Hello,
I apologize in advance for typos, I have to use a translator.
EPYC build:
CPU(s) - 128 x AMD EPYC 7601 32-Core Processor (2 Sockets)
Kernel Version - Linux 6.5.11-8-pve (2024-01-30T12:27Z)
Boot Mode EFI
MB - Supermicro H11DSi-NT
proxmox:
in - 8.1.4
VM - Win 10, Win 11
We are testing gpu passthrough and we have a problem with the maximum gpu load in the VM on the main build with EPYC CPU. In the benchmark, it is not a problem to use the graphics at 100% (UNIGINE Benchmarks), but with more demanding games, we are somewhat limited. When testing in passmark, all values seem to be normal (cpu, hdd, ram, gpu). We use 2 GPUs for testing, both have different workloads:
E.g. Gta 5 :
Rx580 - uses gpu max at 50%
1080ti - uses gpu max at 10%
It is interesting that if we set the details to low or ultra, only the minimum fps always changes, but the gpu usage remains within the specified (or smaller) values. When testing battlefield 4, there is less gpu usage and less fps at low details than on ultra. This means - the more details, the higher fps. When the detail is set to low, it is not possible to play. But this only applies to BF4.
We tried overclocking the CPU because this performance would limit the clock speed of a single core, but the difference was minimal.
In the first attempts, only the rx580 had 25% less performance in the benchmark. After checking with the second PC, we found an inactive Resizable BAR in the VM. After turning on 4g decode in the BIOS, the benchmark in the VM already used the gpu at 100%, but the resizable BAR - BAR0 remained at 256MB (H11DSi-NT does not have a setting option in the BIOS). 1080ti doesn't support resizable bar, that's why it ran at 100% in our benchmark even with 4g decoding turned off, but it also causes problems with usage in games.
With resizable BAR turned off on the second test PC with an i7, the gpu runs at 100% even in games with only 2 threads assigned. This ruled out BAR causing the problem.
We use i7 for testing proxmox and windows. Verification of gpu performance etc. Then we test and compare with EPYC. The Proxmox version and settings are identical (except for the Intel/AMD setting differences) and the gpu usage problem is only on the EPYC build.
Does anyone have an idea or explanation as to why this is? Thanks
I apologize in advance for typos, I have to use a translator.
EPYC build:
CPU(s) - 128 x AMD EPYC 7601 32-Core Processor (2 Sockets)
Kernel Version - Linux 6.5.11-8-pve (2024-01-30T12:27Z)
Boot Mode EFI
MB - Supermicro H11DSi-NT
proxmox:
in - 8.1.4
VM - Win 10, Win 11
We are testing gpu passthrough and we have a problem with the maximum gpu load in the VM on the main build with EPYC CPU. In the benchmark, it is not a problem to use the graphics at 100% (UNIGINE Benchmarks), but with more demanding games, we are somewhat limited. When testing in passmark, all values seem to be normal (cpu, hdd, ram, gpu). We use 2 GPUs for testing, both have different workloads:
E.g. Gta 5 :
Rx580 - uses gpu max at 50%
1080ti - uses gpu max at 10%
It is interesting that if we set the details to low or ultra, only the minimum fps always changes, but the gpu usage remains within the specified (or smaller) values. When testing battlefield 4, there is less gpu usage and less fps at low details than on ultra. This means - the more details, the higher fps. When the detail is set to low, it is not possible to play. But this only applies to BF4.
We tried overclocking the CPU because this performance would limit the clock speed of a single core, but the difference was minimal.
In the first attempts, only the rx580 had 25% less performance in the benchmark. After checking with the second PC, we found an inactive Resizable BAR in the VM. After turning on 4g decode in the BIOS, the benchmark in the VM already used the gpu at 100%, but the resizable BAR - BAR0 remained at 256MB (H11DSi-NT does not have a setting option in the BIOS). 1080ti doesn't support resizable bar, that's why it ran at 100% in our benchmark even with 4g decoding turned off, but it also causes problems with usage in games.
With resizable BAR turned off on the second test PC with an i7, the gpu runs at 100% even in games with only 2 threads assigned. This ruled out BAR causing the problem.
We use i7 for testing proxmox and windows. Verification of gpu performance etc. Then we test and compare with EPYC. The Proxmox version and settings are identical (except for the Intel/AMD setting differences) and the gpu usage problem is only on the EPYC build.
Does anyone have an idea or explanation as to why this is? Thanks
Last edited: