Huawei Told Nvidia NO — 7.2 Terabits Outside the Chip Changes Everything
Community Score: 50% | 2.2K views | 1w
0 community ratings: null thumbs up, null thumbs down
#huawei #nvidia #chipwar Huawei just told Nvidia “NO” — and the reason is hiding outside the chip. Huawei has unveiled its Hi-ONE near-packaged optics platform with a claimed 7.2 Tbps optical engine, using 36 lanes at 200 Gbps each. Instead of placing optical engines directly inside the chip package like co-packaged optics (CPO), Huawei is pushing its NPO approach — keeping the optics extremely close to the compute package while making them easier to replace and service. But the bigger story is what this means for massive AI systems. Huawei says its Atlas 960E SuperPoD can connect 4,096 NPUs, reach up to 16 EFLOPS at FP4, and use around 5,500 Hi-ONE optical engines. The architecture is designed to tackle one of AI's biggest hidden problems: moving enormous amounts of data between thousands of processors without turning the data center into a power and cooling nightmare. In this video, we break down: • Why copper is becoming a bottleneck for massive AI clusters • CPO vs NPO — and why H
Tags: Huawei, Nvidia, Huawei Hi-ONE, Huawei 7.2 Tbps, 7.2 terabits, Huawei optical engine, Huawei NPO, near packaged optics, NPO optics, co packaged optics, CPO, CPO vs NPO, optical interconnect, optical computing, silicon photonics, AI data centers, AI infrastructure, Huawei Atlas 960E, Atlas 960E SuperPoD, 4096 NPUs
More from Quantum Silk Route
- China Just Ran AI on Light. NVIDIA Can't Compete With This. — Score: 50%
- Why TSMC, Intel & Samsung Are Racing to Build This Sub-2nm Chip. — Score: 50%
- China Doesn't Need Nvidia Anymore — DeepSeek V4 Just Proved It — Score: 50%
- TSMC's Worst Nightmare: China's $1B Glass Secret Is Already Running. — Score: 50%
- China Banned This Metal. Intel Made It Irrelevant. — Score: 50%
- China Just Solved the Biggest Problem With 2D Chips — Score: 50%