hardware · 2026-11-19 · Tier 1

AscendGYM and TeleChat Xing4: the Ascend stack prices its own kernels

AscendGYM (2611.17621) and TeleChat Xing4 (2611.22115)

TL;DR. The two most-upvoted papers on the 11-18 Hugging Face list are both about running frontier work on Huawei Ascend NPUs instead of NVIDIA GPUs. AscendGYM (SLAI) is a kernel-optimization benchmark, not a kernel-generation one: 48 existing Ascend kernels across three difficulty levels, each tested on dtypes and tensor shapes collected from real training and inference jobs, so a speedup maps directly to a workload. It ships a full execution environment, a built-in harness with search orchestration and experience accumulation, and a new metric, the Speedup-Token Score (STS), which divides speedup by the token cost of getting it. Xing4 (China Telecom's TeleChat lab, 154 upvotes, top of the day) is a 285B agentic MoE pretrained on 8,192 Ascend 910C cards with MindSpore HyperParallel, with no CUDA anywhere, plus an open 29B-A4B model with 256K to 1M context (per AI Weekly Espresso).

A kernel win only counts if you can afford the search
AscendGYM scores speedup per token spent, on shapes from real jobs.
flowchart LR
  W["Real workloads<br/><small>shapes and dtypes</small>"] --> K["48 kernels<br/><small>3 difficulty levels</small>"]
  K --> H["Agent or harness<br/><small>search + experience</small>"]
  H --> X["Ascend runtime<br/><small>real invocation path</small>"]
  X --> C{"Correct and faster?<br/><small>tests on real shapes</small>"}
  C -->|yes| S["STS score<br/><small>speedup per token</small>"]
  C -->|no| H
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class W,K input
  class H,X core
  class C loop
  class S exit
  linkStyle 4 stroke:#2f9e44,stroke-width:2px
  linkStyle 5 stroke:#f08c00,stroke-width:2px
Blue is the benchmark input, purple the agent and hardware, amber the check-and-retry loop, green the cost-normalized score.

Xing4 systems details (from the abstract)

  • Architecture study of sparse activation, attention design, depth/width and speculative decoding, framed as capacity versus compute, memory access and execution overhead on Ascend.
  • Pretraining: topology-aware hybrid sharding, distributed optimizer execution, long-sequence and activation-memory management; CPU affinity and core isolation to cut host scheduling interference.
  • Post-training: TeleGym, a unified agent platform with disaggregated asynchronous execution, bounded oversampling and partial rollout for tasks of uneven length.

How it relates to prior wiki pages

  • Same lab lineage as SLAI T-Rex, which post-trained DeepSeek-V4 end-to-end on an Ascend SuperPOD. SLAI now ships the benchmark for the kernel layer underneath.
  • STS is the Ascend version of the cost-normalized kernel metric the wiki has been tracking on NVIDIA: CompoundingKernel (11-18) found that trustworthy timing runs per hour, not model quality, limited a 1,000-agent kernel search. AscendGYM makes the token bill of that search part of the score.
  • Policy context: the same day, Tencent's president said the new US remote-access chip rule "sets everyone back in China" and BIS was reported to have rushed H200 export approvals. Xing4 is the counter-signal: a 285B frontier-class training run that needed neither.
  • Concept pages: gpu-kernels, compute-economics.

Gaps

AscendGYM's initial suite is 48 kernels; no leaderboard numbers are in the abstract. Xing4's abstract gives no benchmark scores or MFU, so "frontier" is a claim about scale, not quality.

Raw sources: AscendGYM · Xing4 · arXiv AscendGYM · arXiv Xing4 · GitHub