Huawei’s NPO Optical Engine Packs 5,500 Chips Into a 4,096-Card Supernode

Huawei’s self-developed optical engine Hi-ONE has surfaced at its first landing point, a kilo-card-scale supernode. At a recent optoelectronics expo, HiSilicon showed the engine’s bandwidth of up to 7.2 terabits and said Hi-ONE would enter the next-generation supernode in NPO (near-packaged optics) form. Huawei calls it the world’s first AI compute supernode to use NPO.

Huawei's NPO Optical Engine Packs 5,500 Chips Into a 4,096-C
Huawei’s NPO Optical Engine Packs 5,500 Chips Into a 4,096-C (Source: LeiPhone)

Why NPO. The answer sits in the electrical signal path. As on-chip high-speed interface rates climb, the longer the electrical link from chip to optical engine, the more the signal attenuates and distorts. Unlike traditional pluggable modules that let high-speed signals travel far across a PCB, and unlike CPO (co-packaged optics) that binds the engine tightly to the main chip, Huawei’s NPO seeks balance between bandwidth, latency and maintainability, compressing the high-speed electrical link to about five centimetres, the length of two one-yuan coins placed end to end.

A few days later at Huawei Connect, Huawei detailed Hi-ONE’s deployment: replacing 48,000 800G pluggable modules, 5,500 Hi-ONE engines enter a 4,096-card Ascend 960 supernode; the cross-cabinet Lingqu interconnect supports 176 ports of 1.6 terabits, with about 280 terabits of all-optical bandwidth in a single machine.

Mystery one: how is the shared light source shared?

Hi-ONE is not new this year. As early as the 2025 IFOC optoelectronics conference, HiSilicon disclosed that the engine uses a built-in shared integrated light source and silicon-photonics chip. By the 7.2-terabit capacity, one Hi-ONE equals nine 800G pluggable modules, sharply cutting the light sources and optical components needed. Multiple transmit paths can share the light source via WDM (wavelength division multiplexing), reducing fibre count.

But the public information does not explain how many emitters or wavelengths serve the 36 transmit channels. More importantly, building a light source into silicon photonics is extremely rare in the industry. Most schemes place the laser outside the engine to keep it away from server heat and allow individual replacement. An industry figure notes the built-in source is not a common NPO choice, yet sees it as logical given Huawei’s push to maximise integration.

Mystery two: how is 7.2T split and used?

Hi-ONE carries 36 optical transceiver channels, each up to 200 gigabits, for 7.2 terabits total. The industry’s common 6.4T NPO uses 32 by 200G; Hi-ONE chose 36 by 200G. One analyst suspects a redundant design: 32 channels for normal traffic, 4 on standby. Yet 5,500 engines at 7.2T total about 39.6 petabits per second, slightly above the 38.4 petabits of 48,000 800G modules; at 6.4T each the total is only 35.2 petabits, insufficient. Huawei’s wording therefore implies all 36 channels carry traffic.

NPO need not unify

Huawei did not build a wall everywhere. While keeping full-stack self-development at the core, it launched the OPEN NPO project with China Mobile Research, JD Cloud, Baidu and leading component and connector makers to build a common industry standard. That forces Huawei to redraw a line: which designs stay in its grip, which interfaces must sit on the table for all to follow. This confidence comes from nearly 30 years in optical networking.

The industry’s NPO standard activity is intensifying. Beyond Huawei’s OPEN NPO, an Open CPX MSA backed by Coherent, Marvell and Molex included 6.4T, 7.2T and various light-source schemes in its 1.0 spec in September. The OIF also launched a formal next-generation NPO project led by China’s CAICT with Huawei, Tencent, Alibaba Cloud, Meituan and China Mobile. But an ecosystem needing common rules does not mean NPO must become one unified standard. For the relatively closed Scale-Up scenario inside a supernode, multiple interface specs can coexist without breaking separate systems.

Huawei's NPO Optical Engine Packs 5,500 Chips Into a 4,096-C
Huawei’s NPO Optical Engine Packs 5,500 Chips Into a 4,096-C (Source: LeiPhone)
Huawei's NPO Optical Engine Packs 5,500 Chips Into a 4,096-C
Huawei’s NPO Optical Engine Packs 5,500 Chips Into a 4,096-C (Source: LeiPhone)

Editor’s note: This is an adapted translation of the original LeiPhone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/category/chips/kgknicdWxIIoulL9.html.

Leave a comment