您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：同一个模型、同一块芯片，换一套更优的算子，有时会带来数倍的性能提升。
国产大模型的竞争，正从模型层延伸到算力层：再强的模型，最终都要基于算子在芯片上执行。同一个模型、同一块芯片，换一套更优的算子，有时会带来数倍的性能提升。算子的效率决定了同样的显卡能跑多快，也决定了每个 token 的价格。
DeepSeek 近日全面开源推理框架 DeepInfer。继模型权重和 FlashMLA、DeepGEMM 等算子库之后，他们把模型运行的链路开放了。其中的 FlashMLA，有工程师在芯片上逐行手调的版本，算子性能跑到了硬件极限的 90% 以上，这也是人类手艺目前在部分芯片上达到的极限成绩。
而这个极限，同一周就被 AI 刷新了。据悉，截至 10 月 22 日，完全由 AI 编写算子、无人工逐行调优的 ForgeKernel，同时登顶两个算子评测榜：英伟达主场的 SOL-ExecBench FlashInfer-Bench 合集第一，以及昇腾 CANN 官方仓库维护的 CANNBench A3 硬件榜第一，基准加速比 1.835×，领先第二名 25%。在名次之外，AI自主优化算子的能力不容小觑：在DeepSeek v4.1 Flash的算子上，相比DeepSeek开源的英伟达版本，ForgeKernel也通过从0用AI迭代的方式实现了性能超越。
这是智能进化的接力。无论是DeepSeek的开源精神还是算子优化榜单的焦灼竞争，最终都导向同一个方向：让国产智能在每一块芯片上都跑得更快。
英伟达 SOL-ExecBench FlashInfer-Bench
昇腾 CANNBench A3 硬件榜
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: The same model, the same chip, and a better set of algorithms will sometimes result in several times the performance increase.
Competition for the national production of LLM is extending from the model layer to the computational power layer: the stronger model is ultimately based on the algorithm for the chip. The same model, the same chip, and a better set of algorithms, sometimes result in several times the performance increase. The efficiency of the algorithm determines how fast the same GPU can run and the price of each token.
DeepSeek's recent comprehensive open source reasoning framework. Following the model weights and the algorithms of FlashMLA, DeepGEMM and others, they opened up the link in which the model operated. One of them, FlashMLA, has an engineer-to-man version of the chip, and the algorithm's ability to reach more than 90% of the hardware limit, which is the limit of what human crafts currently achieve on some of the chips.
And this limit was updated by AI that same week. As of 22 October, it was known that Forge Kernel, which had been prepared by AI entirely and was no-man-by-man, was on the top of the two algorithms: Sol-ExecBench FlashInfer-Bench, first at the main Yin Weida site, and CANNBech A3 hardware list, which was maintained at the Canadian official warehouse, with a benchmark acceleration of 1.835 x and a second 25 per cent. In addition to their ranking, the AAI's ability to optimize their algorithms on an autonomous basis cannot be underestimated: in DeepSeek v. 4.1 Flash, Forge Kernel also outperformed it from 0 to 0 in an AI-over.
This is the relay of the evolution of intelligence. The open-source spirit of DeepSeek and the calculator-optimizing competition ultimately lead in the same direction: to make domestic intelligence run faster on every chip.
SOL-ExecBench FlashInfer-Bench
CANNBENCH A3 Hardware List
For your account security, please verify the mailbox.
Please verify your mailbox.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.