NVIDIA has released the first performance preview of its next-generation Blackwell B100 GPU, which will be more than twice as powerful as the Hopper H200 in 2024. In a special SC23 speech, NVIDIA previewed the performance of its next-generation GPU, codenamed Blackwell, which will debut in 2024 and will have more than 2 times the AI performance of the Hopper GPU. The GPU used is the next generation B100, which will succeed the Hopper H200. It can be seen that it has just crushed the GPT-3175B inference performance benchmark, demonstrating its huge AI performance potential.
In the past two years, Nvidia has relied on its Hopper H100 and Ampere A100 GPUs to work with different partners to serve the needs of global AI and HPC customers, but all this is about to change with the arrival of Blackwell in 2024. Nvidia's data centers and overall company revenue have been boosted by the artificial intelligence boom, and it looks like the train is accelerating, with the goal of launching two new GPU series by 2025.
NVIDIA's first new AI/HP CGPU series will be Blackwell, named after David Harold Blackwell (1919-2010). This GPU will be the successor to the GH200Hopper series and will use the B100 chip. The company plans to launch a variety of products, including GB200NVL (NVLINK), standard GB200 and B40 for visual computing acceleration. The next-generation product lineup is expected to be unveiled at the next GTC (2024) and released later in 2024.
Current rumors estimate that Nvidia will use TSMC's 3nm process node to produce Blackwell GPUs, and the first customers will receive this chip by the end of 2024 (fourth quarter). The GPU is also expected to be the first HPC/AI accelerator designed with a chipset and will compete with AMD's InstinctMI300 accelerator.
Another chip that was disclosed is the GX200, which is a follow-up to Blackwell and is scheduled to be launched in 2025. Nvidia's AI and HPC products have always been on a two-year cycle, so we will likely only see the release of the chip in 2025, with actual shipments not until 2026.
The product line will be based on the X100 GPU and include the GX200 product line and a separate X40 product line for enterprise customers. NVIDIA is known for naming its GPUs after famous scientists, and it already uses the Xavier codename for its Jetson series, so we can expect the X100 series to have a different scientist name. Beyond that, we know very little about the X100 GPU.
Nvidia also plans major "doubling" upgrades to Quantum and Spectrum-X with new Bluefield and Spectrum products, providing 800Gb/s transmission speeds by 2024 and 1,600Gb/s transmission speeds by 2025. These new network and interconnect interfaces will also greatly help the high-performance computing/artificial intelligence field achieve the required performance.