RTX 3080 vx RX 6900XT vs RTX3090

[GTX56O]

RTX3080 739$

https://www.techpowerup.com/gpu-specs/g ... 3080.c3621 739$

Pixel Rate
164.2 GPixel/s

Texture Rate
465.1 GTexel/s

FP16 (half) performance
29.77 TFLOPS (1:1)

FP32 (float) performance
29.77 TFLOPS

FP64 (double) performance
465.1 GFLOPS (1:64)


RX6900XT 999$
https://www.techpowerup.com/gpu-specs/r ... 0-xt.c3481

Pixel Rate
288.0 GPixel/s

Texture Rate
720.0 GTexel/s

FP16 (half) performance
46.08 TFLOPS (2:1)

FP32 (float) performance
23.04 TFLOPS

FP64 (double) performance
1,440 GFLOPS (1:16)



RTX3090 1500$
https://www.techpowerup.com/gpu-specs/g ... 3090.c3622


Pixel Rate
189.8 GPixel/s

Texture Rate
556.0 GTexel/s

FP16 (half) performance
35.58 TFLOPS (1:1)

FP32 (float) performance
35.58 TFLOPS

FP64 (double) performance
556.0 GFLOPS (1:64)
-----------------------------------------------------------------


Reference:

RX5700XT 1.200.000 PPD

Pixel Rate
121.9 GPixel/s

Texture Rate
304.8 GTexel/s

FP16 (half) performance
19.51 TFLOPS (2:1)

FP32 (float) performance
9.754 TFLOPS

FP64 (double) performance
609.6 GFLOPS (1:16)


Ya pueden hacer con OPEN CL toda la programacion para que folding at home funciones en la RX6900XT, por en FP64 saca 1.44TFLOPS

Demasiada potencia

[JimboPalmer]

For Folding@Home we can ignore Pixel Rate, Texture Rate, and FP16, we don't do them. The vast majority is FP32 with a sprinkling of FP64.

I am unclear what the effect of CUDA vs OpenCL will be.

Para Folding@Home podemos ignorar Pixel Rate, Texture Rate y FP16, no los hacemos. La gran mayoría es FP32 con una pizca de FP64.

No tengo claro cuál será el efecto de CUDA vs OpenCL.

[Jupis_folder]

The easiest way to see the FAH performance of a GPU is to look at this site:
https://folding.lar.systems/folding_dat ... pd_overall

Some day you will discover the RX6900XT there. I assume not in one of the TOP 5, but the TOP10. As time goes by, (in terms of years), the performance of the new GPUs will rise. This is mainly because
a) The vendors will optimize drivers
b) The manufacturers will optimize their processes
c) The programmers of FAH will find ways to utilize the additional features/stream processors
Look at the current values for RTX 3070 and RTX3080. These low numbers (compared to RTX2080TI) are not the end of the story, just the beginning

[bruce]

a)
b)
c)

d) The planned FAH project to make assignments more consistent with the specific hardware will reach fruition and be rolled out to the FAH servers.

Of course (d) can only work if there is a wide assortment of projects for the server to choose from. There currently is a shortage of new projects so non-optimal assignments are often the only choice. That will always vary from time-to-time.

[LAR_Systems]

First posting of the NAV 21 cards has appeared https://folding.lar.systems/folding_dat ... xt_6900_xt

The challenge is that going forward we will not have specific numbers for each model of the new AMD cards as they are grouped in the same GPU entry on the F@H client. (at least for now)

[foldy]

RTX 3080 . . .4.5M PPD $739
RX6900 XT . . . 3M PPD $999
RTX 3090 . . . . 6M PPD $1500

I would say RTX 3080 wins. But maybe it is possible to run 2 work units in parallel on RX6900 XT as AMD does not use all shaders for only one work unit. That could improve RX6900 XT again

[Joe_H]

The challenge is that going forward we will not have specific numbers for each model of the new AMD cards as they are grouped in the same GPU entry on the F@H client. (at least for now)

From information posted elsewhere, AMD distinguishes between models using the same chipset by a revision ID number. F@h only uses the manufacturer and device ID number fields of a PCI device ID in GPUs.txt. So for now there is no way to tell which model is being used.

This may or may not change at some point in the future. There is some discussion on changing how the client detects a GPU for use and that may show up in a future version or update to the client software.

[tulanebarandgrill]

@foldy - where are you getting the PPD numbers ? I"m doing about 4M PPD (total 5.5M but that includes multicore CPUs) on my 2080Ti. It IS overclocked tho, to about 1980 MHz.

I'd be interested to know who out there is running at about a 20M PPD level and what hardware it takes to do that.

Or frankly anyone > 5M PPD, what are hardware combinations that are known working for 8, 10, 12, 15, or 20M PPD ?

I see some people getting those kinds of numbers and I'm not sure if it is one of these blockchain miner setups or if they just have several computers running like I do.

[Jupis_folder]

The easiest way to see the FAH performance of a GPU is to look at this site:
https://folding.lar.systems/folding_dat ... pd_overall

Some day you will discover the RX6900XT there. I assume not in one of the TOP 5, but the TOP10.

Rank 9. Check.

[Jupis_folder]

I'd be interested to know who out there is running at about a 20M PPD level and what hardware it takes to do that.

Or frankly anyone > 5M PPD, what are hardware combinations that are known working for 8, 10, 12, 15, or 20M PPD ?

I see some people getting those kinds of numbers and I'm not sure if it is one of these blockchain miner setups or if they just have several computers running like I do.

1. Look for a user with enough PPD, you will find some here: https://folding.extremeoverclocking.com ... ist.php?s=
2. Type in the user name here: https://apps.foldingathome.org/cpu

[gordonbb]

@foldy - where are you getting the PPD numbers ? I"m doing about 4M PPD (total 5.5M but that includes multicore CPUs) on my 2080Ti. It IS overclocked tho, to about 1980 MHz.

I'd be interested to know who out there is running at about a 20M PPD level and what hardware it takes to do that.

Or frankly anyone > 5M PPD, what are hardware combinations that are known working for 8, 10, 12, 15, or 20M PPD ?

I see some people getting those kinds of numbers and I'm not sure if it is one of these blockchain miner setups or if they just have several computers running like I do.

I’ve tried it both ways. I have a mining frame but instead of PCIe3 x1 mining risers I use PCIe bifurcation and m.2 to PCIe3 x4 adapters and Desktop x570 motherboards to run up to 6 GPU.

Running 1 big system is great until the Download bug hits then it becomes a PITA having to reboot to restart stuck slots.

Mostly I run 5 dual GPU rigs on x470/570 and z370/390 “SLI-Capable” motherboards with Mid range (2070 Super) GPUs.

I run everything headless under Linux as it’s just easier and I power limit the GPUs with a slight overclock to increase efficiency.

The mining frame is easier to manage heat with. In conventional cases I usually run one Hybrid GPU and one with dual-axial Fans to keep the thermals in check.

I can produce 25-26MPPD at full throttle but generally run power-limited and for 12 hours on weekdays and all weekend to take advantage of lower electricity rates yielding an average of about 15MPPD.

[Skajaquada]

@foldy - where are you getting the PPD numbers ? I"m doing about 4M PPD (total 5.5M but that includes multicore CPUs) on my 2080Ti. It IS overclocked tho, to about 1980 MHz.

I'd be interested to know who out there is running at about a 20M PPD level and what hardware it takes to do that.

Or frankly anyone > 5M PPD, what are hardware combinations that are known working for 8, 10, 12, 15, or 20M PPD ?

I see some people getting those kinds of numbers and I'm not sure if it is one of these blockchain miner setups or if they just have several computers running like I do.

I build one special F@H Setup with two watercooled RTX2080Ti and i have 2 "normal" Gaming PC. One with 2080Ti and another currently with a RX6800xt.
The twin 2080Ti machine runs on Linux 24/7 and do 8M+ PPD. Easy to handle and no errors for month.
Maintenance is only required for the watercooling round about 3-4 times a year. https://forum.alphacool.com/threads/fra ... 43&slide=2

[ajm]

@foldy - where are you getting the PPD numbers ? I"m doing about 4M PPD (total 5.5M but that includes multicore CPUs) on my 2080Ti. It IS overclocked tho, to about 1980 MHz.

I'd be interested to know who out there is running at about a 20M PPD level and what hardware it takes to do that.

Or frankly anyone > 5M PPD, what are hardware combinations that are known working for 8, 10, 12, 15, or 20M PPD ?

I see some people getting those kinds of numbers and I'm not sure if it is one of these blockchain miner setups or if they just have several computers running like I do.

I need one powerful machine for my work and I salvaged old material and built 3 other PCs for FAH when I discovered it.
Main is now an i9-10980XE, a 2080ti and a 2080S (all under water, external, 2 1080x45 rads)
X299 is an i9-10900X and a 2080S (under water, internal, 1 360 and 1 280 rads) (well, and a 1030 right now, but temporary)-> ready (RAM, storage, 10G Ethernet, constant data BU, ...) to serve as a replacement for the main machine in case of emergency
Fold00 is an i9-7940X, a 2070S and a 1080ti (AIO for the CPU, 2070S standard, 1080ti hybrid)
Fold01 is an i7 4930k and 2 1050ti, all air-cooled
All 24/7, win 10. I manage the 3 secondary machines via RDP.


I have a few errors on the main machine as I sometimes have to switch a bit brutally. Sorry about that.

[MeeLee]

I've mentioned before, AMD GPUs have less shaders. The 6900XT has only 5120 stream processors, which is 14% less than an RTX 3070, and should be about as fast, even if the 6900XT shaders run 10-20% faster than the Nvidia Cuda cores.
From my experience folding and doing other projects, core frequency appears to affect performance less than core count.

[foldy]

How much faster is Linux compared to Windows for the new GPUs?