不靠台积电，华为靠什么做到世界算力最强？

Core Viewpoint - Huawei is advancing its AI chip technology with a roadmap that includes the release of several new Ascend chips, aiming to enhance computing power and interconnect bandwidth significantly over the next few years [3][4][5]. Chip Development Roadmap - The Ascend 910C was launched in Q1 2025, followed by the Ascend 950PR and 950DT in Q1 and Q4 of 2026 respectively. The Ascend 960 and 970 are expected in Q4 2027 and Q4 2028 [3][4]. - The Ascend 910C features a computing power of 800 TFLOPS (FP16) and an interconnect bandwidth of 784 GB/s, while the Ascend 950PR and 950DT will achieve 1 PFLOPS (FP8) and 2 PFLOPS (FP4) respectively, with bandwidths of 2 TB/s [4][5]. Technical Specifications - The Ascend 950PR will have a memory capacity of 144 GB and a bandwidth of 4 TB/s, while the Ascend 950DT will have 128 GB and 1.6 TB/s [4]. - The Ascend 960 will double the computing power to 2 PFLOPS (FP8) and 4 PFLOPS (FP4), with an interconnect bandwidth of 2.2 TB/s and a memory capacity of 288 GB [5]. - The Ascend 970 will further enhance performance to 4 PFLOPS (FP8) and 8 PFLOPS (FP4), maintaining a memory capacity of 288 GB but increasing bandwidth to 14.4 TB/s [5]. Competitive Landscape - Compared to NVIDIA's Blackwell Ultra GB300, which has a computing power of 15 PFLOPS (FP4) and a bandwidth of 8 TB/s, Huawei acknowledges a gap in single-chip performance due to U.S. sanctions affecting production capabilities [8]. - Despite this, Huawei emphasizes its strengths in connectivity technology, which it believes will enable it to achieve ultra-node capabilities in AI infrastructure [8]. AI Infrastructure Development - Huawei has deployed over 300 CloudMatrix 384 ultra-nodes, serving more than 20 clients, and plans to launch the Atlas 950 SuperPoD with a computing scale of 8192 cards in Q4 2023 [9]. - The upcoming Atlas 960 SuperPoD, expected in Q4 2027, will have a computing scale of 15488 cards, while the TaiShan950 SuperPoD, set for Q1 2026, will support up to 16 nodes and 48 TB of memory [9]. Industry Collaboration - Huawei expresses a commitment to collaborate with the industry to build a robust foundation for AI computing needs both in China and globally [10].