At the 2026 RISC-V China Summit, Alibaba DAMO Academy reported four different kinds of XuanTie C950 performance results: multicore linearity above 75%, an advantage above 10% across four agentic workloads, inference rates of 43 and 38 tokens per second for two named models, and a 30% server-Java improvement. Each figure describes a different workload or measure, not one all-purpose performance score.
NeoTeo’s earlier coverage of XuanTie shipments examined a separate milestone. C950 itself was announced in March 2026; ShanghaiEye’s English-language report from that month covered its RISC-V server and AI-agent positioning.
Multicore linearity measures scaling across cores
DAMO reported C950 multicore linearity above 75% using its XT-LINK interconnect. The comparison is with the sum of the cores’ performance when each runs independently. It describes how well performance scales across C950’s cores, rather than a speed advantage over another processor.
DAMO’s March 2026 announcement also reported a 3.2 GHz clock and a single-core SPECint2006 score above 70 points, or more than 22 points per GHz. These are separate single-core figures, not another measure of XT-LINK’s multicore scaling.
The agentic result covers four workloads
DAMO reported more than a 10% performance advantage across four agentic workloads against processors it characterized as mainstream traditional architectures. Agentic workloads involve AI agents coordinating multistep tasks. The result applies to those four workloads; it does not describe a general performance lead across processors or tasks.
Inference rates belong to separate model runs
DAMO reported inference throughput of 43 tokens per second for Qwen3.8-Flash-Next and 38 tokens per second for DeepSeek V4-Flash. Each rate belongs to its named model, so the two figures do not rank the models against each other. DAMO attributed its inference approach to a TPE matrix extension, storage-architecture and special-function work, and software optimizations.
DAMO separately reported a 30% performance improvement after deep adaptation and optimization of C950 for server Java. That figure concerns Java server workloads, not model inference.
C950A has a separate automotive target
The fivefold performance target belongs to C950A, planned automotive processor IP, and is relative to existing RISC-V automotive products. Summit coverage reported an expected introduction by the end of 2027. That roadmap target is separate from the C950 results.
