RTX cores for folding@home

[warm]

I'm wondering about the potential of the RTX cores playing a role in folding@home, thinking of picking one of these to side fold quite often, any idea what kind of performance we could see to compare ppd with the previous generation? from the RTX 2060/2070/2080/2080ti ?

[Joe_H]

Until these cards actually get into the hands of testers, no idea how the changes in GPU's will contribute to folding performance. Most of the changes appear to be in areas related to image creation on the GPU, that may not help folding which just uses the compute power of the shaders. General improvements in the design over Pascal may contribute, but how much waits to be seen.

[foldy]

From the shader count and little more shader performance
RTX 2080 Ti = Titan V
RTX 2080 = GTX 1080 Ti
RTX 2070 = GTX 1080

The additional Raytracing or AI cores with lower than single precision will currently not be used by FAH.
(But maybe in future if these cores benefit an updated FAH and most users will use these GPUs)

[JimboPalmer]

[This is my interpretation of information publicly available, we all await details from Nvidia.]

Current software uses what Nvidia calls CUDA Cores, although it uses them with OpenCL software.
I am seeing claims that each CUDA Core in Turing (RTX 20x0) is 10% faster than Pascal (GTX 10x0) https://en.wikipedia.org/wiki/Turing_(m ... terization
It also reports much greater memory bandwidth. In the past, memory has not been a limiting feature of F@H. (either size or speed)

Volta had Tensor Cores as well. " A tensor core is a unit that multiplies two 4×4 FP16 matrices, and then adds a third FP16 or FP32 matrix to the result by using fused multiply–add operations, and obtains an FP32 result that could be optionally demoted to an FP16 result." -https://en.wikipedia.org/wiki/Volta_(mi ... e)#Details

Traditionally F@H has used FP32 math augmented where needed by (much slower) FP64 math. It is not clear to me how much FP16 math F@H uses that could be executed on the Tensor cores.

The Ray Tracing Cores are new with Turing and their capabilities to do Compute server math is not known yet.

So lots of well understood CUDA cores which may have increased IPC, Faster memory which may not be important to Compute math, (Fast memory is great for texture mapping in games) Tensor Cores which seem to be too simple for Compute math, and unknown RT Cores.

The good news is that those CUDA cores will be available to F@H as soon as folks can get them whitelisted. Faster memory may make it so memory continues not to be a bottleneck. The Tensor cores and Ray Tracing will be future projects to find some use for, to free up some CUDA cores for the heavy Math.

[foldy]

Even if the new RTX 2080 GPUs may not have much benefits above GTX 1080 for folding they will reduce the price of the GTX 1080. So the gamers buy the RTX 2080 and for folding we get cheap refurbished GTX 1080.

[JimF]

The new cards don't have any real advantage for Folding that I can see, and mainly just increase the core count at a corresponding increase in power (and a disproportional increase in price).

For example, an RTX 2070 has 2304 cores and a TDP of 175 watts (https://einsteinathome.org/content/pasc ... ent-166562). But a GTX 1070 has 1920 cores and a TDP of 150 watts, so the power per core is about the same. And the base clock of the RTX 2070 is 1410 MHz, but the base clock of the GTX 1070 is 1506 MHz.

So unless the cores of the RTX 2070 are sufficiently more efficient (more instructions per cycle) to justify the higher price, you are probably ahead with the GTX 1070, considering that you don't need the extra memory bandwidth of the new card for Folding.

I will wait until the 7nm cards come out before considering an upgrade.

[JimboPalmer]

My current best guess compared to GTX 1080. Eventually we will have real cards to test.

[kiore]

Looking forward to seeing those real numbers. Interesting they are releasing the Ti from the start rather than after a delay, but I guess the opening price makes the no Ti versions competitive.

[bruce]

My current best guess compared to GTX 1080. Eventually we will have real cards to test.

What is the source of your information?

All of the benchmarks that I've found are based on game performance (either with or without ray tracing) and I doubt those benchmarks can accurately predict how the hardware will perform for FAH -- where changes to the performance of the video generation hardware components are inapplicable.

[tcaud]

The price of a video card is no longer determined by its suitability to run games, but by its ability to mine (or even fold). As such, I doubt the price of these cards will come down until a lot of people have them. The advent of mining has made it such that it is computing power relative to other machines that determines the price, because that is the determinant of difficulty. As ability to compute increases in the pool, difficulty rises, which makes cards of a given performance level less profitable.

A reduction in coin payout here at Folding@home, like as not, would do more to lower the price of cards than simply trotting out new ones (though likely insignificantly compared to mining). Coin has made it so that anyone can make a living as an miner/folding admin IF they have enough starting capital and there are a minimum of competitors (of course Bitcoin's price is another factor but there seems to be a lot of effort to keep its return comparable, if not worse, than stocks/bonds). As is, the card makers are determined to get as big a piece of our returns as possible.

I'm frankly unimpressed with the new cards, although to be fair at this point the only thing that could impress me is photo realism. There are political hurdles to consider there (see Final Fantasy movie controversy of 2000), or else we'd probably already be there. The PS3 was a huge leap over the PS2... the PS4 and PS4 pro, and even newer games for PC, not much I'm seeing improved over the PS3 except for higher resolutions of the same basic levels of detail. As for 4k... it's crisper than my actual vision and not something I require for gaming.

[SteveWillis]

Currently the exchange rates are so bad that the coin payout doesn't even pay for the electricity used in folding, or maybe just barely and it's just getting worse.

[bruce]

The price of a video card is no longer determined by its suitability to run games, but by its ability to mine (or even fold).

My point, exactly. The RTX 2000 series provide features that support ray tracing acceleration which add nothing to the ability to mine or fold, While they did adjust the number of shaders slightly (which helps FAH/Mining), the prices went up disproportionately to that change. They're really pushing new developments in gaming.

[tcaud]

My current best guess compared to GTX 1080. Eventually we will have real cards to test.

What is the source of your information?

All of the benchmarks that I've found are based on game performance (either with or without ray tracing) and I doubt those benchmarks can accurately predict how the hardware will perform for FAH -- where changes to the performance of the video generation hardware components are inapplicable.

My understanding is that the top end AMD cards already have raytracing as a feature, so the question might be: how adaptable are ray processors to molecular modeling? I'm skeptical because the advantage of polys is that they fill space... these rays are actually lines running between polygonal surfaces. Carbon is a lattice so it's triangular by nature, that's why GPUs are good for folding. The raytracers might be useful in calculating explosions of force, possibly, but particularly in that the ray tracer itself is an automatic black box that the programmer has no access to... its function is to calculate the angles of virtual subatomic particles (particularly photons) for purposes of shading very small polygons and nothing else.

[bruce]

FAH uses Newton's laws to calculate atomic positions.

Proteins are hydrocarbons, so pure carbon (e.g. in a lattice) probably will never be modeled. Instead, carbon is bonded mostly to hydrogen and to some other atoms. Each atom has a mass and a position and exerts forces on other atoms. A lot of science goes into creating an appropriate force-field and then a lot of calculations go into adding up those forces and figuring out how fast that atom is moving and in which direction. As it moves, forces change so everything has to be repeatedly recalculated. No visual image is generated.

A FAH model has no facets or rays or reflections. The FAHViewer program can create a visual image of the protein from snapshots containing lists of atoms and positions, but that's a NONessential part of the science and very few people even run that program.

[tcaud]

What are the odds that streaming gaming will kill the consumer GPU as we know it?