Hasty Briefsbeta

A Look into Intel Xeon 6's Memory Subsystem

8 hours ago
  • #Server Processors
  • #Chiplet Design
  • #Intel Xeon 6
  • Intel's Xeon 6 server platform introduces a scalable chiplet strategy to compete with AMD and Arm.
  • Xeon 6 uses a side-by-side chiplet arrangement with up to three compute dies and two IO dies, scaling to 128 cores per socket.
  • AWS offers Xeon 6 instances (r8i) with the Xeon 6 6985P-C, featuring 96 Redwood Cove cores and AVX-512 support.
  • Xeon 6's mesh interconnect and Modular Data Fabric (MDF) enable high cross-die bandwidth but with higher L3 latency.
  • Memory latency on Xeon 6 is higher than AMD's Turin platform, despite Intel's SNC3 NUMA configuration.
  • Xeon 6 provides massive DRAM bandwidth (691.62 GB/s) but lags behind AMD in L3 bandwidth and latency.
  • Intel's logically monolithic design offers advantages in L3 capacity and core-to-core latency but faces scalability challenges.
  • AMD's Turin platform uses a different interconnect strategy, focusing on cluster-level efficiency rather than a monolithic design.
  • Xeon 6's single-core performance in SPEC CPU2017 is comparable to AWS's Graviton 4, with a slight lead in floating-point performance.
  • Future Intel server designs, like Lion Cove, aim to better compete with Zen 5, but tradeoffs in monolithic designs remain a challenge.