TileLang

Headlines · 4

1. Oct 22
   SemiAnalysis examines Huawei-DeepSeek reliance and Ascend 950PR microbenchmarks

   SemiAnalysis published an analysis of the mutual reliance between Huawei and DeepSeek, in which Huawei provides compute hardware while DeepSeek contributes model workloads and kernel optimization, according to a Techmeme summary of the piece. The analysis includes a deep dive and microbenchmarks of Huawei's Ascend 950PR, an analysis labeled DSv4.1 Flash, and coverage of DeepSeek's support for Ascend, AICPU and Tilelang Ascend. The supplied summary does not include the benchmark results themselves and the DSv4.1 Flash reference is truncated and unclear.

   Techmeme · 🔥 0
2. Oct 21
   DeepSeek open-sources its internal DeepInfer inference framework

   DeepSeek announced on October 20 that it is fully open-sourcing DeepInfer, its internal inference framework, following earlier open releases of models, the TileLang compilation tool and inference operator libraries, according to a Geek Park report. DeepSeek says DeepInfer already powers high-throughput, low-latency large-scale serving for DeepSeek V4.1 Flash, and that inference code and future optimizations for new DeepSeek models will continue to be updated there. The report cites DeepSeek's own statements and gives no technical details, benchmarks or license terms for the release.

   极客公园 · 🔥 0
3. Oct 19
   DeepSeek helps shape Huawei AI chips; Ascend 950DT pretraining eyed for late 2026

   DeepSeek is working closely with Huawei to shape the design of future AI chips, according to The Information, which cited two people familiar with the matter. Huawei engineers have been stationed at DeepSeek's offices for an extended period and are folding the model maker's requests into design decisions, such as how models calculate and move data between processors, for chips due next year and beyond. Chinese reports citing the same reporting add that DeepSeek plans to begin using Huawei's Ascend 950DT for pre-training as soon as the end of 2026, with testing still underway, and that Huawei intends to allocate most of the capacity of its Ascend 960DT, due in the first quarter of 2027, to DeepSeek for both training and inference; neither company has confirmed the accounts.

   The Information · 🔥 0
4. Oct 7
   DeepSeek open-sources Ascend ports of DeepGEMM and DeepEP kernels

   DeepSeek has published two new repositories, DeepGEMM-Ascend for matrix multiplication and DeepEP-Ascend for the all-to-all communication used by mixture-of-experts models, alongside Ascend-targeted updates to TileLang, TileKernels, DeepSelect and FlashMLA. The DeepGEMM changelog dates the Ascend release to 2026.09.30; DeepGEMM itself is a tensor-core kernel library covering FP8, FP4 and BF16 GEMMs, fused MoE and MQA scoring, with all kernels compiled at runtime through DeepJIT. SemiAnalysis reported that at the time of writing only TileLang was usable, because DeepSeek requires CANN 9.2.0 and a C++20 compiler that the beta toolkit did not provide, while TechRadar said DeepGEMM-Ascend needs Ascend 950 series silicon and CANN 9.20. The README's claim that performance matches or exceeds expert-tuned libraries across matrix shapes is the maintainers' own and has not been independently measured.

   GitHub Trending（每日） · 🔥 0

Experience and discussion from the community

Share my TileLang experienceAsk about TileLang