China bypasses US GPU bans with 1.54-exaflops 'LineShine' supercomputer — CPU-only monster packs 2.4 million Huawei-designed Armv9 cores
China has developed the LineShine supercomputer, a CPU-only system achieving 1.54 exaflops, to circumvent US restrictions on GPU exports. The system relies on 2.4 million Huawei-designed Armv9 cores across 40,960 custom LX2 processors optimized for AI and high-performance computing. This marks a strategic shift toward domestically developed, CPU-centric architectures for advanced computational workloads.
- ▪The LineShine supercomputer delivers 1.54 exaflops using 20,480 nodes, each with two LX2 processors totaling 40,960 CPUs.
- ▪Each LX2 processor contains 304 Armv9 cores with SVE and SME units to accelerate AI and HPC workloads using FP64, BF16, FP16, and INT8 formats.
- ▪The LX2 CPU integrates 32 GB of on-package HBM with 4 TB/s bandwidth and up to 256 GB of DDR5 memory, a design similar to Fujitsu's A64FX.
- ▪A dedicated SDMA engine manages data movement between HBM and DDR memory to optimize performance for AI training workloads.
- ▪The system's architecture required co-design of kernels, scheduling, and memory management to sustain high utilization of matrix processing units.
Opening excerpt (first ~120 words) tap to expand
Tech Industry Artificial Intelligence China bypasses US GPU bans with 1.54-exaflops 'LineShine' supercomputer — CPU-only monster packs 2.4 million Huawei-designed Armv9 cores News By Anton Shilov published 17 May 2026 CPUs can successfully do GPU jobs, but there is one important caveat. When you purchase through links on our site, we may earn an affiliate commission. Here’s how it works. (Image credit: Google) Copy link Facebook X Whatsapp Reddit Pinterest Flipboard Email Share this article 0 Join the conversation Follow us Add us as a preferred source on Google Newsletter Subscribe to our newsletter The vast majority of leading supercomputers and AI clusters today use CPUs for general-purpose tasks and orchestration and AI GPUs for massive parallel computing workloads to achieve…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Tom's Hardware.
Discussion
0 commentsMore from Tom's Hardware
