I wasn’t able to get those properties to reveal themselves to me (with my kernel config) so I hardcoded pf_get_min_spare_lmem to SZ_1G instead of SZ_128M. I tried SZ_2G but got an out of space error. Been playing with it too long to go much deeper than that.
If you go the debugfs route, there appears to be a way to provision memory for VFs individually:
/sys/kernel/debug/dri//sriov/vf1/tile0/vram_quota
/sys/kernel/debug/dri//sriov/vf1/tile0/ggtt_quota
2 Likes
Does it exist only in 6.18?
1 Like
Great recommendation. I’m using it with looking glass with my B580 and it works great. Can’t wait for my B50 to come in so I can test it out.
1 Like
All of the quirky issues I’ve been experiencing with the Pro B50 in Windows and/or with PEG bifurcation on my Asus Pro WS X570-ACE board went away with the latest motherboard BIOS (5302). Possibly a coincidence… maybe clearing the CMOS or reflashing the old version would have led to the same result… but I thought I’d mention it just in case it helps anyone else.
Anyone having this happen when they launch a vm attached to one of the vfio-pci devices?
[ 210.325676] xe 0000:0e:00.0: [drm] Xe device coredump has been created
[ 210.325682] xe 0000:0e:00.0: [drm] Check your /sys/class/drm/card0/device/devcoredump/data
[ 210.784184] xe 0000:0e:00.0: [drm] [CRTC:145:pipe B] DSB 0 queue setup failed, will fallback to MMIO for display HW programming
I have no idea how to read this core dump so any assistance would be great.
Were you able to figure out how to leave a set amount of memory behind for the host?
Not reliably yet. It seems to work if you somehow allocate some before creating the VFs, but there is no kernel option (yet?) to control this explicitly.
I haven’t tried yet but perhaps loading a smallish LLM onto the card first could work?
So I tried ollama but keep in mind you have to build ollama from source and point it to the vulkansdk to make it build with intel gpu support. This did work. I have 4G of vram free for the host machine after turning off ollama. I used the codellama model to take up the vram then create the VFs. I’ll keep testing today and see if I notice less issues with the host machine having some vram to play with.
1 Like
Anyone with a B50 and noticing coil whine when idle? Just stood up a proxmox 6.17 kernel instance and with the card sitting idle, there’s an intermittent coil whine which is driving me nuts. I’m guessing if I put some load on the card it might go away but has anyone observed this behaviour? Sensors output -
xe-pci-0400
Adapter: PCI adapter
fan1: 1049 RPM
pkg: +49.0°C
vram: +48.0°C
card: N/A (crit = 0.00 W, cap = 70.00 W)
card: 9.01 kJ
pkg: 29.30 kJ
I haven’t noticed anything… coil whine can vary from ample to sample I believe (even if some products are more or less prone to it). Or depend on your home wiring and/or PSU etc.
Well, it seems go away when I pass the card to VM which is a positive. One thing I noticed is that fan control disappears when passed through in Proxmox. I hooked up the card to a bare metal and the intel software enables me to control the fan speed - just not in a VM. Also, gpu-z says rebar enabled but intel software says no.
Mine also has some WICKED coil whine at idle. honestly… I hadn’t noticed if it was only at idle but I think it was… I’ll have to look into that.
2 Likes
in Wendel’s video he said he had wrote as script to setup proxmox properly.
does anybody have a copy of it so that I can adapt it ?
I left my card running for a few hours to the point i had to take it out, im so sensitive to coil whine! Could have thrown the card in the bin tbh.