New Projects & AMD Vega 64

[Sven]

Seeing very high PPD for this Project on my Vega 64. On most projects I get around 800k PPD, but with the 14229 it's around 1,4 Million PPD.

So also AMD GPU-Hardware comes into good use for Folding@Home, hopefully the problem with the 5700 Series can be fixed soon.

[toTOW]

Vega likes when there's a lot of atoms in a WU ...

[FaaR]

Getting WUs from project 14184 on my Vegas; they give LESS THAN 250k PPD... Kinda wasteful running a GPU which regularly slurps around 200W for just the main ASIC for so few points. Electricity is a precious resource after all, we shouldn't squander it.

Can't you steer big projects only to big GPUs...?

[bruce]

FAH does make attempts steer small projects to clients with smaller hardware. Unfortunately there are many degrees between big and small and many steps in between.

What rule would you suggest?(and do NOT mention PPD. Use shader count and atom count. What are N and K so that Atom count> N for shader count > K ; atom count <N for shader count < K

Assuming (probably incorrectly) there is a shortage of big WUs assignable to your device, the AS will assign whatever is available.

What are the GPUID codes for your Vega64 and devices you consider to be similar?

[FaaR]

FAH does make attempts steer small projects to clients with smaller hardware.

Ok, good then. Just have to hope more projects release more big WUs! With GPUs getting faster and faster, surely it would be beneficial to dole out bigger and bigger WUs, methinks?

What are the GPUID codes for your Vega64 and devices you consider to be similar?

Not sure I understand. What GPUID code is it you want and where do I find it? Not sure what you mean by what devices would be similar either; I have 2x ASUS Strix RX Vega 64...so any other AMD Vega 64 or even 56 would therefore be similar? *shrug*

My cards:
PCI Device ID is 1002 687F
Subsystem ID is 1043 04C4

...In case that helps, or means anything to you.

[bruce]

The PCI device ID 1002 687f is called a "Vega 10 XL/XT [Radeon RX Vega 56/64]" in FAH's gpus.txt file.

There are currently 2 different devices called a Radeon RX Vega 64 (or a Vega10 XT) with different clock rates. There is also a Radeon RX Vega 56 (which is also known as a Vega10 XL). I expect that the 687f device is only one of the three.

[Sven]

Project 14286 => 1,5M PPD

Being curious about the 14283 with its even higher Atom-Count. Got one through night, but haven't seen the PPD. I will report, when I see the PPD of this workunit on the Vega 64.

Edit: So now I got a 14283, nearly 1,7 Million PPD. Good to see Radeon's are now also in good business for Folding.

[foldy]

Interesting so AMD Vega GPUs with 4000 shaders is similar to rtx 2080 ti with 4000 shaders which also needs big projects with many atoms to be fully used.

And Vega 64 gets the same PPD as a RTX 2080 but Vega 64 at $400 vs RTX 2080 at $700

Maybe also Vega Frontier Edition and Radeon VII with also 4000 shaders have the same issue.

If we had GPU categories at assignment server we could put all these GPU types together to get preferable high atom count work units.

[bruce]

If the shader count and and clock speed are identical, the peak GFLOPS will be the same (unless the PCIe bus can't supply enough data for the shaders to maintain peak speed).

[MeeLee]

Interesting so AMD Vega GPUs with 4000 shaders is similar to rtx 2080 ti with 4000 shaders which also needs big projects with many atoms to be fully used.

And Vega 64 gets the same PPD as a RTX 2080 but Vega 64 at $400 vs RTX 2080 at $700

Maybe also Vega Frontier Edition and Radeon VII with also 4000 shaders have the same issue.

If we had GPU categories at assignment server we could put all these GPU types together to get preferable high atom count work units.

More like a 2070 to 2080. I wouldn't say it's like a 2080 Ti, because that one gets into the 3.3M PPD (on some WUs), but 2,4M PPD on most Wus.
The 2070 gets 1,21-1,7M PPD on most WUs, and the 2080 gets 1,4 M PPD (I haven't yet seen it do a fast WU, but wouldn't be surprised if some WUs run at 2.2M PPD).

[foldy]

I meant it has the same issue like rtx 2080 ti having too many shaders for small atom count work units. With large atom count work units it matches RTX 2080 from Sven's report.

[Zangetsu]

I have the same situation,

Project 14280 & 12465 so far has had this bump in PPD (and a lower TPF).
is a part of collection 14260-14289 (can't post link to project as i'm a new poster)

It even surprised me as i'm using a ex-miner pc:
3 X Vega64 stock settings, no overclock.
Celeron G4400 (2 processors).
8GB RAM
PCIe gen2 1X connection (with risers v007).
On core21 (adv flag on)

The bottlenecks of this setup (CPU&PCIe gen2 1X) gives me 600K PPD normally, with the high atom numbers projects (i think), 1.1-1.2M PPD (for each GPU)
Difference of TPF is 4 min ("normal" project) to 1min 22sec ("big atom number" project)
I haven't gotten a high atom number project so far on more then 1 GPU, so i can't give any info about that yet.

I think the Radeon VII would benefit from these "big atom" projects as well, as they share a similar architecture(part of the Vega line)

[foldy]

FAH currently tests new FAHCore_22. Maybe that gives a speedup for GPUs with high shader count too?

[MeeLee]

Ok, good then. Just have to hope more projects release more big WUs! With GPUs getting faster and faster, surely it would be beneficial to dole out bigger and bigger WUs, methinks?

Higher atom count isn't always beneficial.
What they're really aiming for, is the interaction of only a few atoms in a cluster. The rest of the atoms are there, because they don't know where the trigger will manifest.

Like in an extreme example, they could fold an entire human being with a couple of exatrillion atoms, just to find that the cancer only affects the liver.
It would make more sense to simulate the cancer on liver tissue, instead of an entire human.
What would improve the process, is if scientists knew where they had to Target the folding to.
Currently, over 99.9999% of our effort, are lost due to inefficient programming. But since scientists don't know where to start looking, it's the best we currently have.

[Zangetsu]

I've had the project on 2 GPU's now,
They both had the same increase of PPD and lower TPF as i had when only 1 was assigned to it.
the 3th GPU got a "normal WU" and did not suffer from lower PPD or higher TPF.
So no extra bottleneck from the PCI-e risers , nor the 2 core processor.
The GPU's do have to work harder => they use more watt , and generate more heat.
I'm using a very aggressive cooling fan graph so my temp rose from 50 to 55°C.

FAH currently tests new FAHCore_22. Maybe that gives a speedup for GPUs with high shader count too?

Core22 gives an 20% increase in PPD overall, as i can tell by the beta testers forum.
This project (big atom projects) gives an increase of X2 PPD.