Nvidia cuts Vera Rubin NVLink bandwidth to 3.0 TB/s per GPU, SemiAnalysis reports
SemiAnalysis reports that Nvidia lowered Vera Rubin's NVLink bandwidth from 3.6 to 3.0 TB/s per GPU because of signal-integrity limits.
Nvidia has reduced Vera Rubin's NVLink bandwidth from 3.6 to 3.0 TB/s per GPU, according to SemiAnalysis. The report attributes the cut to limits in signal integrity in the SerDes circuits, a drop of 0.6 TB/s per GPU, or about a sixth of the original figure.
SerDes, short for serializer/deserializer, is the circuitry that converts the parallel data inside a chip into a fast serial stream sent over a wire or link, and back again. Signal integrity is whether that high-speed signal stays clean enough across the connection to be read correctly. According to SemiAnalysis, those limits are why the lower figure applies. The report gives no further detail on how the change affects customers or workloads.
What changed
Nvidia reduced Vera Rubin's NVLink bandwidth from 3.6 TB/s to 3.0 TB/s per GPU because of signal-integrity limits in the SerDes circuits.
Type: Breaking change · Applies to: Vera Rubin, NVLink
Why it matters
Lower per-GPU NVLink bandwidth could reduce scaling efficiency for multi-GPU training and inference on Vera Rubin systems.
What you can do
Ask hardware suppliers how the NVLink change affects the Vera Rubin systems you are planning to buy.
Original source
We have no link to the original source for this story yet.
Where we saw it
Latent Space
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works