DeepSeek engineer says AI could match his GPU-kernel work within a year

A DeepSeek engineer says AI-written GPU kernels could match or exceed his own work within six to 12 months, while a separate NVIDIA experiment shows a narrower version of that workflow already producing numerically correct code. The important shift is from AI assisting with low-level optimisation to agents searching through hardware-specific implementations themselves.

DeepSeek Watch analysis

What happened

Shengyu Liu’s forecast, reported by NeoTeo, concerns GPU kernels, the specialised operations that control how AI models use memory, arithmetic units and parallel threads. He says his work has moved from writing and tuning kernels towards directing agents, setting objectives and judging their output.

Liu’s prediction is personal, not a measured forecast of the whole profession. But the underlying technique is not merely theoretical. NeoTeo describes an NVIDIA experiment in which DeepSeek-R1 generated candidate CUDA C++ attention kernels for Hopper hardware. A verifier tested each candidate, refined the prompt when it failed and accepted code only when it met the stated criteria.

That experiment reached 100 per cent numerical correctness on KernelBench Level 1 and 96 per cent on Level 2, according to the report. Those results answer a narrow question. They do not show that generated kernels are better across production workloads, GPU generations or optimisation goals.

DeepSeek’s TileKernels project covers operations including mixture-of-experts routing, quantisation, transpose and gating. The project targets NVIDIA SM90 and SM100 hardware and requires recent versions of Python, PyTorch, TileLang and CUDA, according to the supplied report. The practical lesson is less magical than the marketing: the agent generates possibilities, while the verifier and engineer decide what counts as usable.

Why it matters

GPU-kernel engineering sits below the chatbot interface but directly affects how quickly models respond and how much hardware they need. A faster operation can reduce training or inference time, but only if it preserves numerical behaviour and works within the surrounding model architecture.

This is therefore a story about computing craft, not a general claim that software jobs are about to vanish. AI may take over more of the search and implementation work while leaving people responsible for objectives, profiling, verification and the awkward question of whether a local speed-up survives contact with a real system.

Our read

The striking development is not that a model can write CUDA. It is that code generation is being joined to an automated test-and-retry loop aimed at specialised hardware. That is a much more useful milestone than another leaderboard flourish.

The sensible response is neither panic nor applause on autopilot. Engineers should judge these systems by reproducible performance, correctness and failure handling across their own workloads. The verifier is the star of this show, and it has no patience for vibes.

What to watch

• Production gains:
  Whether generated kernels improve complete model workloads, rather than only benchmark cases.
• Hardware coverage:
  Whether the approach transfers beyond the reported Hopper setup and listed NVIDIA targets.
• Verification quality:
  Whether tests catch numerical errors, regressions and fast-but-wrong implementations.
• Changing roles:
  Whether kernel specialists spend more time directing automated searches and less time writing every implementation by hand.

Discussion spark: If AI agents can search and verify GPU kernels, should engineers trust them with performance-critical code now, or does the verifier still need far more independent testing before the craft can safely shift from writing to supervision?

not affiliated with or endorsed by DeepSeek