WORLDTECHNVIDIA Lowers Rubin GPU Memory and NVLink Bandwidth in Revised Specs
NVIDIA has quietly revised the shipping specifications of its Rubin GPU as production hardware begins to arrive. According to ServeTheHome, the chipmaker now quotes less memory bandwidth and less scale-up NVLink bandwidth than it had promoted, while scale-out networking bandwidth has gone up.
What changed in the specifications
NVIDIA, a semiconductor company based in Santa Clara, posted updated specifications for its Rubin GPU and several Rubin-based products in September, according to ServeTheHome, which first spotted the change through IT Home. The figures appeared on NVIDIA's product webpages and in its architecture datasheet.
Memory bandwidth is the largest reduction. NVIDIA had initially targeted 22TB/second per Rubin GPU, which implied each of the eight HBM4 memory stacks running at about 10.75Gbps per pin. The revised shipping specification gives a single GPU 19.2TB/second, or about 9.375Gbps per pin, a reduction of 1.375Gbps or roughly 11%.
NVLink, the multi-GPU connection that lets many physical GPUs act as one, has also been scaled back. The preliminary figure of 3.6TB/second per Rubin GPU is now a flat 3.0TB/second, a cut of 600GB/second, or 17%. Scale-out networking moves the other way, rising from a preliminary 400GB/second to 450GB/second per GPU.
Context of the HBM4 targets
The change comes as NVIDIA had been projecting aggressive data rates for HBM4, the high-bandwidth memory stacked next to the processor. According to the report, the specification already doubled the memory bus width per stack from 1024 bits to 2048 bits. The official HBM4 specification had also called for a slightly lower clock rate than HBM3E, targeting 8Gbps per pin.
Against that baseline, NVIDIA's original 10.75Gbps figure was well above the standard, and even the reduced 9.375Gbps remains higher than the rate targeted by some other HBM4 customers.
Unexplained reduction and possible causes
NVIDIA has not said why it lowered the memory bandwidth figures. ServeTheHome notes that the immediate assumption is that NVIDIA and its memory partners cannot yet deliver the original specifications, whether because of memory controller speeds, stack speeds, stack yields or a combination of these. The report stresses that this remains an assumption.
Whatever the cause, customers will have less GPU memory bandwidth available, which the report says carries performance consequences.
Effect on the Vera Rubin NVL72 rack
The revisions also flow into the products built on Rubin. The flagship Vera Rubin NVL72 rack-scale system combines Vera CPUs, BlueField-4 DPUs and Rubin GPUs. According to the report, cumulative GPU memory bandwidth for the rack falls from 1,580TB/second to 1,400TB/second, a loss of 180TB/second or 11%.
The theoretical compute throughput of the Rubin GPU is unchanged at 50 PFLOPS of FP4 inference and 35 PFLOPS of FP4 training, and the report says NVIDIA has met its clock speed and memory-capacity targets.
Open questions
- Why did NVIDIA reduce the memory bandwidth specification for Rubin GPUs?
- Will the lower NVLink and memory bandwidth affect real-world performance of Vera Rubin NVL72 systems?