Huawei has showcased its Atlas 950 SuperPoD servers, a system comprising 1,024 NPUs designed for AI workloads. The system features extensive inter-rack connectivity, with each of the 32 LRS units utilizing 4x100G connections to neighboring racks, employing Huawei's StarMatrix pluggables. The Atlas 950 boasts significant memory capacity, including 98TB of HiZQ memory (a HBM3E equivalent) and an additional 158TB of DRAM, totaling 256TB. The system is available in versions optimized for decode and training (Ascend 950DT) or prefill tasks (Ascend 950PR). AI
IMPACT This large-scale NPU system from Huawei could accelerate AI training and inference capabilities, particularly for organizations requiring massive compute power.
RANK_REASON The cluster reports on the public showcase of Huawei's Atlas 950 SuperPoD, a significant AI infrastructure product.
- 800G LPO
- NPU
- Ascend 950
- Ascend 950DT
- Ascend 950PR
- Atlas 950 SuperPoD
- DRAM
- HBM3E
- HiZQ memory
- Huawei
- Larsen syndrome
- SemiAnalysis
- StarMatrix
- SuperPoD
AI-generated summary · Google Gemini · from 4 sources. How we write summaries →