RTX 4090 is getting 26M PPD

[psaam0001]

I would be curious to see results from a system using AMD's upcoming RX 7xxx series RDNA3 cards (using OpenCL & no overclocking).

Paul

[Mxyzptlk]

I don’t see AMD being able to fold as well as the Nvidias.

[FaaR]

This is the memory latency effect. Latency between CPU and main/system memory is probably 65-75ns (depending). Latency between GPU and its memory is usually higher, 100ns?, because the special memory used is designed for high bandwidth, not lower latency.
FaH only uses a small fraction of the GPU memory. Its too bad we do not have a special GPU card with smaller low latency memory

Memory latency seen from the perspective of a shader core out to GDDR memory is in the 1000+ cycles range (perhaps several thousand even), due to the huge numbers of outstanding memory I/O operations generated by thousands of shader processors, texturing units, ROPs and so on, all of which are held in on-chip queues, leading to these long service times. So GDDR memory is not inherently high in latency by the way, this is not true. All modern forms of DRAM have similar-ish base latency, counted in milliseconds. GDDR in of itself is not hugely higher in latency than DDR4/5, same with HBM DRAM by the way.

So this high effective latency is normal for a GPU and doesn't affect performance as such as long as there's sufficient work for the GPU to handle. Where a consumer CPU today runs one or two threads per core, modern GPUs spin off thousands, or tens of thousands of threads, parking those which are pending memory I/O and running those which have data available, that's what creates this very high effective latency, but as the GPU is constantly producing huge amounts of work (running a lot of stuff in parallel), you're not suffering from that.

By the way, the 4090 has 96MB on-chip last level cache, far more than any other geforce GPU up until this current generation. So it does have access to a rather substantial amount of comparatively very fast, very high bandwidth memory.

[GTX56O]

Is interesant know how are the values of mV and Mhz of the GPU core and MHz and mV of the memoria for the overclock in the RTX4090, because i see cards about 36M PPD...

[knopflerbruce]

My 4090 only gets about 9-18M PPD.

Will there be more WUs later that actually make this GPU worth the money? At this point, those of us who buy equipment to support the project as our hobby might just as well go for a 4080.

[AZBrandon]

My 4090 only gets about 9-18M PPD.

Will there be more WUs later that actually make this GPU worth the money? At this point, those of us who buy equipment to support the project as our hobby might just as well go for a 4080.

Which OS are you running? I noticed over on the lar systems page there's a pretty big gap between Windows and linux, showing about 15M vs 21M PPD respectively.

[BobWilliams757]

There are some flaws in the LAR systems info, and trends can't always be directly related when one OS is a drop in the bucket for sample sizes vs another OS. The last 500 samples for one might be a week, for the other two months, so it would skew the points trend lines.

The only real comparison is like for like as much as possible, and more than a few people have reported some fairly low points projects with the newer GPU's. I think this was really something to expect, as even the 3 series cards seemed to have more variance. So to some extent for now I think it's kind of a feature of there when certain projects can take advantage of it, but otherwise not guaranteed to give high points returns.

But to some extent this is true going back a couple generations of GPU's.

[knopflerbruce]

I do run Win 10. That said, I don't mind it if the point inflation stops - it's just weird that this 450W capable card doesn't even hit 300W power consumption with F@H. That tells me that there's alot of unused computing power here.

[BobWilliams757]

Even the very lesser cards sometimes will not reach power limits of the hardware. It's quite possible they might with more intensive assignments designed for more powerful series cards.

I'm not sure how F@H will "scale" towards the newer high end cards, but I do feel your pain. Really short of luck, none of us can get the "perfect" projects for any given hardware.


It doesn't seem to be hurting your progress much though, I knew I had recognized the user name.

[Mxyzptlk]

The extra $ and power draw for the 4090 never appealed to me. So I went the 4080 route....

[toTOW]

I do run Win 10. That said, I don't mind it if the point inflation stops - it's just weird that this 450W capable card doesn't even hit 300W power consumption with F@H. That tells me that there's alot of unused computing power here.

There are a lot of circuitry in GPUs that are useless for FAH (everything related to display features) ...

[thiblec71]

People with RTX 4090, could you post some PPD updates?
on linux or Win?
and which project: Cancer, Huntington...

Thanks

[gordonbb]

https://folding.lar.systems/gpu_ppd/bra ... e_rtx_4090

[MeeLee]

Even the very lesser cards sometimes will not reach power limits of the hardware. It's quite possible they might with more intensive assignments designed for more powerful series cards.

I'm not sure how F@H will "scale" towards the newer high end cards, but I do feel your pain. Really short of luck, none of us can get the "perfect" projects for any given hardware.


It doesn't seem to be hurting your progress much though, I knew I had recognized the user name.

It can't and won't. For instance, RTX 4090 has both int and float registers, where only one can be chosen. For folding that would be mostly float, leaving the 'int' parts of the registers unused.

Also ray tracing cores will be unused.

Vram might only use some vram modules (not all modules are needing to be used).

All these things save some power.

In theory, some folding calculations could be done on the smaller rt cores, but I don't think fah has any resources to adopt the code for those cores.

[Joe_H]

In theory, some folding calculations could be done on the smaller rt cores, but I don't think fah has any resources to adopt the code for those cores.

No, that is not going to happen, even in theory. The rt cores are 16-bit, not enough accuracy to be useful for the folding calculations. As best as I can find out, there is no support for using them in a mode where a 32-bit calculation is done using two 16-bit cores. Even if there was, that would be slower than using an actual 32-bit core for the calculation.