GPU questions.

Message boards : GPUs : GPU questions.
Message board moderation

To post messages, you must log in.

AuthorMessage
BoincSpy

Send message
Joined: 28 Oct 21
Posts: 7
Message 110976 - Posted: 28 Jan 2023, 21:33:11 UTC

1. I have observed if you have one RTX 30x or RTX 40x and have the build in intel GPU processing working units the rate of processing goes down ~10 %. IE if not running an RTX 3070 ti without intel GPU the processing rate ~31% if we add the intel processor the rate drops to ~19-21%. Anyone know what the underlying cause of might be the slow down.

2. I just purchased a RTX 4070Ti and noted the GPU rate is not much higher than the RTX3070 Ti. IE 32.0 % /min vs 29.0 % /minute. I know it might be related to what WUs processing but thought I would get a better rate / minute IE almost double based on the GFLOPS peak numbers. Here are the specs of the 2 graphics cards. Again any thoughts on way the 4070ti is not really outperforming the 3070Ti.

Computer 1


1/28/2023 10:50:09 AM | | CUDA: NVIDIA GPU 0: NVIDIA GeForce RTX 3070 Ti (driver version 528.02, CUDA version 12.0, compute capability 8.6, 8192MB, 8192MB available, 21934 GFLOPS peak)
1/28/2023 10:50:09 AM | | OpenCL: NVIDIA GPU 0: NVIDIA GeForce RTX 3070 Ti (driver version 528.02, device version OpenCL 3.0 CUDA, 8192MB, 8192MB available, 21934 GFLOPS peak)
1/28/2023 10:50:14 AM | | Windows processor group 0: 12 processors
1/28/2023 10:50:14 AM | | Processor: 12 GenuineIntel Intel(R) Core(TM) i5-10400 CPU @ 2.90GHz [Family 6 Model 165 Stepping 5]
1/28/2023 10:50:14 AM | | OS: Microsoft Windows 11: Professional x64 Edition, (10.00.22621.00)
1/28/2023 10:50:14 AM | | Memory: 7.82 GB physical, 16.16 GB virtual

Average GPU rate: 29% / minute.

Computer 2.

1/28/2023 10:42:31 AM | | CUDA: NVIDIA GPU 0: NVIDIA GeForce RTX 4070 Ti (driver version 528.24, CUDA version 12.0, compute capability 8.9, 12282MB, 12282MB available, 42624 GFLOPS peak)
1/28/2023 10:42:31 AM | | OpenCL: NVIDIA GPU 0: NVIDIA GeForce RTX 4070 Ti (driver version 528.24, device version OpenCL 3.0 CUDA, 12282MB, 12282MB available, 42624 GFLOPS peak)
1/28/2023 10:42:46 AM | | Windows processor group 0: 20 processors
1/28/2023 10:42:46 AM | | Processor: 20 GenuineIntel 12th Gen Intel(R) Core(TM) i7-12700K [Family 6 Model 151 Stepping 2]
1/28/2023 10:42:46 AM | | OS: Microsoft Windows 11: Professional x64 Edition, (10.00.22621.00)
1/28/2023 10:42:46 AM | | Memory: 31.78 GB physical, 63.78 GB virtual

Average GPU rate: 32 % / minute
ID: 110976 · Report as offensive
ProDigit

Send message
Joined: 8 Nov 19
Posts: 718
United States
Message 111640 - Posted: 23 Apr 2023, 0:04:45 UTC - in response to Message 110976.  

1. I have observed if you have one RTX 30x or RTX 40x and have the build in intel GPU processing working units the rate of processing goes down ~10 %. IE if not running an RTX 3070 ti without intel GPU the processing rate ~31% if we add the intel processor the rate drops to ~19-21%. Anyone know what the underlying cause of might be the slow down.



Cpu bus or ram bandwidth bottleneck. Many tests have shown that 20-22 threads on older ryzen cpus start to bottleneck the ram at 3600-3800Mhz ram speed.
Faster ram (5000Mhz DDR4) somewhat alleviates the issue, (can run up to 30 threads fine), but using the IGP aggravates it in 2 ways:

1- An IGP heavily makes use of ram memory both in read and write, and it bottlenecks the cpu interconnects. This causes delays and data collision in ram access, read/writes. Additionally, the igp uses up 4pcie lanes. Not a lot of study is done to see if pcie lanes can cause other lanes to slow down at an interconnect point to the processors, as in some cases it takes 1 to 2 full cpu threads to process 16 lanes. So the more pcie lanes used, the less cpu threads will be available.

2- using the IGP causes the cpu to increase in internal temperature, which in turn lowers boost frequency of both the cpu and igpu core frequencies, and returns lower CPU and IGP PPD. Lower cpu frequency may also adversely effect pcie transfer speeds, causing additional lag on the dgpu, resulting in lower PPD, though the slowdown should be barely registerable. Still, it's a slowdown nonetheless.

My recommendation would be, to just use a DGPU for boinc, and use up to ~80% of your available cpu threads (depending from cpu to cpu), and use the IGP only for display out. If you plan on installing 2 dGPUs, it might be best to turn off the IGP in bios, to free a few PCIe lanes for the dGPUs. Running the monitor from your primary dGPU is ok, if you're not using the pc. But if you also want to play games or watch videos, you might want to keep the igp enabled and connect the monitor through that.
ID: 111640 · Report as offensive

Message boards : GPUs : GPU questions.

Copyright © 2024 University of California.
Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.