July 22, 2026
Bridge too far?
When Is NVLink Worth It?
Turns out the pricey bridge is mostly a flex unless your setup is a total mess
TLDR: Testing found the costly NVLink bridge only gives a big win in a narrow set of AI tasks, and a simpler direct connection can often do nearly as well. In the comments, readers mostly treated this as confirmation that the bridge is overpriced, while others used it as one more reason to mourn how much worse GPU upgrades have gotten.
The big reveal in this NVLink test is almost deliciously petty: that very expensive little connector people have been eyeing for old Nvidia cards may not be the miracle upgrade many hoped for. The tester ran two RTX 3090 graphics cards through a pile of AI jobs and found that for most everyday use, the fancy bridge barely changed anything. It helped a lot in one heavy training case, but once the setup was fixed to allow direct card-to-card communication without the bridge, the magical gap mostly vanished. Translation for normal humans: the expensive add-on often isn’t worth it.
And the comments? Instant reality check. One reader basically declared the case closed, saying this confirms the bridge only saves you if your computer’s layout is "nonsensical PCI-E topology"—which is the nerd version of calling someone’s wiring cursed. Another commenter came in with pure retro sadness, mourning the old days when you could buy the same graphics card years later for cheap and get a huge gaming boost. It’s less a benchmark discussion, more a support group for people grieving dead upgrade paths.
The juiciest undercurrent is the community’s ongoing side-eye toward Nvidia itself. The article points out that some limits look artificial, and that sparked the familiar mood: not outrage exactly, but a tired, knowing "of course they did that." Even the shortest reply—just "Thanks"—somehow reads like a calm survivor walking away from a $400 purchase decision.
Key Points
- •The article benchmarks NVLink on a dual-RTX 3090 system for AI inference and training workloads in April 2026.
- •Initial results showed no change for layer-split inference or DDP training, about a 30% boost for tensor-parallel prompt processing, and nearly 3x better FSDP training with NVLink.
- •The author discovered the original non-NVLink tests were falling back to shared host memory because NCCL_P2P_DISABLE=1 disables PCIe peer-to-peer as well as NVLink.
- •After physically unplugging the bridge and patching drivers to restore PCIe P2P on consumer GPUs, the author confirmed Nvidia driver-level restrictions were blocking P2P by default.
- •With actual PCIe P2P enabled, most of the difference between SHM and NVLink was nearly eliminated, and several workloads remained effectively unchanged within test noise.