Note: The English version is provided below.
摘要
近期,阿里旗下平头哥半导体的 PPU芯片、华为昇腾910B和壁仞104P产品,以及国际厂商英伟达的A800/H20对比数据在央视节目中的亮相,标志着国产AI算力芯片在高端市场的实质性突破。
显存容量:
平头哥PPU和英伟达H20 均为 96G,远超英伟达A800的80G,能支持更大规模的 AI 模型(如千亿参数大模型)训练与推理。国产厂商华为昇腾910B和壁仞104P相对较弱。
显存类型:
平头哥PPU 的HBM2e与A800相当,弱于H20采用的HBM3(第四代高带宽显存),其单在超高速数据读写场景中更具优势。国产厂商壁仞104P也与A800相当,华为昇腾910B较弱。
片间带宽:
平头哥PPU的700GB/s远超A800的400GB/s,接近 H20 的900GB/s,这意味着多卡互联时数据交换更高效,适合构建大规模算力集群。国产厂商华为昇腾910B和壁仞104P相对较弱。
?
点评:
整体来看,国产AI芯片平头哥PPU和华为昇腾910B已在部分核心参数上接近或比肩国际厂商(NV),且在能效比、自主可控等维度形成差异化竞争力;而 NV 的产品则在极致性能上仍有领先场景。
国产AI芯片技术已经不在是备选,而是更优选。未来,国产厂商将在 AI、算力、数据要素等领域持续突破,将重塑全球产业格局。
Summary
The recent appearance of comparative data on Alibaba's T-Head Semiconductor's PPU chip, Huawei's Ascend 910B and BiRen 104P products, and international manufacturer Nvidia's A800/H20 on a CCTV program marks a substantial breakthrough for domestically produced AI computing chips in the high-end market.
Memory Capacity
Both the T-Head Semiconductor PPU and the NVIDIA H20 offer 96GB of RAM, far exceeding the NVIDIA A800's 80GB, enabling the training and inference of larger AI models (such as those with hundreds of billions of parameters). Domestic Huawei Ascend 910B and Biren 104P are relatively weaker.
Memory Type
The T-Head Semiconductor PPU's HBM2e is comparable to the A800's, but weaker than the HBM3 (fourth-generation high-bandwidth memory) used in the H20. This performance advantage is greater in ultra-high-speed data read and write scenarios. The domestic Biren 104P is also comparable to the A800, while the Huawei Ascend 910B is weaker.
Inter-chip bandwidth
The T-Head Semiconductor PPU's 700GB/s far exceeds the A800's 400GB/s and approaches the H20's 900GB/s. This means data exchange is more efficient when interconnecting multiple cards, making it suitable for building large-scale computing clusters. Domestic vendors Huawei's Ascend 910B and Biren 104P are relatively weaker.
Comment
Overall, domestic AI chips like the T-Head Semiconductor PPU and Huawei's Ascend 910B are approaching or on par with international vendors (NV) in some core parameters, and have differentiated themselves in areas such as energy efficiency and proprietary control. NV's products, however, continue to lead in extreme performance scenarios.
Domestic AI chip technology is no longer an alternative, but rather a preferred option. Going forward, domestic vendors will continue to achieve breakthroughs in AI, computing power, and data elements, reshaping the global industrial landscape.
来源: 与非网,作者: 王兵,原文链接: /article/1894685.html