New RTX3xxx cards

[Neil-B]

Watched the launch live ... hugely more entertaining than the last launch I watched/attended (but that was Windows 95) !!

[gordonbb]

Back of napkin calculations show a 3070 172% more FLOPs than a 2070 (1.3 to >3.5MPPD) and a little more than twice as efficient. So more than the folding power of a 2080Ti at less than half the cost.

Now all we’ll need is the scientists to release lots of large molecule models to leverage all the shaders. I’m sure they’ll be up for the challenge.

[Neil-B]

Good job work is ongoing to manage assignment at a finer granularity than currently ... Hopefully it will arrive in time to keep RTX3xxx folders happy (and hopefully address some of the challenges the RTX2xxx folders have had ... But until an improved assignment regime arrives there might be some frustrated/disappointed RTX3xxx owners

[foldy]

Nvidia GPUs . . . . . . CUDA Cores
GeForce RTX 3090 . . 10496
GeForce RTX 3080 . . 8704
GeForce RTX 3070 . . 5888

FAH performance was bad on GPUs with large shader count and work units with low atom count. So these new cuda cores almost doubled compared to previous GPU generation maybe cannot fully be used by FAH small work units. If true then running 2 or more FAH work units concurrently on GPU may be a workaround. But that also uses more CPU and pcie bandwidth.

For silent PC the Watt usage is too high on RTX 3090, it will make some noise on full load

[Frontiers]

It looking like not old CUDA cores with FP32, it looking like Shader Cores with core count based on FP16 gaming rendering.
So FP32 compute performance wouldn't be doubled compared to previous gen, but will have some x1.2-1.4 bump.
F@H using single precision FP32, not half precision FP16.
But if it's real FP32 cores in these very high counts - this will be sort of miracle, with very high PPD at many atoms projects.

[markdotgooley]

It looking like not old CUDA cores with FP32, it looking like Shader Cores with core count based on FP16 gaming rendering.
So FP32 compute performance wouldn't be doubled compared to previous gen, but will have some x1.2-1.4 bump.
F@H using single precision FP32, not half precision FP16.
But if it's real FP32 cores in these very high counts - this will be sort of miracle, with very high PPD at many atoms projects.

I think they're claiming that it's FP32 on their "CUDA cores" so maybe there will be a good boost in performance.

AMD does seem to be building the new GPUs so that there is no dedicated RT hardware: shaders can be used for RT as desired. That means that for folding there would be no big idle area full of RT units: all the shaders would be used as shaders. This might make those AMD cards better value for money for folding -- if the prices are good, if the drivers are reliable, if performance/watt is competitive. 7nm process rather than 8nm, but exactly what those mean is a bit nebulous nowadays. Wait and see, I guess, and, if you're brave or wealthy or both, risk taking a loss to help clear the way for others...

[foldy]

As I understand the technical details, nvidia previously had FP32+Int32 in one shader. Now they have 2xFP32+Int32 in one shader and that is why it doubled.

[gordonbb]

It looking like not old CUDA cores with FP32, it looking like Shader Cores with core count based on FP16 gaming rendering.
So FP32 compute performance wouldn't be doubled compared to previous gen, but will have some x1.2-1.4 bump.
F@H using single precision FP32, not half precision FP16.
But if it's real FP32 cores in these very high counts - this will be sort of miracle, with very high PPD at many atoms projects.

The FLOPs mentioned are Single Precision:
3070 20.31 TFLOPS 220W TDP
3080 29.77 TFLOPS 320W TDP
3090 35.58 TFLOPS 350W TDP

[MeeLee]

From a power perspective, I think those cards will probably run more the way some of us have capped power on them.
Their OC/power capping potential may be a lot lower than the 2000 series.

That being said, Boinc still has some GPU projects that can utilize those GPUs to the fullest!

Despite them having a higher core count, their frequency is also lower, so the 3070 should match up closer to between a 2080 Super and a 2080Ti. I don't expect it to outperform a 2080Ti.
Nivida has never followed that line, unless they rename it to a 3080. Then it could outperform a 2080Ti.
I'm just speculating here.

Another thing to take note of, no more GPUs running side by side, due to it's cooling design.
The front is blowing the hot air out of the case, but the back fan will be blowing the hot air straight in the next GPU.
And if you have one of these mini itx cases, it'll suck cool air in, and blow hot air straight on the CPU and mobo.
I wished they'd reverse the fan direction. I'd rather have it suck air out of the case, and blow it out of the side of the case.

Also, another problem might be if you use risers, it won't be possible anymore to tilt them horizontally.
That means only 1 GPU per PC max, in most scenarios.
Then again, doubling the core count, a 2080Ti needs PCIE 3.0 x4 in Linux, x8 in Windows,
Chances are you'll need an x8 slot in Linux, and a full size x16 slot in Windows.
Perhaps good for those that run Linux, no more worries about running 18.04 for multi GPUs. Now a single one will do.
It's such a big change compared to running 8 to 14x GTX 1060s or P106 GPUs!
Not only a lot more simple to maintain, it's easier to cool.

Lastly, they're supposed to have different connectors than the 6+2 power pin most GPUs use.
It'll probably take a while before adapter cables are on the market. And hopefully they're affordable!
I wouldn't mind running 2x GPU cables (4x 6+2 connectors) to one GPU if it needs it.
I just don't want to buy new PSUs!

The 3090 would be my goto card, unless the price is too high, then I'll go with a 3080.

[foldy]

Power connectors 2x 8pin(6+2) is needed. And custom boards from Asus, MSI, ... maybe have different fan cooling design. And GPUs now support pcie 4.0 so RTX 3000 even with doubled performance can run without bottleneck in PCIE 4.0 x4 in Linux, PCIE 4.0 x8 in Windows.

[MeeLee]

Yeah, I noticed the 3070 has 2 normal fans. The special fan design is for the reference gpus 3080 and 3090

[HaloJones]

Given that SLI is not possible any more with these cards, NVidia assume you won't have multiple cards any more. Those of us building FAH multi-cpu rigs aren't realy a market consideration for Nvidia.

What's bizarre right now is pricing. Nvidia's online store still has 2070S at £489 for delivery now and the 3070 on "notify" at £469! I'm tempted by a 3070 at that price if it lets me retire multiple 1070s.

[gunnarre]

The 3090 will actually have an NVLink connector for SLI - maybe that makes sense for some compute specific workloads, but it's not utilized in many games anymore.

[JimboPalmer]

"One of the key design goals for the Ampere 30-series SM was to achieve twice the throughput for FP32 operations compared to the Turing SM. To accomplish this goal, the Ampere SM includes new datapath designs for FP32 and INT32 operations. One datapath in each partition consists of 16 FP32 CUDA Cores capable of executing 16 FP32 operations per clock. Another datapath consists of both 16 FP32 CUDA Cores and 16 INT32 Cores. As a result of this new design, each Ampere SM partition is capable of executing either 32 FP32 operations per clock, or 16 FP32 and 16 INT32 operations per clock. All four SM partitions combined can execute 128 FP32 operations per clock, which is double the FP32 rate of the Turing SM, or 64 FP32 and 64 INT32 operations per clock." - https://www.reddit.com/r/buildapc/comme ... om_nvidia/

So if F@H can minimize the INT32 operations, they can approach twice the FP32 throughput of earlier cards. I am not overjoyed calling this twice as many CUDA cores, but it may have the same effect.

Pascal: 1 FP32 OR 1 INT32 per clock cycle
Turing: 1 FP32 AND 1 INT32 per clock
Ampere: 1 FP32 AND (1 FP32 OR 1 INT32) per clock

In an ideal word, Hopper: 2 FP32 AND (1 or 2) INT32 per clock, but that is yet to be seen.

[MeeLee]

Given that SLI is not possible any more with these cards, NVidia assume you won't have multiple cards any more. Those of us building FAH multi-cpu rigs aren't realy a market consideration for Nvidia.

What's bizarre right now is pricing. Nvidia's online store still has 2070S at £489 for delivery now and the 3070 on "notify" at £469! I'm tempted by a 3070 at that price if it lets me retire multiple 1070s.

Even if SLI is supported, how will you connect them?
2 GPUs of ~250-350Watts each, is 500-700W of heat generated in an area smaller than a shoe in size, can't be good.
Most motherboards having 3x full sized PCIE slots, can only use the first and the last one, often the first running at 8x, and the last one running at 4x.
Most modern motherboards run the first slot at Gen 4, and slot 2 and 3 at Gen 3.0 speeds, save for the high end boards of $250 and up.