HBM MASSIVE CONTENT DOWNGRADE ALERT🚨🚨: Nvidia Rubin Ultra will ship with 192GB of HBM4 8-hi. Not only is this a huge downgrade from the original Rubin Ultra which was previewed with 1TB of HBM, but it is even lower than regular Rubin which has 288GB. How did HBM content for Rubin Ultra get cut to 1/5th of the original? The number of cubes got cut to 8 when 4 die Rubin Ultra was scrapped. Then the stack height got reduce to 8-hi instead of 16-hi. Last it is cut to HBM4 which uses 24Gb dies instead of 32Gb dies with HBM4E, although an 8-hi HBM4E upgrade could com later. What was Nvidia's reasoning behind this? (1/2)🧵
This and more is contained in the SemiAnalysis Accelerator Model, where we first broke this news to subscribers in July. Subscribe to find out more on the implications, what Nvidia is doing to offset lower capacity per GPU, and our views on where HBM specs are heading in the future.
https://t.co/H1RSZwHlSA (2/2)
HBM per accelerator caps model and KV-cache size. Rubin Ultra landing at 192GB — below regular Rubin's 288GB and a fifth of the previewed 1TB — changes what a single GPU will hold.
