Pins matching tag:"NVIDIA GPU"

NVIDIA unveils its Hopper architecture and first Hopper GPU, the H100: 80 billion transistors on TSMC 4N, the first GPU with HBM3 (3 TB/s) and PCIe Gen5, a Transformer Engine and fourth-generation NVLink. Availability is stated as later in 2022.

Tom's Hardware reports that Nvidia is selling the H100 to China as the H800, with the chip-to-chip data transfer rate cut. Nvidia published no specification sheet for the part.

NVIDIA unveils Blackwell: a 208-billion-transistor GPU on TSMC 4NP, the GB200 Grace Blackwell Superchip (two B200 GPUs and a Grace CPU) and the 72-GPU GB200 NVL72 rack. Partner products are stated for later in 2024.

NVIDIA announces Blackwell Ultra, the GB300 NVL72 rack (72 Blackwell Ultra GPUs, 36 Grace CPUs) and the HGX B300 NVL16 system, claiming 1.5x the AI performance of GB200 NVL72. Partner products are stated for the second half of 2025.

NVIDIA announces Rubin CPX, a GPU for million-token context processing, and the Vera Rubin NVL144 CPX rack: 8 exaflops of AI compute, which NVIDIA says is 7.5x a GB300 NVL72, with 100TB of fast memory and 1.7 PB/s of memory bandwidth. NVIDIA states availability at the end of 2026.

At CES on 5 January 2026 NVIDIA launched the Rubin platform: six new chips, the Vera Rubin NVL72 rack (72 Rubin GPUs, 36 Vera CPUs) and the HGX Rubin NVL8 board. NVIDIA claims up to 10x lower inference cost per token and 4x fewer GPUs to train MoE models than Blackwell, with partner products due in the second half of 2026.

At GTC on 16 March 2026 NVIDIA said seven chips, including the Groq 3 LPU, are in full production and described five Vera Rubin racks: NVL72, a 256-CPU Vera rack, Groq 3 LPX, BlueField-4 STX storage and Spectrum-6 SPX Ethernet. NVIDIA says the NVL72 delivers up to 10x inference throughput per watt at one-tenth the cost per token versus Blackwell.