/AI22h ago

Arcee AI's Cody Blakeney says NVIDIA's Rubin NVL8 transfers adjacent GPU memory faster than an H100 reads local memory

The architecture offers a 48x jump in memory bandwidth.

--0--
Original posts
Cody Blakeney#999
Quote posts
wh#1430
Comments
wh#1430
Original post
Cody Blakeney@code_star#999inAI

This isn't private information, but considering the first Vera Rubin systems are rolling out and many people are still using Hoppers (at record prices) It's crazy to think how insane of an upgrade moving from Hopper -> Vera Rubin is.

48x memory bandwidth?! 31x NVlink bandwidth?! 3.5x more memory

At face value, this means it's actually faster to transfer and read memory from another GPU inside a node of Rubin NVL8 than it is to just read the memory inside an H100

Shit is going to be actually wild in the next couple of years. The cost of training monster models is about to go way down, and the cost of tokens is going to go way down as well.

9:36 AM · May 31, 2026 · 35K Views
Sentiment
Sentiment unavailable for this story.
Cluster Engagement
-
Views
-
Comments
-
Reposts
-
Bookmarks
Expand data
Posts from X
Most Activity
Most ActivityTimeline
VIEWS6.3KBOOKMARKS26LIKES60RETWEETS6REPLIES3
wh@nrehiew_

Comparison of the specs between Hopper, Blackwell and now Rubin.

Rubin NVLink bandwidth is now faster than H100's HBM bandwidth. This also doesn't include a 3-5x FP4 FLOPS increase between Blackwell and Rubin.

Cody Blakeney@code_star

This isn't private information, but considering the first Vera Rubin systems are rolling out and many people are still using Hoppers (at record prices) It's crazy to think how insane of an upgrade moving from Hopper -> Vera Rubin is.

48x memory bandwidth?! 31x NVlink bandwidth?! 3.5x more memory

At face value, this means it's actually faster to transfer and read memory from another GPU inside a node of Rubin NVL8 than it is to just read the memory inside an H100

Shit is going to be actually wild in the next couple of years. The cost of training monster models is about to go way down, and the cost of tokens is going to go way down as well.

2hViews 6.3KLikes 60Bookmarks 26