Rumor

Potential next-gen AI GPU (VR200) specs leak suggests major power and memory upgrades to compete with MI450X.

Sunday, November 23, 2025 at 01:27 PM

Leaked specifications for the unannounced VR200 AI accelerator indicate a substantial TGP increase to 2200-2300W and a memory bandwidth upgrade to 20TB/s via higher-bin HBM to better compete with the MI450X. The leak also suggests the VR200 may employ advanced 448G BiDirectional SerDes technology for its interconnect.

Context

Recent rumors suggest Nvidia is aggressively upgrading its next-generation VR200 AI accelerator in direct response to competitive pressure from AMD's upcoming MI450X. The revisions indicate a fierce battle for AI hardware dominance heading into the 2H 2026 launch window for both products. This escalating "spec war" involves significant last-minute design changes, highlighting the high stakes of the next product cycle. To maintain a performance lead, Nvidia has reportedly increased the VR200's total graphics power (TGP) from 1800W to ~2200-2300W. To match the MI450X, memory bandwidth is also rumored to be boosted from 13 TB/s to ~20 TB/s using higher-grade HBM4 memory. These power-hungry adjustments underscore the immense compute demands of future AI models and will have significant implications for data center power and cooling infrastructure. A key technological shift is the rumored use of 448G bidirectional SerDes for interconnects. This would allow Nvidia to send and receive data simultaneously over a single copper cable, effectively halving the cable count compared to the current GB200 NVL72 backplane architecture. This innovation aims to improve rack-scale efficiency and density, a critical factor as GPU power consumption continues to soar.

Related Companies

Nvidia
Nvidia
NVDA
US
AMD
AMD
AMD
US