Core Viewpoint - The release of DeepSeek V3.1 and the mention of a new architecture and next-generation domestic chips have caused significant excitement in the AI industry, leading to a surge in stock prices for domestic chip companies like Cambricon, which saw an intraday increase of nearly 14% and became the top company on the STAR Market [4][22]. Group 1: UE8M0 FP8 Concept - The term "UE8M0 FP8" can be broken down into two parts, with "UE8M0" representing a scaling factor in the MXFP8 path, which is defined in the Open Compute Project's specification for 8-bit micro-scaling formats [7][8]. - MXFP8 is based on FP8, compressing conventional floating-point formats to 8 bits, allowing for a significant expansion of the dynamic range while maintaining an 8-bit width [8][15]. - The scaling factor in UE8M0 consists of 8 bits, which can be allocated to sign, exponent, and mantissa bits, with the "U" indicating unsigned [11][12]. Group 2: Benefits of UE8M0 FP8 - UE8M0 allows processors to restore data using simple operations, significantly reducing the complexity of floating-point multiplication and normalization, thus shortening critical clock paths [15][17]. - The dynamic range of UE8M0 spans from 2^(-127) to 2^(128), providing ample space for subsequent block scaling and reducing information loss while maintaining 8-bit tensor precision [15][17]. - The adoption of UE8M0 can lead to a 75% reduction in data traffic compared to traditional FP32 scaling, making it a crucial optimization direction for next-generation architectures [18][27]. Group 3: Domestic Chip Manufacturers - Several domestic chip manufacturers, including Cambricon, Hygon, and Moore Threads, are preparing to support FP8, with Cambricon's chips already being compatible with FP8 calculations [22][23]. - The market has reacted positively to the potential of these domestic chips, with the STAR 50 index rising by 3%, marking a three-and-a-half-year high for the chip industry [24][27]. - The collaboration between DeepSeek and domestic chip manufacturers represents a shift towards a more self-sufficient AI ecosystem in China, reducing reliance on foreign computing power [27][28].
DeepSeek V3.1 专为国产芯片设计的 UE8M0 FP8 到底是什么?