RTX 3000 Preliminary price/performance (estimated)

[MeeLee]

Performance Ratio is based on Price/performance (est);
Real life numbers will differ slightly.

2080 Ti: 4352 Cuda Cores, 352 bit GDDR6, $1200 ~3.6-4.1M PPD ; RATIO: 3.2
3070: 5888 Cuda Cores, 256 bit GDDR6, $499 ~3.5-4.0M PPD (EST) ; RATIO: 7.5
3080: 8704 Cuda Cores, 320 bit GDDR6X, $699 ~6.4M-7.2M PPD (EST) ; RATIO: 9.7
3090: 10496 Cuda Cores, 384 bit GDDR6X, $1499 ~9.4M-10.5M PPD (EST) ; RATIO: 6.6

Assuming the est. PPD is correct,
Price/performance, the 3080 seems like it's ahead of the rest, by far!
This thanks to it's high core count, high memory bandwidth, and lower price.
This will probably be the most popular GPU, estimated to have about twice the performance of the 2080 Ti.



Efficiency Ratio is based on PPD (est)/watt * 100 (to get an easier to read number):

2080 Ti: ~3.6-4.1 MPPD, 250W ; RATIO: 1.54
3070: ~3.5-4.0M PPD (EST), 225W ; RATIO: 3.33
3080: ~6.4M-7.2M PPD (EST), 320W; RATIO: 2.12
3090: ~9.4M-10.5M PPD (EST), 350W ; RATIO: 3.01

The 3080 seems lower, that could be due to a too highly rated TDP of 320W. If I deduct the TDP from the core count, it looks like the 3080 will be closer to 300W max.
Meanwhile the 3070's ratio seems rather high (efficient), especially considering it's not using the faster memory bandwidth, and has lower core count.

Either way, these GPUs can be power capped, just like the GTX, and GT series do.
The higher the model (3080, 3090), usually the higher the power savings will be possible (more efficient); making the higher end models more desirable for folding (so long all cores can be utilized).
For the 3090, I would definitely recommend the NVidia reference design.
Because it's halfway a blower, halfway a pass through heat sink design (aftermarket designs won't blow the hot air out of the back of the case, as far as I know).
3 fan aftermarket designs will be better for the 3070 (which could probably be ran at 150-180W in most PC cases just fine).
But for the 3080, I expect that a triple fan design won't be able to cool the GPU sufficiently in a closed case, if the GPU utilizes more than 225-250W.
If you do go with aftermarket designs for 3080 and 3090, I'm expecting you'd have to run the GPU in either an open case, or hanging suspended with PCIE risers.
Unless aftermarket will pick up on Nvidia's design, and will also use a semi-blower style design.
The PCB of the GPUs make this design very interesting, to be ran in smaller mini ITX home PCs (smaller form factor) at a reduced wattage.
The length of the PCB is literally half of a full 2000 series design.

Especially when paired with water cooling blocks, these 3000 series GPUs will be phenomenal! And probably the only way you could run 3 GPUs on one motherboard for folding.

[bikeaddict]

It would be wise to wait and see what Super and Ti variants of these GPUs come out. There have already been leaks of Lenovo machines with RTX 3070 Super and Ti models.
https://www.tweaktown.com/news/74916/le ... index.html
https://www.tweaktown.com/news/74935/le ... index.html

Anyway, it sounds like availability of the 3000 series will be limited until the end of the year, so there may be no choice but to wait.
https://www.tweaktown.com/news/74915/ex ... =tweaktown

[HaloJones]

ti versions will likely just be larger VRAM specs.

but I agree that the 5888 core count etc, looks a little like double counting but we will find out on the 17th.

[markdotgooley]

3080 with a factory “hybrid” setup (water loop to a radiator) perhaps? I think I saw a picture of one; can’t recall the brand.

[foldy]

There were already some CUDA and OpenCL benchmarks and despite having doubled shader count for FP32 they only gave a 50% speedup. Still 5 Mill PPD for a RTX 3080 at $700 and 350 watts is great.

[MeeLee]

The video here, shows that there's a 50% increase of the 3080 over the 2080 (in gaming).
OpenCL might be lower, due to it not making use of a lot of the shader enhancements (specifically RT cores, and possibly some parts of tensor cores).
https://www.youtube.com/watch?v=cWD01yUQdVA

[kiore]

I have an issue with your numbers for current cards:

2080 Ti: 4352 Cuda Cores, 352 bit GDDR6, $1200 ~3.6-4.1M PPD ; RATIO: 3.2

Maybe under Linux, on some units, with the wind at your back this is 'average' ppd for an RTX 2080ti.
I am cautious about over estimating PPD, but am hopeful that moves to utilize the higher performing GPUs will come to pass.

[tomc001]

Performance Ratio is based on Price/performance (est);
Real life numbers will differ slightly.

2080 Ti: 4352 Cuda Cores, 352 bit GDDR6, $1200 ~3.6-4.1M PPD ; RATIO: 3.2
3070: 5888 Cuda Cores, 256 bit GDDR6, $499 ~3.5-4.0M PPD (EST) ; RATIO: 7.5
3080: 8704 Cuda Cores, 320 bit GDDR6X, $699 ~6.4M-7.2M PPD (EST) ; RATIO: 9.7
3090: 10496 Cuda Cores, 384 bit GDDR6X, $1499 ~9.4M-10.5M PPD (EST) ; RATIO: 6.6

Assuming the est. PPD is correct,
Price/performance, the 3080 seems like it's ahead of the rest, by far!
This thanks to it's high core count, high memory bandwidth, and lower price.
This will probably be the most popular GPU, estimated to have about twice the performance of the 2080 Ti.

I don't understand your Ratio value.

For Price/Performance a lower value is better. How is Ratio calculated? Is it supposed to be Price/Performance?

Taking the 3080: $699 / 7.2 = 97. Perhaps $699 / 7,200,000 = 0.000097
Nope. Neither computation matches your Ratio value of 9.7.

[HaloJones]

For those looking at water-cooling a Founders Edition, be aware that it is not "reference design". It is in fact a custom board and as yet, the blocks have not been released from the major manufacturers. Minor manufacturers have been offering reference blocks as suitable for the FE which they are not.

[MeeLee]

I have an issue with your numbers for current cards:

2080 Ti: 4352 Cuda Cores, 352 bit GDDR6, $1200 ~3.6-4.1M PPD ; RATIO: 3.2

Maybe under Linux, on some units, with the wind at your back this is 'average' ppd for an RTX 2080ti.
I am cautious about over estimating PPD, but am hopeful that moves to utilize the higher performing GPUs will come to pass.

Yes, I got 4.4M PPD out of one, with the first Advanced batches of a while back.
A person just had to be lucky to get a few of them in a streak. Most get a combination of fast and slow WUs, lowering the overall performance by a lot.
I presume the same will be true for the higher performing GPUs.

[MeeLee]

Performance Ratio is based on Price/performance (est);
Real life numbers will differ slightly.

2080 Ti: 4352 Cuda Cores, 352 bit GDDR6, $1200 ~3.6-4.1M PPD ; RATIO: 3.2
3070: 5888 Cuda Cores, 256 bit GDDR6, $499 ~3.5-4.0M PPD (EST) ; RATIO: 7.5
3080: 8704 Cuda Cores, 320 bit GDDR6X, $699 ~6.4M-7.2M PPD (EST) ; RATIO: 9.7
3090: 10496 Cuda Cores, 384 bit GDDR6X, $1499 ~9.4M-10.5M PPD (EST) ; RATIO: 6.6

Assuming the est. PPD is correct,
Price/performance, the 3080 seems like it's ahead of the rest, by far!
This thanks to it's high core count, high memory bandwidth, and lower price.
This will probably be the most popular GPU, estimated to have about twice the performance of the 2080 Ti.

I don't understand your Ratio value.

For Price/Performance a lower value is better. How is Ratio calculated? Is it supposed to be Price/Performance?

Taking the 3080: $699 / 7.2 = 97. Perhaps $699 / 7,200,000 = 0.000097
Nope. Neither computation matches your Ratio value of 9.7.

All values are multiplied by a constant to make it more readable.
And yes, it's performance/price.
We're so used to saying 'price/performance', but then the ratio of high price, low performance should be higher than low price, high performance indeed.

A higher overall score means more performance/dollar.
Perhaps this would have made more sense.