Redwood团队宣布其AI系统成功实现端到端芯片设计突破,自主发现、设计并验证了一款名为Frostbite的新型通用CPU架构 1。该架构由AI从零开始构建指令集和微架构,无需人工预设先验知识 1。这一成就标志着人工智能完成了从理论探索到物理布局的全流程自动化设计 1。
Frostbite架构基于“交互网络”(Interaction Nets)理论,利用局部步骤、显式依赖和单值所有权等特性进行开发 1。在测试中,该架构展现出显著的性能优势:在7nm工艺节点下,与采用超标量双发射顺序执行模式的现代硅验证CPU核心相比,Frostbite实现了平均4.59倍的每瓦性能提升 1。具体数据显示,其在搜索集工作负载上的性能提升了6.02倍,而在未调优的保留测试集上也有3.83倍的提升,同时每面积性能提高了8.77倍 1。
该AI系统在迭代过程中自主处理了编译器、ISA、架构、微架构、RTL代码生成、验证以及最终GDS布局等环节 1。值得注意的是,设计中涌现出了“窃取工作”(work-stealing)等硬件特性,这种并行性直接源于程序结构而非传统线程模型,且系统中不存在中心调度器 1。为确保设计的正确性,团队使用内部形式化规范语言ASL,并最终生成了超过80万行Lean证明代码 1。在物理参数方面,单个Quad加L3节点的面积为0.792平方毫米,综合频率达到800 MHz,而最终的GDS布局频率为500 MHz 1。
Redwood has announced that its artificial intelligence system autonomously discovered, designed, and verified a new general-purpose CPU architecture named Frostbite. 1 This development represents a significant breakthrough in end-to-end chip design, as the AI constructed the instruction set and microarchitecture from scratch without relying on any human-provided prior knowledge. 1
The architecture is grounded in Interaction Nets theory, utilizing characteristics such as local steps, explicit dependencies, and single-value ownership. 1 Through an iterative process, the AI independently handled the compiler, Instruction Set Architecture (ISA), architectural design, microarchitecture, Register-Transfer Level (RTL) code, verification, and the final GDS layout. 1 During this autonomous design phase, emergent hardware features such as "work-stealing" were identified, allowing parallelism to arise from program structure rather than threads, eliminating the need for a central scheduler. 1
In performance benchmarks conducted at a 7nm process node, Frostbite demonstrated an average 4.59 times improvement in performance per watt compared to modern silicon-verified superscalar dual-issue sequential execution CPUs (TCPU). 1 The geometric mean of results across 25 workloads showed specific enhancements, including a 6.02 times boost for search set workloads and a 3.83 times gain for the holdout test set, which was not tuned during development. 1 Additionally, the architecture achieved an 8.77 times increase in performance per area. 1
To ensure correctness, the project employed an internal formal specification language called ASL, ultimately generating over 800,000 lines of Lean proof code. 1 Physically, a single Quad plus L3 node occupies an area of 0.792 mm², with a synthesis frequency reaching 800 MHz, while the final GDS layout operates at 500 MHz. 1
评论
还没有评论,欢迎留下第一条。