H20卖不动？英伟达最新“特供版”芯片曝光

Core Viewpoint - NVIDIA is developing a China-specific AI chip, B30A, based on its latest Blackwell architecture, which is expected to outperform the currently available H20 model in China [1][4]. Group 1: Chip Specifications - The Blackwell architecture was introduced at NVIDIA's GTC conference last year as an upgrade to the Hopper architecture [3]. - The B30A chip will feature a single-die design, integrating all major circuits on a single silicon wafer, enhancing overall performance and stability [3]. - B30A's original computing performance is anticipated to be about half of the flagship Blackwell Ultra B300's dual-chip configuration [3]. - The Blackwell Ultra B300 is manufactured using TSMC's 4NP process, supports fifth-generation NVLink with a chip-to-chip interconnect bandwidth of 1.8TB/s, and has a typical power consumption of 1400W [3]. Group 2: Performance Comparison - The B30A is likely a "cut-down" version of the B300A, which is a general product for the global market, featuring 144GB HBM3E and approximately 4 PFLOPS of FP8 performance [4]. - The H20 model has 96GB HBM3, a bandwidth of 900GB/s, and an FP8 performance of 296 TFLOPS, which is significantly lower than both B300A and B300 [4]. - The B30A is expected to exceed the performance of H20 and may approach that of H100, which has an FP8 performance of 4 PFLOPS [4]. Group 3: Market Strategy - NVIDIA is also planning to launch the RTX 6000D, which will be priced lower than H20 and designed for AI inference tasks, featuring traditional GDDR memory with a bandwidth of approximately 1.398 TB/s [5]. - The RTX 6000D is expected to be delivered in small batches to customers as early as September, further expanding NVIDIA's product offerings in the Chinese market [5].