Huawei launches Ascend 960 SuperNode: world's first NPO supernode with 4,096-card support

The most memorable part of this launch is a set of comparison numbers: 5,500 optical engines replacing 48,000 800G optical modules. But the question Huawei really wants to answer is more plain—the cards are already fast enough, yet the whole system cannot keep up. Data hasn't arrived, so you wait; other nodes haven't finished synchronizing, so you wait again. The more expensive the cards, the harder that waiting is to bear.

One quantified backdrop Huawei provides is that under traditional server architectures, intra-cluster communication can account for over 40% of training time. It must be emphasized that this ratio cannot be applied to every training task; it shows that communication overhead does constitute a bottleneck when scaling up clusters, rather than slapping the same label on all workloads.

The Ascend 960 SuperNode was officially launched on September 17, supports 4096 cards, and mainly targets training and inference for large models at the hundred-trillion-parameter scale. According to Huawei's definition, a SuperNode tightly connects multiple physical compute nodes through efficient interconnection, supports unified memory addressing across physical nodes, and logically possesses the characteristics of "one computer"—the devices are still those devices, but the way software sees and uses resources has changed.

A clear head is needed here: unified memory addressing does not erase physical distance, nor does it make all accesses equally fast. A logical "one computer" and physical equivalence are two different things.

The most noteworthy technical change in the Ascend 960 SuperNode is NPO (Near-Packaged Optics). Huawei claims it is the world's first SuperNode to adopt NPO technology—this "world's first" claim comes solely from Huawei and lacks third-party verification, so it should be treated as a vendor claim rather than an established fact.

What NPO does is very specific: it places the optical engine closer to the main chip package, shortening the path that high-speed electrical signals must travel. High-speed electrical signals do not behave the same no matter how far they run; the longer the path, the harder it is to handle loss, signal integrity, and power consumption issues. If transmission rates continue to rise, engineering compromises will only multiply. NPO's approach is to hand part of the distance over to light.

But turning it into a product is another matter. With the optical engine close to the main chip, layout and thermal design become more constrained; performance must improve, and reliability cannot be sacrificed. When introducing Hi-ONE, Huawei specifically mentioned the balanced design of optical, mechanical, electrical, magnetic, and thermal factors, which precisely shows that these constraints must be handled simultaneously rather than defeated one by one.

A single Hi-ONE engine achieves a transmission capacity of 7.2T, uses a built-in light source, and together with Lingqu forms the interconnection foundation of the Ascend 960 SuperNode. The Ascend 960 SuperNode adopts an orthogonal architecture and a full liquid-cooling design, providing a maximum of 8 EFLOPS FP8 compute power. At this scale, chips, interconnection, and cooling can hardly be handled separately.