Rumor
Nvidia Rubin Ultra reportedly features both 2-die and 2+2 die configurations utilizing HBM4E
Monday, March 30, 2026 at 10:22 PM
Nvidia is reportedly developing two versions of the Rubin Ultra architecture: a standard 2-die version and a 2+2 die configuration. TSMC will manufacture the 2-die units, which can be assembled into 2+2 configurations on a single board. The Rubin Ultra 2-die setup will feature 512GB of HBM4E (8 cubes of 64GB), while the 2+2 configuration will reach 1,024GB of memory capacity. This design choice maintains high memory density and rack counts but may impact packaging and testing requirements.
Context
Reports from GF Securities suggest Nvidia is utilizing a dual-track die strategy for its next-generation Rubin Ultra GPU to manage manufacturing complexity. While the market initially expected a 4-die native architecture, TSMC is reportedly producing 2-die units that can be deployed as a standard 512GB HBM4E configuration or assembled into a 2+2 die configuration on a CoWoS board to reach 1,024GB (1TB) of memory. This modular approach allows a single Kyber compute blade to house either four 2-die units or two 2+2 die units, maintaining the performance targets showcased at GTC 2025.
Analysts believe the overall value impact is limited as memory capacity remains in line with original expectations, though there may be minor negative pressure on specialized packaging and testing margins. The Rubin Ultra platform, scheduled for a 2027 launch, aims to deliver massive density gains, scaling up to 144 GPU packages per rack. This flexibility in die configuration is likely a strategic move by Nvidia and TSMC to ensure yield stability and supply chain readiness given the extreme manufacturing demands of the Kyber architecture.
Sources (7)
Samsung Unveils HBM4E, Showcasing Comprehensive AI Solutions, NVIDIA Partnership and Vision at NVIDIA GTC 2026 | Samsung Semiconductor GlobalBest of NVIDIA GTC: AI Breakthroughs and Recommended SessionsExclusive: TSMC reshapes advanced packaging expansion, shifting AP8, AP7, and US fab plansCo Packaged Optics (CPO) – Scaling with Light for the Next Wave of InterconnectNvidia demonstrates Rubin Ultra tray, the world's first AI GPU with 1TB of HBM4E memory — new chips will slot into Kyber racks | Tom's HardwareGfhk - Nvda Preview | PDF | Equity (Finance)NVIDIA Rubin R100: Specs, Architecture, and GPU Cloud Availability | Spheron Blog
Related Companies
TSMC
2330
Nvidia
NVDA