Folding Forum

IMHO, zero.

Some components of a GPU do computations and other components convert that information into a video image.

The components that put the vidwo image on the display are not used by FAH, but the components that are used by FAH are essential to gaming, as well as other applications for stream computing such as AI. All video cards have to be able to compute 3D geometry.

In the Intel GPU request thread, it was discussed that AMD and Nvidia graphics cards only have 30% of the market; Intel has the rest. Only gamer PCs and professional render workstations (owned by a handful of companies) have non-low profile cards. Low profile cards are intended only to provide additional video-out options and are terrible for folding (I don't think they are even listed). When 5G appears, it will be feasible to stream lossless video over services like Steam and PSNow. My point is, we are assuming that graphics will get "better and better" and folding power will increase proportionally, but the whole foundation depends on fickle gamer taste (not to mention, 5G will make low power "thin clients" worthwhile alternatives for gaming, particularly in poorer regions). One day everyone is waiting for an announcement of the next consumer GPU gen, then one day John Caramack at Id announces that the next Quake will be streaming only, and its all over. We assume that the project's folding capacity will increase, but the reality is that this project and others like it are all hanging by a very thin thread.

I've used LP GPUs from NVidia. They're supported by FAH and are respectable performers. I've got a GTX 1050 (two slots, half height) and I'm ordering a GT 1030 (one slot, half height). The slots on those two PCs are interchanged, so I'm currently running a GT 710 (one slot, half height) which is a poor performer, but does meet the deadlines.

For Single-Precision math, the 710 has only 192 shaders [yielding 366 GFLOPS] whereas the 1030 has 384 shaders [yielding 942 to 1059 GFLOPS]. All support double precision math.

Most of the Intel GPUs do not support Double-Precision, so their percent of the market means almost nothing. For a reasonable approximation of the FAH performance, compare the single-precision GFLOPS number for those GPUs that do support double-precision math (and to a lesser degree, the double-precision GFLOPS number). Ignore ALL of the other advertised features.

All the game benches being published suggest 50% improvement in frame rates for 2080 ti over 1080 ti (without raytracing), so I suppose the project can be expected, participation constancy assumed, to increase its throughput by 50% over the next 3 years.

Frame rates are a measure of the GPU's ability to rapidly generate and refresh screen images from pre-computed 2D or 3D geometry, a capability that FAH does not use. Computation of 3D geometry is done by an independent segment of the GPU. As I said, SP GFLOPS is what you should consider, NOT FRAMERATES.The Framerate is good measure of game responsiveness, whether or not ray-tracing is available. (FAH would be happy with a GPU that has a framerate of zero and which supported zero monitors.)

A 50% increase in throughput may happen but it will depend in increases in shader count and/or shader frequency.

I make games so I understand their tech. The 3D transforms are the largest drain on the CPU. If you've ever watched Futurama where the Doctor creates hyperdrive by "moving the universe around the traveler", that's the same thing games and simulations do. All the polygons in view must be rotated (either that or primitive raytracing ("eye beams") must be used, usually a mixture of the two). The performance of games is a factor of the physical (the CPUs handle the physics and what is usually very primitive AI). Anyway we'll know in a couple weeks what the performance of the cards is.

I found some interesting benchmarks on the power of Iris graphics and also PS4 Pro, Switch, and XBox One X.
https://en.wikichip.org/wiki/intel/micr ... tures/gen9
https://www.gamespot.com/gallery/consol ... 0-1334/16/

Iris graphics aren't that great but nothing bad compared to CPUs when it comes to double precision. PS4 Pro, Switch, and XBox One X all rival top end GPUs in terms of total processing ability. All things considered they are about even with new PCs.

FAH doesn't use facet or beams of rays. No visualization is calculated, just the motions of up to hundreds of thousands of atoms (points) interconnected by force fields that depend on the location of all of the other points. There are a lot of proteins that are intractable because they're bigger than can reasonably be studied with today's hardware.

Stick figures are easy, but I'm sure there are some real challenges for a game to simulate a foamy ocean or fire or clouds (Oh, and you're not allowed to use repeated textures since every atom matters.)-- and those can all be simulated on a 2D surface rather than 3D cloud of atoms like FAH.

Hi,

the first RTX 2080 benchmark ist finaly available:

Nanoseconds per Day

RTX 2080 TI @ 260W TDP 157,70 NpD = 0,606 NpD / Watt
RTX 2080 @ 225W TDP 140,70 NpD = 0,625 NpD / Watt
GTX 1080 TI @ 250W TDP 130,60 NpD = 0,522 NpD / Watt
GTX 1080 @ 180W TDP 108,78 NpD = 0,604 NpD / Watt
AMD RX Vega64 @ 300W TDP 104,90 NpD = 0,349 NpD / Watt

Source :
anandtech

Results are in, there is around 21.4% increase in Cuda cores going from 1080ti to 2080ti, (3584 vs 4352) and the single precision increase is pretty much exactly the same.

https://www.anandtech.com/show/13346/th ... -review/15

In other words, you're paying more for a GPU that has a small increase in FAH performance due to an increase in the shader count @ clock rates but the biggest improvements are in the speed increases and capabilities of the Video and Display Engine used by your favorite game or by VR ... plus you're getting version 0.0.1 of a ray-tracing engine which will be improved (or actually deliverer ) by the next generation(s) of GPUs.

Exactly

And it uses 20% more power to get that hash rate than the 1080Ti

If you could overclock your 1080Ti by 24% then you'd probably match it.

I heared they may release a 7nm shrink gtx 2080 Ti in late 2019 which then will have lower power usage or higher GPU clock. For those not impressed by the current gtx 2080 Ti the next shrinked version may be the GPU to go for.

Hi
Here is some benchmark. Are you surprised?
https://www.phoronix.com/scan.php?page= ... i-FAH-Perf

beer wrote:Hi
Here is some benchmark. Are you surprised?
https://www.phoronix.com/scan.php?page= ... i-FAH-Perf

Nice!
Quick back of envelope calculations suggest 2 million PPD. Thanks to the QRB these might just be worth the outlay!

I've missed the boat on the 2080ti FE, but have just ordered a 2080 FE to play with. Hope to upgrade a few rigs with partner 2080Ti's when (if!) the price drops a bit.

My review of both cards show the same number of CUDA Cores.