AI专用芯片御三家：Celebras vs Groq vs Taalas（今天的主角）
https://www.youtube.com/watch?v=_OoB3WxBQq0 这期由 基地 发布的视频，深度剖析了 AMD 在 2026 年 8 月收购芯片初创公司 Taalas 及其推出的 HCE（Hardcoded Engine） 芯片背后的技术细节与商业赌局。 主要内容梳理如下： 一、 疯狂的硬件设计：代码即物理形状 [ 00:00 ] 固化模型 ：普通 GPU（如 H100、MI300）运行 AI 时需要不断从显存中提取模型权重；而 Taalas 芯片 没有外挂高带宽内存（HBM） ，它直接将 Llama 3.1 8B 等大模型的权重以物理金属互连（过孔）的形式“焊”死在了芯片硬件上 [ 00:34 ]。改代码/模型就必须重新改芯片物理形状 [ 00:48 ]。 降本核心（只改最上 2 层金属） ：芯片底层 90 多层电路（逻辑门、累加器等）全是一样的通用晶圆 […
A recent video discusses the recent acquisition of AI chip startup Taalas by AMD and the development of its HCE (Hardcoded Engine) chip. The video delves into the unique technology behind Taalas' chip, which replaces traditional GPUs with a chip that has its model weights physically "welded" into the hardware, eliminating the need for high-bandwidth memory (HBM).
This design results in significant cost savings and efficiency improvements, but also limits the chip's ability to run different models. The video contrasts Taalas' approach with Cerebras' approach of using a whole silicon wafer to create a single large chip, and examines the strategic motivations behind AMD's decision to acquire Taalas.
Overall, the video provides a detailed analysis of the cutting-edge AI chip market and its key players, including AMD, Taalas, Cerebras, and Groq.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
