tilpner has quit [Remote host closed the connection]
tilpner has joined #spectrum
cole-h has quit [Quit: Goodbye]
cole-h has joined #spectrum
cole-h has quit [Quit: Goodbye]
cole-h has joined #spectrum
cole-h has quit [Ping timeout: 258 seconds]
cole-h has joined #spectrum
cole-h has quit [Quit: Goodbye]
<zgrep>
puck: What counts as modern-ish?
<qyliss>
Broadwell
<puck>
zgrep: hrmm, iommu and 4th gen and up?
<puck>
oh, they dropped 4th? hrm
<qyliss>
puck: anyway yeah that's also a possibility
<qyliss>
There are two GPU sharing technologies of interest, as far as I can tell
<puck>
qyliss: i wouldn't be surprised if google used virtio-gpu + gvt-g at some point
<qyliss>
puck: don't they already?
<qyliss>
crosvm supports it
<puck>
i don't quite know what backs virtio-gpu on cros
<puck>
they don't trust the nvidia gpu virtualisation, that's for sure
<qyliss>
One is dma-buf, which would allow VMs to write directly to GPU buffers, rather than copying to shared memory first and letting another VM do it
<qyliss>
And the other is virtio-gpu, which allows VMs to do render on the GPU
<puck>
well, like
<qyliss>
Using either of those would mean GPU drivers on host, though.
<puck>
the two are shared
<puck>
DMABUF is just the internal linux mechanism for giving gpu memory
<puck>
to a single devic
<puck>
process* or something
<qyliss>
You can use dma-bufs in virtio_wl without virtio-gpu I think
<qyliss>
But actually maybe I'm wrong about that
<qyliss>
No, looks like I'm right
<puck>
right, separate protocol, hrmm
<qyliss>
But you do need dma-bufs for virtio-gpu I think
<qyliss>
puck: wdym by they don't trust nvidia gpu virt -- do you have a source?
<puck>
right, virtwl supports virtio-gpu dmabufs
<qyliss>
I thought nvidia wouldn't let you do virt on consumer GPUs anyway
<puck>
qyliss: well, you can get GPUs on google cloud, and nvidia supports dividing up a GPU
<puck>
but google always passes through the entire GPU
<qyliss>
ah, cloud things, right
<qyliss>
I'm pretty sure virtio-gpu in crosvm is backed by a GPU, at least
<qyliss>
I don't know if they use GVTg
<qyliss>
puck: a maybe-interesting observation is that I noticed today intel_iommu=on breaks resuming from suspend on the pixelbook
<puck>
oof
<qyliss>
So I guess Google must not use the iommu? Or I imagine that would have been fixed.
<puck>
right, they might not right now
<qyliss>
(Or they're using it with igfx_off, which fixes it)