Nvidia Reportedly Lowers Rubin HBM4 Pin Speed From 11.7 Gbps to 9.3 Gbps and Trims NVLink Per GPU Bandwidth
Holy smoke NVIDIA has officially lowered the HBM4 pin speed. 11.7-> 9.3Gbps
Context
Dialing back HBM4 memory pin speeds and scale up interconnect bandwidth on Nvidia's upcoming Rubin architecture eases yield bottlenecks for memory suppliers while narrowing Nvidia's raw bandwidth lead over AMD's MI450.
Over recent months Nvidia had pushed memory makers SK hynix, Samsung, and Micron to hit aggressive 11.7 Gbps pin speeds on next generation HBM4 stacks for its Vera Rubin platform, above the JEDEC standard baseline, raising supply and thermal concerns across the memory supply chain.
Deep dive
Supply chain analyst Jukan reported on Sunday night that Nvidia has lowered the target HBM4 memory pin speed on its upcoming Rubin GPU platform from 11.7 Gbps down to 9.3 Gbps, sharing an updated hardware specification table. Hardware researcher Zephyr noted from the same specification update that Nvidia also reduced per GPU NVLink scale up bandwidth from 1.8 TB per second unidirectional, or 3.6 TB per second bidirectional, down to 1.5 TB per second unidirectional, or 3.0 TB per second bidirectional.
In a separate note citing South Korean brokerage Daishin Securities, Jukan reported that Nvidia also revised Rubin's stack configuration amid base die supply constraints and customer demand for capacity, dropping an earlier 8 high HBM4 option on standard Rubin in favor of 12 high stacks, while initial variants of the 2027 Rubin Ultra platform are undergoing tape out with 8 high HBM4 before moving to 12 high HBM4E.
According to Zephyr's comparison of the updated specification tables, the lower HBM4 pin rate and trimmed NVLink speed leave AMD's competing 2027 MI450 accelerator, which packs 12 HBM4 stacks per CoWoS package for 288GB of memory, with higher projected HBM memory bandwidth at 23.3 TB per second and higher scale up bandwidth at 1.8 TB per second on paper.
What to expect next
Relaxing Rubin's HBM4 pin speed target from 11.7 Gbps to 9.3 Gbps should improve manufacturing yields and qualification timelines for SK hynix and Samsung as they prepare volume shipments. Hardware buyers at hyperscalers and frontier AI labs will wait for official production silicon benchmarks to see how Rubin's real world training and inference throughput compares against AMD's 12 stack MI450.