GeForce RTX 3080 and 3090 support enabled !

[MeeLee]

Sub 3070/2080Ti.
Many tests say it's close to a 2080 Super.

[Gnomuz]

Hi all,

Just a quick feedback on my first WUs with my brand new RTX 3060Ti FE. I was lucky enough to get one of the few pieces available here in France, it was delivered this morning and it's been folding since 4pm UTC, for circa 6h30. F@h is running as a docker container under Unraid 6.9.0 (linux kernel 5.8.18), AsrockRack X470 server motherboard, Ryzen 7 3700X CPU.

It's of course too early to draw conclusions, there's no optimization at all, the drivers are as new as the card (v455.45.01) but so far it looks rather promising (all figures from HFM.NET, consistency checked against FAHControl) :

I completed 3 WUs from project 13428, and the first one was somehow long to start, I think there was some kind of initialization or benchmarking step at the very beginning.
This first WU was completed in 1h55, TPF 1'07", PPD 3,038k
The next two WUs were roughly identical, around 1h53 elapse, TPF 1'06", PPD 3,123k avg, so a bit better
The fourth WU is @43%, 3,131k PPD so far.

The GPU is running stable at 77 C° (in an under-optimal case for the moment), and the power draw is 191-194W (power limit @200W, sw_power_cap throttling permanent, power state P2).

Compared with my previous folding GPU (Quadro P2000), it's clearly not the same league !

Then I just had a look at folding.lars.systems GPU database, and I'm a bit surprised by the figures for this same project 13428 :

RTX 3060 Ti - 2,659k PPD
RTX 2080 Super - 2,891k
RTX 3070 - 3,082k
RTX 2080 Ti - 3,597k
RTX 3080 - 4,963k

Perhaps most of their samples come from Windows rigs, but I think their database is clearly not large enough on RTX30 to be reliable (88 data samples for all projects from RTX 3060 Ti, 5,973 from RTX 3080, 8,723 from RTX 3070). I hardly believe the RTX 3060 Ti will be on par with RTX 3070 (mine at 3,125k vs 3,082k), with a TDP of 200W vs 220W, and an MSRP 100€ lower, it's too nice to be true !

I'll keep watching, do not hesitate to ask if you'd like to have other information.

[kiore]

Thanks for that Gnomuz, there will be lots of variation between how different work units look on different cards, but from your initial data the 3060ti looks like a nice one for folding. NVIDIA does weird stuff with models especially ti models so I would not be surprised if an xx60ti was not so different from an xx70.

[MeeLee]

Perhaps some people remember the 1070 Ti, vs the 1080?

[chabgood]

i am getting 3.4M on my 2080 Supers on ubuntu. so 4 is not that great, for the next gen cards.

[Gnomuz]

Back after 40h of continuous folding with the new RTX 3060 Ti !

So far it's rock stable (not a single WU failed, no temps issue), and the performance reported by HFM.NET is extremely stable. On 20 relevant data samples so far, the average is 3,123,048 PPD (standard deviation 2,756 PPD or 0,09%). TPF is reported as 1'06" on all WUs.

Still I have a question for veterans here. Before getting this new GPU, I folded with a Quadro P2000, and I completed 119 WU from 12 different projects (81 from project 13428, 38 from others). Since I started with the RTX 3060 Ti, it was assigned all WUs from project 13428.
Is it a pure coincidence due to the WUs availability and the short observation period, or is there any determinism behind ?
I had a look at GPUs.txt, and noticed P2000 is in group 7, whereas all RTX are in group 8. I wonder whether that explains I'm assigned WUs considered as more "GPU-intensive" ?

Thanks in advance for your feedback.

[toTOW]

When a Moonshot Sprint is running, you get almost exclusively WUs from p134xx ... which has been the case for the past weeks ...

[Gnomuz]

OK I got it, I'm still new to F@H, but I try to learn
Bon dimanche à toi

[MeeLee]

i am getting 3.4M on my 2080 Supers on ubuntu. so 4 is not that great, for the next gen cards.

Considering that the RTX 3060 Ti is the lowest of the RTX 3000 GPUs, and costs like $500, vs RTX 2080 Super being $850 at the moment (up from $700 before Covid), even this 3060 TI is quite a good deal!
It is possible that the results for the 3060 Ti are well balanced, because it falls in about the same category as the higher end RTX 2000 series GPUs.
The 3080 and 3090 might not even hit maximum load, as they could fall into an entirely different category (the 3090 having more than twice the amount of Cuda cores compared to the 2080 Ti).
There may just simply not yet be many WUs available that can make full use of GPUs with +5000 cuda cores?

[bruce]

There may just simply not yet be many WUs available that can make full use of GPUs with +5000 cuda cores?

I'm sure this is true, though what really matters is how many cuda cores process FP32 instructions. Half-precision cores, RT cores, and Tensor cores don't help FAH and FP64 cores only contribute a small percentage.

[foldy]

RTX 3080 e.g. has 8.704 FP32 units (nvidia doubled them compared to previous GPU gen). If not all cores can be used by a single FAH work unit then workaround could be to run 2 work units concurrently.

[chabgood]

my 2080's get more than 2X what my 1080's were doing. I am not comparing a 2080 to a 3060.

[MeeLee]

my 2080's get more than 2X what my 1080's were doing. I am not comparing a 2080 to a 3060.

That's not because the card is 2x faster.
It's because of improved GPU cores and cuda support.
Also, because the early return bonus disproportionally rewards points.
Eg: If your GPU is exactly twice the cores, returning the WU in exactly half the time it took your first GPU to process, you'll get much more than twice the PPD. Sometimes up to 7x the PPD for halving the process time would be my guess.

[Joe_H]

Eg: If your GPU is exactly twice the cores, returning the WU in exactly half the time it took your first GPU to process, you'll get much more than twice the PPD. Sometimes up to 7x the PPD for halving the process time would be my guess.

Bad guess. If you do WUs in half the time, the PPD will always be 2.8 times doing them in the same time. This goes right back to the QR formulas - https://foldingathome.org/support/faq/points/, see the How is QRB Determined? entry.

Basically, if WUs are processed in 1/N of the time, the ratio for the increase in PPD reduces to N times the square root of N. Applies to both CPU and GPU folding.

[Bastiaan_NL]

On my cards I see the following with P13428 on Windows:

40 seconds tpf on a 3090 with 6.4 million ppd
52 seconds tpf on a 2080ti with 4.5 million ppd
64 seconds tpf on a 2080 super with 3.3 million ppd

So if your 3060ti is at 3.1 million ppd with 190W, that sounds pretty good to me considering that it's brand new (drivers etc).