Hello everybody,

i'm new to this project and have a question:
My system is a dual GPU setup using a Nvidia GT 640 an a Nvidia GT 430. In other projects the GT 640 is about 2.5 times faster than the GT 430 but not so in Milkyway. Here the GT 430 completes a standard task in about 20 minutes, where the GT 640 uses about 25 minutes. Also the GPU load is ~80-90% on the GT 430 where the GT 640 reaches only ~55-60& GPU load. Is there a way to improve the performance and load on the GT 640?

---

Aloha, Uli

The applications via this project depend on FP64. In general, consumer Kepler excluding Titan lags behind Fermi architecture due to Kepler having FP64 locked down to 1/24 of the FP32 compute rating. The GT 430 is set at 1/12 of the FP32 rating. By my calculations, the GT 640 still comes ahead due to having a much higher FP32 GFLOPS rating than the GT 430 and should run at approximately 25.5-33.5 GFLOPS FP64 compared to 22.4 GFLOPS FP64 of the GT 430. There are several variants of the GT 640 with different compute ratings.

Based on this, my guess would be that the GT 640 is not boosting to full clock speed due to the GPU not being fully loaded. You can setup an app_config.xml file in the milkyway project folder and run two tasks per GPU to see if this increases the GPU load and resulting frequency.

<app_config>
<app>
<name>milkyway</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.05</cpu_usage>
</gpu_versions>
</app>
</app_config>

Even 'only' running 1 WU at a time the 640 should be achieving a much higher GPU load, maybe there isn't enough spare CPU resources for the GPUs?

Ulrich
If you pause crunching on the 430 does the GPU load on the 640 go up?
Do you leave a spare CPU core for the 2 GPUs when running MW?

---

Team AnandTech - SETI@H, DPAD, F@H, MW@H, A@H, LHC, POGS, R@H, Einstein@H, DHEP, WCG

Main rig - Ryzen 5 3600, MSI B450 G.Pro C. AC, RTX 3060Ti 8GB, 32GB DDR4 3200, Win 10 64bit
2nd rig - i7 4930k @4.1 GHz, HD 7870 XT 3GB(DS), 16GB DDR3 1866, Win7

Hi there,

thanks for the suggestions, i tried them. I freed one core, but the GPU load on the GT 640 remained at about 55-60%. The GT 640 is running at max speed, there is no throttling. Excluding the GT 430 does not influence the crunching on the GT 640. The only way to max out the GT 640 is running two instances via app_config.xml, but then screen lags occur and the GT 430 starts thrashing memory. I think i'll have to live with that. :?

---

Aloha, Uli

Hi there,

thanks for the suggestions, i tried them. I freed one core, but the GPU load on the GT 640 remained at about 55-60%. The GT 640 is running at max speed, there is no throttling. Excluding the GT 430 does not influence the crunching on the GT 640. The only way to max out the GT 640 is running two instances via app_config.xml, but then screen lags occur and the GT 430 starts thrashing memory. I think i'll have to live with that. :?

You could run the 640 here and the 430 on seti gpu units and the 640 could run 2 units at once while the 430 could run only one unit at a time. That should help the thrashing and optimize things.

Hi there,

i finally found the optimal configuration for my GPU setup. I now run two instances of Milkyway per GPU and one instance on CPU. For that i use the following app_config.xml:

<app_config>

<app>
<name>milkyway</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.25</cpu_usage>
</gpu_versions>
</app>

</app_config>

Additionally i raised the "Frequency (in Hz) that should try to complete individual work chunks." from 60 to 100, to prevent screen lags. Now the GPU usage is a constant 99% on both GPUs and the system is responsive and usable without deficits. Thanks everyone for the suggestions! :)

---

Aloha, Uli

Hi there,

i finally found the optimal configuration for my GPU setup. I now run two instances of Milkyway per GPU and one instance on CPU. For that i use the following app_config.xml:

<app_config>

<app>
<name>milkyway</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.25</cpu_usage>
</gpu_versions>
</app>

</app_config>

Additionally i raised the "Frequency (in Hz) that should try to complete individual work chunks." from 60 to 100, to prevent screen lags. Now the GPU usage is a constant 99% on both GPUs and the system is responsive and usable without deficits. Thanks everyone for the suggestions! :)

+1