Nvidia NVHBM Targets 30% Higher Bandwidth and 15% Lower Power Than HBM4e
Nvidia is positioning a custom high-bandwidth memory stack called NVHBM as a step beyond commodity HBM4e, with Tom's Hardware reporting targets of roughly thirty percent higher bandwidth and fifteen percent lower power. The design pairs a custom base die and PHY, and Nvidia plans to make those building blocks available to NVLink Fusion partners rather than keeping them solely in-house. For builders planning large-scale training or inference clusters, that combination matters because memory bandwidth and power per bit increasingly set the ceiling on throughput and facility economics. If the figures hold in production silicon, NVHBM could widen Nvidia's advantage in tightly coupled GPU-memory systems. The packet offers headline performance deltas and partner availability, but no independent benchmarks, workloads, or ship dates, so treat the claims as directional until validated.
Sources
- Nvidia custom 'NVHBM' promises 30% higher bandwidth, 15% lower power than commodity HBM4e — custom base die and PHY will be available to NVLink Fusion partners - Tom's HardwareTom's Hardware AI · primary
- Nvidia Has Earnings Pricing Power, BlackRock's Li SaysBloomberg Technology · secondary
- Nvidia has been in talks to acquire Hugging Face for more than $13 billion - Business Insiderr/LocalLLaMA Top · secondary
- Meta says its AG deal requires it to create an age assurance standard that is subject to independent testing, with a false positive rate no higher than 10% (Lauren Feiner/The Verge)Techmeme · secondary