Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@ for the original description of the approach.
| RTX 3090 | Pairwise NVLink where available, PCIe BAR1 otherwise |
| RTX 4090 | PCIe BAR1 |
| RTX 5090 | PCIe BAR1 |
| RTX 5060 Ti / 5060 (GB206) | PCIe BAR1, including with a display attached |

P2P also works between different devices of the same generation, for example RTX 5090
to RTX PRO 6000 Blackwell.
Expand All @@ -23,6 +24,10 @@ This enables BAR1 P2P on consumer GPUs where NVLink isn't available, and falls b
NVLink where it is. For PCIe pairs, transfers write directly to the other GPU's physical
address over DMA.

On GB206 cards (RTX 5060 Ti / 5060) BAR1 P2P works even when a display is attached: the
static BAR1 window is placed above the console reservation instead of requiring all of
BAR1, and allocations that don't fit in the window fall back to dynamic mappings.

> [!WARNING]
> IOMMU must be in passthrough mode (`iommu=pt`), not translating, or DMA will go through
> IOMMU page tables and transfers will fail. This is very dangerous if you run untrusted
Expand Down
Loading