Iluvatar CoreX picked an unusual yardstick for its latest accelerator. The Shanghai company staged a media day on Oct 28 to show off the Tiangai 300 and a rack-scale supernode, then published benchmark numbers that measure the part against Nvidia’s H200, a chip Washington has spent years keeping out of Chinese data centres.
The numbers come from Shanghai Yidian, a third party the vendor commissioned. In a disaggregated inference setup that splits prompt processing from token generation, per-card throughput reached 1.05 times the H200. On a long-context test it reported 15 percent better throughput on a mixture-of-experts model and 10 to 30 percent lower first-token delays. An eight-card training demonstration using FP8 math landed at 1.33 times the H100.
The commercial pitch sits above the silicon. Iluvatar sells a 144-GPU rack and claims its configuration cuts the cost of one million useful tokens by about 40 percent. That framing treats throughput per dollar as the buying criterion rather than peak floating-point rates, which is a reasonable bet while memory and power constrain Chinese clusters.
What the benchmarks do not settle is delivery. The figures hold for named models at named context lengths, not for an average workload, and the components tested are pre-production samples. Questions about software maturity and how many racks a customer can actually take delivery of are not answered by a throughput ratio.
Iluvatar listed in Hong Kong this year, which gives it a currency for expansion but also a quarterly audience. Comparing itself to a chip it cannot buy is good marketing; whether cloud operators agree will show up in shipments.
