Memory is too expensive to afford! Nvidia Rubin slashes configuration: CPU memory is directly halved

Kuai Technology reported on July 26 that due to high memory chip prices and supply shortages, Nvidia will significantly reduce the memory configuration of the Vera Rubin NVL72 rack-level AI system. Judging from the specific data, the capacity of the SOCAMM memory module in the rack was directly halved from 192GB to 96GB, the total memory supporting the Vera CPU was cut from 54-55TB to 28TB, and the CPU rack part was also switched to 96GB SOCAMM. The only thing that remains unchanged is the HBM4 memory on the GPU side, which remains at 20.7TB per rack. The direct reason for Nvidia's move is the tight supply in the LPDDR5X market. Affected by rising memory prices, the cost of a Rubin NVL72 rack may be as high as $9.1 million. The earlier estimate of $7.8 million is outdated, and the price of HBM4 may rise to $53 per GB by 2027. If no adjustments are made, the memory cost will account for 29% of the VR200's total material cost, which is approximately US$2.1 million, far exceeding the ideal level of 20%. After the configuration was cut in half, the effect was immediate: the cost of VR200's LPDDR5X dropped from the earlier estimated US$1.2 million to US$586,000. If it is further reduced to a quarter of the original, it can even be reduced to US$293,000. You know, NVIDIA has been the company that can best withstand this wave of price increases by relying on long-term storage agreements signed in advance. Now even it has to go under the knife, which shows how much pressure there is on the storage market. It should be noted that Vera Rubin NVL72 itself is one of the most powerful AI systems currently, and was just released in January this year. According to actual measurement data, when processing MoE large model loads, the Blackwell platform's throughput per megawatt of power consumption is 80,000 tokens/s, while the Vera Rubin NVL72 reaches 800,000 tokens/s, a 10-fold increase.