AMD's Zen 6 Family: Feeding the Beast

AMD’s upcoming Zen 6 server and consumer CPUs prompt debate over how far core counts and microarchitectural tricks—like deeper out-of-order execution and advanced branch prediction—can go when main memory latency remains a key bottleneck. Commenters weigh trade-offs between adding more memory channels versus the physical, cost, and pinout limits of current sockets, and note that many workloads are latency- rather than bandwidth-bound. The thread also touches on the positioning of Zen 6 “c” efficiency cores, comparisons to Apple’s M-series and other architectures, and frustration with current AM5 memory controller limitations for high-speed, multi-DIMM setups.

Memory Controllers, Bandwidth, and Latency

  • Multiple commenters argue that cores are not typically “too fast” for memory controllers; most workloads are latency-bound, not bandwidth-bound.
  • Adding more controllers/channels is constrained by pin count, package size, board routing complexity, and cost (especially for motherboards).
  • Current AMD offerings already scale memory channels up to 16 on high-end EPYC; per-core channels would require tens of thousands of pins and extreme board complexity.
  • Some point out that “few workloads are bandwidth-bound” is partly self-fulfilling and that AI and image workloads can saturate bandwidth.
  • NUMA topology and cache organization often limit real-world gains from more bandwidth.

Zen 6 / Zen 6c Core Design and Scaling

  • Zen 6c increases out-of-order structures to better hide memory latency, making it attractive for high-core-count, bandwidth-constrained servers.
  • The jump from 12 to 32 cores per CCD via node shrink and cache/layout changes is seen as impressive but raises concern about rising L3 latency.
  • Unified L3 per CCD benefits gaming; partitioning CCD L3 again (as in Zen 1/2) is seen as unlikely while games are a key target.
  • Rumors mention future, costlier interconnects to reduce CCD-to-CCD latency.

Branch Prediction, ILP, and Microarchitecture

  • The sophisticated “2-ahead” branch prediction and instruction-level parallelism reflect how slow memory is relative to core speed, making complex speculation worthwhile.
  • Several comments note that modern cores are dominated by cache, reorder logic, and prediction, not ALUs; compute units are relatively small.
  • Historical contrast: earlier eras sometimes used lookup tables in RAM because memory was comparatively fast.

AMD vs Apple and Other CPUs

  • Some criticize charts omitting Apple’s newer cores, noting their strong single-thread SPEC scores.
  • Others respond that the article is about server parts, where Apple lacks an equivalent product and server-style I/O (many PCIe lanes, ECC, etc.).
  • Benchmarks like Stockfish show Zen 5/6 dominating multi-threaded throughput via many wide AVX-512-capable cores, while Apple does well given its narrower SIMD.
  • One side claims Apple’s efficiency cores could theoretically power a server chip beating Zen 6 in perf/watt/area; others stress Apple doesn’t target that market, so it’s hypothetical.

Consumer Platforms, Memory Overclocking, and IO Dies

  • The AM5 IO die and DDR5 memory controller are widely criticized, especially with four DIMMs: many users report difficulty exceeding ~3600 MT/s, despite official 5600 MT/s (2 DIMMs) specs.
  • Others report 4-DIMM success at 6000 MT/s+ with specific boards and kits, suggesting variability and tuning sensitivity (voltages, timings, AGESA versions).
  • There’s anticipation for a new, smaller-node IO die in Zen 6 consumer parts, especially for better memory behavior; some mention earlier Zen 5 IO-die plans were delayed/rolled back.
  • Enthusiasts express desire for more memory channels on consumer platforms (8–12) to exploit “ramageddon,” though others note it conflicts with socket and board constraints.

C-cores, E-cores, and Product Segmentation

  • Zen 6c is seen as increasingly optimized and potentially deserving its own branding as its design diverges.
  • In contrast, Intel’s E-cores are rumored to be converging toward P-core capabilities.
  • Discussion notes differing design goals: high-clock, lower-core Threadripper vs high-core EPYC, and frustration at incompatible sockets preventing easy CPU swaps between platforms.
  • Some say certain boards could electrically support both EPYC and Threadripper, but it would require complex and costly board-side logic.

ISA, Micro-ops, and RISC vs CISC

  • One commenter muses that x86 micro-ops should effectively be the ISA, preferring “pure” RISC designs like RISC-V.
  • Others respond that even RISC architectures now define instructions that intentionally decode into multiple micro-ops, blurring classical RISC/CISC distinctions.

Branch Prediction vs Branchless Code

  • A strong disagreement arises over whether branch prediction is a “bane” forcing branchless programming.
  • Some argue for instructions that disable prediction for specific branches; others reply this would just stall pipelines and likely hurt performance.
  • Experienced SIMD programmers emphasize that a good branch predictor helps more often than not; branchless code primarily shines for truly unpredictable branches or highly vectorized inner loops.
  • Hints like likely/unlikely or __builtin_expect are mentioned, but one side claims branchless programming exists because such hints and predictors still fail in many real cases.
  • There is consensus that the best approach is workload-dependent, and no single strategy (always branchless vs always branching) dominates.