The competition between NVIDIA and AMD on AI chips is becoming increasingly fierce.To meet the challenges of AMD's Instinct MI455X AI chip, NVIDIA is making upgrades to its next-generation Vera Rubin platform, especially in terms of memory bandwidth.SemiAnalysis revealed that at CES 2026, NVIDIA revised its Vera Rubin NVL72 memory bandwidth specification to exceed AMD MI455X chip by 10%.
The memory bandwidth of AMD MI455X is 19.6TB/s. In March 2025, Huang Renxun announced that Rubin would be 13TB/s, and then in September of the same year it was announced that it would be 20.5TB/s. In January 2026, NVIDIA announced that its memory bandwidth would be 22.2TB/s.
NVIDIA's upgrade strategy reflects the current AI system's urgent need for higher memory bandwidth. As generative AI systems gradually dominate in 2026, memory bandwidth has become a key factor in determining chip performance.
This is regarded by the industry as a direct counterattack to AMD, which has achieved a lead of 19.6 TB/s with the 12-Hi HBM4 stacking technology adopted by the MI455X.
In order to achieve overtake, NVIDIA adopted a more radical plan:Based on the 8-layer stacked HBM4, suppliers are required to increase the pin speed to 11Gbps, far exceeding the JEDEC standard rating.
This strategy of increasing bandwidth through extreme overclocking is designed to ensure NVIDIA's absolute dominance among ultra-large-scale cloud service providers.
