Die Shrink 185: UDNA, Diamond Rapids, Wildcat Lake, Nvidia Cult (Hot Chips 38)
Summary
Host Tom and guest Carbonryon from Moore’s Law Is Dead recap Hot Chips, analyzing AMD’s CDNA 5 cache hierarchy and RDNA 5 implications. They discuss NVIDIA’s Rubin architecture, Vera CPU’s multi-threading, and Intel’s Crescent Island and Diamond Rapids designs. The episode concludes with an assessment of Intel’s server market challenges and the need for sustained execution to regain momentum against AMD.
Topics discussed
Introduction and Hot Chips 2024 overview
AMD CDNA 5 cache hierarchy and broadcast arbitrator
RDNA 5 leaks, L0/L1 cache changes, and ray tracing
AI inference workloads and consumer GPU features
Nvidia Rubin architecture and Vera CPU design
Intel Crescent Island AI accelerator and gaming future
Intel chiplet packaging and Diamond Rapids server CPU
AMD Zen 6 Venice vs Intel performance and market share
Intel Wildcat Lake and Bumblebee low-end CPUs
PCIe lanes and workstation expandability
RISC-V adoption in microcontrollers and handhelds
Networking silicon generations and copper links
IBM dual-ISA Z CPU running ARM and x86
Outro, sponsor thanks, and Nvidia presentation critique
Listen ad-free on Castria