Sunrise showcases S3 inference GPU and 256-card system to cut token costs
The AMW Read
S3's inference-focused architecture and LPDDR6 memory meaningfully update the infrastructure player map, but the claimed cost reduction lacks disclosed validation and its broader impact remains unproven.
Sunrise showcases S3 inference GPU and 256-card system to cut token costs
Sunrise presented its Qiwang S3 inference GPU and a 256-card supernode during CCTV Finance's Dialogue program on October 17. The Hangzhou company said S3 delivers roughly five times the inference performance of its previous S2 generation while reducing inference cost per token by about 90%, according to its own calculations. The article provides no benchmark methodology or independent validation of those estimates. Sunrise said it adopted an all-in inference strategy at the end of 2024 after finding that most customer workloads on S2 involved inference.
The hardware choices put Sunrise in the inference infrastructure race, where memory economics and system utilization matter alongside processor performance. S3 removes training-related modules, redirects transistor and power budgets toward inference, and uses LPDDR6 instead of conventional HBM. It also supports FP4 precision, high-speed interconnects and tiered storage. Sunrise links that design to agents' repeated planning, retrieval and tool calls, which increase token consumption. Its cost argument combines hardware amortization, operating expenses and system maintenance divided by token output, making the entire serving system relevant to the economics.
For builders and investors, the concrete diligence question is whether the reported savings hold for the same model, output quality and latency requirements under a representative workload. The 256-card supernode is described as supporting trillion-parameter and larger multimodal mixture-of-experts models, but the article supplies no measured serving results for that configuration. S3's memory and interconnect design offers a specific proposition to evaluate; the company estimates alone do not establish production economics.