AMD Zen架构首席设计师Mike Clark在采访中详细解析了下一代Zen 6及Zen 6c核心微架构的技术演进。他阐述了Zen 6如何通过优化前端解码、分支预测器及执行引擎来提升IPC,并介绍了针对AI负载增加的VNNI单元和向量加载能力的改进。此外,他还说明了不同核心家族(高频与高密度)在L3缓存配置和调度策略上的差异化设计逻辑。1
在核心性能与兼容性方面,Zen 6c被视为最高性能的核心,每核配备4MB L3缓存,单CCD最多32核,总L3达128MB。1为了保持软件兼容性,Zen 6c虽频率较低,但通过扩大队列容量(如浮点调度器)来提升IPC。1测试显示,Zen 6c在SPECint 2026测试中,尽管时钟频率较低,其性能已匹配最快的Zen 5服务器CPU。1
在执行引擎与向量处理单元的硬件设计上,整数乘法器数量从Zen 5的3个减少至2个,以配合分布式调度器并优化面积效率。1向量侧新增2个VNNI单元,以更好地支持推理等AI工作负载。1同时,向量加载能力从Zen 5的2次提升至4次(支持256位)。1
关于内存子系统与分支预测优化,LPDDR5X内存技术的引入对核心内部设计影响不大,主要关注点在于CCD到IOD的带宽。1L2缓存带宽已优化至可稳定达到64字节/周期。1分支预测器不仅延迟降低,准确率也高于Zen 5。1
Mike Clark, the chief architect for AMD's Zen family, provided a detailed technical overview of the upcoming Zen 6 and Zen 6c microarchitectures. 1 The presentation focused on how Zen 6 improves Instructions Per Cycle (IPC) through optimizations in front-end decoding, branch prediction, and the execution engine. 1 Specifically, the new branch predictor offers lower latency and higher accuracy compared to Zen 5. 1
To support increased AI workloads, the architecture introduces two new VNNI (Vector Neural Network Instructions) units. 1 Vector load capability has been doubled from two operations per cycle in Zen 5 to four operations per cycle in Zen 6, supporting 256-bit vectors. 1 Additionally, L2 cache bandwidth has been optimized to consistently achieve 64 bytes per cycle. 1 In terms of area efficiency and scheduler integration, the number of integer multipliers was reduced from three in Zen 5 to two in Zen 6. 1
The design distinguishes between high-frequency and high-density core variants through different scheduling strategies and cache configurations. 1 The Zen 6c is described as the highest-performance core within its category, featuring 4MB of L3 cache per core. 1 A single CCD with Zen 6c cores can house up to 32 cores, resulting in a total L3 cache capacity of 128MB. 1 Although Zen 6c operates at lower clock frequencies, it maintains software compatibility and achieves performance parity with the fastest current-generation Zen 5 server CPUs in SPECint 2026 benchmarks. 1 This performance is achieved by expanding queue capacities, such as the floating-point scheduler, to boost IPC despite the lower frequency. 1
Regarding memory subsystems, the introduction of LPDDR5X technology has minimal impact on the internal core design. 1 Instead, engineering efforts are primarily focused on maximizing bandwidth between the Core Complex Dies (CCDs) and the I/O Die (IOD). 1
评论
还没有评论,欢迎留下第一条。