Will the 4080 Super's New Price Make it a Better Value for Folding?

[Gorgon]

With a mid-series refresh under way for NVIDIA's Ada GPUs I started to wonder what the changes mean in terms of value for Folding.

 

The short answer is that even with a $200US drop in price of the "mid range" GPUs it is a better value but still can't match the value of the new 4070 Super or even the 4060 in terms of Shaders/$

	Price	Shaders	Clock	TDP	Value	Value
Model	($US)	(#)	(MHz)	(W)	(shClk/$)	(sh/$)
RTX 4090	$1,599	16,384	2,235	450	22,901	10.2
RTX 4080s	$999	10,240	2,295	320	23,524	10.3
RTX 4080	$1,199	9,728	2,205	320	17,890	8.1
RTX 4070ts	$799	8,448	2,340	285	24,741	10.6
RTX 4070t	$799	7,680	2,310	285	22,204	9.6
RTX 4070s	$599	7,168	1,980	220	23,694	12.0
RTX 4070	$599	5,888	1,920	200	18,873	9.8
RTX 4060t	$399	4,352	2,310	160	25,196	10.9
RTX 4060	$299	3,840	1,830	115	23,502	12.8

My personal preference for a new build would now be the RTX 4070 Super as it is showing nearly as good value as the 4060 and is still a dual slot card making it much easier to implement dual GPUs in motherboards with 3-slot spacing for the PCIe expansion slots.

[bl4kers]

Quote

My personal preference for a new build would now be the RTX 4070 Super...

Out of curiosity, what would be your pick on the AMD side?

[Gorgon]

12 hours ago, bl4kers said:

Out of curiosity, what would be your pick on the AMD side?

I don't use AMD for Folding as they refuse to port ROCm to their Consumer GPUs and Folding@Home and thus their GPUs are forced to to OpenCL so their Folding@Home performance is much lower than an Nvidia GPU with similar rasterization performance that can take advantage of the work Nvidia has put into the CUDA libraries.

 

In general though the Folding performance will be directly proportional to the Single Precision Floating Point performance (FP32) so if you divide that by the Cost it should offer a good indication of the value of a particular model for just Folding use.

 

Folding at Home does not use Ray Tracing nor does it use a significant amount of VRAM so VRAM Size nor the size of the GPU Memory Bus Width should have much impact, if any.

 

Some AMD models, however, have severally limited amounts of Double Precision (FP64) cores which may impact their performance on Work Units that require these calculations as they will have to be off-loaded to the CPU with a significant performance penalty.

[Gorgon]

16 hours ago, bl4kers said:

Out of curiosity, what would be your pick on the AMD side?

Also, you can look at @LAR_Systems excellent GPU Folding Performance Database and /or use the FP32 Specifications for various GPUs from sites like TechPowerUP.

 

Simply divide the Performance (PPD from LAR_Systems; OR FP32 from TechPowerUP) by the Retail Price in your area and that should give you a good idea of the Value of a specific model. Of course, if you want/need to use the GPU for other things such as gaming that can influence which GPU is the best for your specific use cases.

 

It's not that I don't like AMD GPUs, I actually prefer them over NVIDIA, as they are, IMHO, much better value  and my daily driver is a FirePro W4100, but for just Folding they just don't make sense. Some older Models, like the Radeon 7, are also really strong performers for other Distributed Computing applications (BOINC) as they have excellent FP64 performance, but as I mentioned AMD, Like NVIDIA, seems to be intentionally "nerfing" FP64 performance these days on their consumer GPUs, likely to drive users needing good compute performance to their much more expensive Enterprise Product Lines.

[vsteel]

On 1/18/2024 at 10:53 AM, Gorgon said:

I don't use AMD for Folding as they refuse to port ROCm to their Consumer GPUs and Folding@Home and thus their GPUs are forced to to OpenCL so their Folding@Home performance is much lower than an Nvidia GPU with similar rasterization performance that can take advantage of the work Nvidia has put into the CUDA libraries.

 

On Linux you can now run ROCm on the 7900 XTX and XT.   

 

I don't have either of those cards so I can't say how they work compared to the OpenCL performance. 

[Gorgon]

13 hours ago, vsteel said:

On Linux you can now run ROCm on the 7900 XTX and XT.   

 

I don't have either of those cards so I can't say how they work compared to the OpenCL performance. 

Thanks, good to know. I haven't dived down the Rabbit Hole of AMD GPUs lack of performance in OpenMM or HIP in Folding@Home in a while but it appears that progress is being made.

 

But the "proof is in the pudding" as they say and looking at a large Atom WU with good results such as Alheimer's p12424 in @LAR_Systems database we still see the AMD GPUs vastly under performing compared to NVIDIA. Granted, a direct comparison is hard to determine due to AMD's annoying habit of using the same GPUID for cards in the similar class (7900XTX/7900XT; 7800XT/7700XT ...)

 

This discussion of AMD versus NVIDIA for Folding should be moved to a separate thread.