of computation merely to be favored in a head-to-head game and lose less than 50% of the time.

[Uberdude]

My suspicion is if you put 100 playouts KataGo on Fox and played an internet game (small main time, 30 second byo yomi) against an unknowing top pro like Shin Jinseo or Ke Jie then KataGo would probably win. But if you told the pro it was KataGo, or they figured it out after several games, then perhaps the top pro would be able to lead the game into the sort of positions bots struggle with at low playouts and manage to win, though could be tough at internet game time limits. It's like when I played as LeelaZero on Fox maybe 2 years ago, so much weaker bots than now, against some Japanese/Taiwanese 8p around goratings #500, LZ just crushed him with superior direction and shape play and he never got to exploit its ladder weakness, whereas there was the occasional game against weaker players who managed to make it mess up a ladder and win, or I could find analysis positions it struggled with.