NVIDIA cuts the memory bandwidth of its Rubin GPU by 11%

NVIDIA put new specifications for its Rubin GPU, an AI chip, on its website in September. The memory bandwidth, which is the data speed of the memory, is at this time 19.2 terabytes each second. It was 22 in the first plan, which is about 11% less. The bandwidth between GPUs in one rack decreases 17%, and the bandwidth to other racks increases. NVIDIA gives no cause.

What changed

The Vera Rubin NVL72 rack has the same changes. Its memory bandwidth decreases 11% and its GPU bandwidth decreases 17%. Its bandwidth to other racks increases. The HGX Rubin NVL8 system has 13% less memory, because each GPU uses 7 of its 8 memory stacks.

The numbers

The compute numbers do not change. The memory capacity of the NVL72 does not change. The memory speed is more than the HBM4 standard of 8 gigabits each second for each pin. The AMD MI455X works at 7.6.

What is not known

NVIDIA gives no cause. The source tells that NVIDIA and its memory partners possibly cannot make the first speed, but this is not sure. NVIDIA tells that the first numbers were not last. No test shows the effect on AI work.

Insight: The specifications come before the measured speed

Google Cloud tells that its A5X cloud servers, which use NVL72 racks, open to all customers next month. It gives no cost and no speed numbers. The racks have less memory bandwidth than NVIDIA first showed. This looks like a purchase where the specifications come first and the measured speed comes after the purchase. The customer has control only with a test on its work.