Home / Technology / Engram Unlocks AI's True Reasoning Power
Engram Unlocks AI's True Reasoning Power
18 Jan
Summary
- Engram decouples AI memory storage from computation, easing hardware demands.
- This method uses knowledge lookups, freeing GPU capacity for complex reasoning.
- Engram promises linear memory scaling across GPUs with minimal overhead.
A novel AI training method named Engram has been introduced by DeepSeek in collaboration with Peking University. This innovative system decouples memory storage from computational processes, directly addressing the performance and cost bottlenecks caused by high-bandwidth memory requirements in traditional large language models.
Engram allows AI models to efficiently retrieve essential information through hashed N-grams, a technique that ensures static memory access independent of the current context. The retrieved data is then refined using a context-aware gating mechanism, enabling models to handle long context inputs more efficiently and supporting system-level prefetching with minimal performance impact.
This approach has demonstrated measurable improvements on a 27-billion-parameter model, offering predictable gains without additional computational cost. Engram's design facilitates linear memory scaling across multiple GPUs, potentially alleviating pressure on expensive memory hardware and mitigating volatile DRAM price swings.
