Video summary

AMD Zen 5深度评测：HX370能效超强！但大小核靠谱吗？

Main summary

Key takeaways

Technology

Product/Platform Overview (AMD “Strix Point” / Zen 5 Notebook SoC)

A review deep-dive into AMD’s Strix Point notebook SoC (Zen 5, early June), focusing on the flagship Ryzen AI 9 HX 370.

Core/SoC Configuration

  • CPU layout
    • 4× Zen 5 “P-cores” + 8× Zen 5C “E-cores”
    • Described as a “homogeneous big/little” approach within the Zen 5 family.
  • GPU
    • 16CU RDNA 3.5
    • Includes an XDNA2 NPU (noted as part of the SoC).
  • Cache differences
    • P-cores share 16MB L3
    • E-cores (Zen 5C) get 8MB L3
    • Implies much smaller per-core L3 on E-cores.
  • Frequency differences
    • P-cores: up to ~5.1GHz
    • E-cores: ~3.3GHz
    • E-cores also have hyper-threading.
  • Physical/CCX organization
    • P and E cores reside in two CCXs
    • Connected via Infinity Fabric
    • Raises concerns about inter-core communication latency

Zen 5 Microarchitecture Improvements (IPC-Focused)

The analysis emphasizes improvements aimed at IPC (instructions per cycle).

Front-End (More Parallel Work)

  • Broader parallelism, including:
    • 2× decode units
    • 12-instruction Op-cache (vs 9)
    • 64 decoded instruction bytes/cycle (vs 32)
    • Higher dispatch bandwidth

Back-End / Execution

  • Changes to integer/FP execution resources
  • Deeper scheduler/queue
  • Adds/expands a 512-bit AVX-512 data path
  • AMD claim noted: less frequency drop vs Intel during AVX-512 workloads

Memory Subsystem

  • More AGUs (3 → 4)
  • L1 improvements
  • Larger/different DTLB
  • L2 bandwidth doubled
    • Capacity remains ~1MB per core

Measured IPC

  • An official IPC uplift is cited at ~16%
  • Reviewer measured about ~17% IPC in SPEC2017 (approx.)
  • Breakdown vs prior gen:
    • ~10% integer
    • ~25% floating-point
  • Context: comparisons reference Intel Meteor Lake big-core variants as a reference point for Zen 5 “big core” discussion

Energy Efficiency Testing (SPEC, Multiple Frequencies)

Testing Approach

  • Reviewer builds single-core SPEC energy-efficiency curves
  • Covers roughly 2.3GHz up to maximum

Key Findings

  • Zen 5 integer energy efficiency:
    • Worse than Zen 4 at low-power points
    • Significantly better at high-performance points
  • Overall characterization:
    • Zen 5 overall efficiency is described as quite good
    • Floating-point efficiency is portrayed as especially strong
  • At full frequency:
    • Single-core performance ~20% ahead of 8845H in SPEC2017
    • This drives excitement for real-world results

Major Performance Caveat: Inter-Core Delay & Game Scheduling Risk

Inter-Core Latency Measurement

  • Reviewer reports inter-core delays around ~180ns
  • Compared against Intel and older AMD parts, this is described as unusually high

Why It Matters (Games)

  • If a game uses only P-cores (4 cores): performance may be fine
  • If a game is scheduled across P/E or on E-cores:
    • performance may drop due to added latency

Interpretation of “Zen 5C” Impact

  • Zen 5C is largely the same architecture as Zen 5, but with:
    • smaller L3
    • lower frequency
  • The reviewer also criticizes scheduling maturity:
    • when games land on small cores, results can plummet
    • especially in multi-core and/or cross-core communication-heavy situations

Bottom line: the CPU may be excellent for some throughput patterns, but not ideal for consistent multi-core gaming behavior.

GPU / Core Graphics Performance + Bottleneck Analysis

Benchmarks

  • 3DMark Time Spy
    • 890M (16CU RDNA 3.5) vs prior gen 780M
  • Steel Nomad Light
    • Also improved vs previous gen

Scaling Issue

  • Despite low-frequency efficiency advantages, performance at higher power didn’t scale as expected vs competitors.

Hypothesis: Memory Bandwidth Bottleneck

  • LPDDR5X-7500 is referenced
  • However, bandwidth appears similar to previous gen
  • No Infinity Cache
  • Supporting experiment:
    • Reducing GPU frequency shows performance increases become frequency-dependent
    • Interpreted as memory-limited behavior

“Game-style” Take

  • Core graphics improvement is described as big vs 780M
  • Some previously unplayable 1080p scenarios become more feasible in select games

CPU Performance in Application Benchmarks

Cinebench R23

  • Multi-core energy efficiency improves substantially for 4+8 HX370
  • Compared favorably to:
    • 8-core 8945H
    • Ultra 9 185H (6+8)
    • (in comparable power scenarios)
  • Also strong in:
    • multi-core and single-core Cinebench 2024

Blender

  • HX370 ahead of 8945H and Ultra 9 185H
  • Both single and multi results described as strong

7-Zip

  • Mentions improvement in compression/decompression versus competitors

Gaming Results (Core Graphics vs CPU Gaming with dGPU)

iGPU-Only (Core Display) at 1080p / 1440p

  • CS2
    • Strong esports performance
    • ~120fps at 1080p on low settings mentioned
    • ~25% faster vs 780M
  • Dauntless
    • ~39% better vs 8845H
    • ~180fps range described
  • Some games show smaller uplifts or even losses:
    • examples include Elden Ring (lower gains)
    • and titles where HX370 trails due to memory sensitivity and CPU↔GPU memory contention

Overall iGPU uplift summary (as stated by the reviewer):

  • vs 8845H / 780M: about +27%
  • vs Ultra 9 185H: about +11%

With dGPU (paired with 4070)

  • CPU gaming gains are described as subtle and inconsistent
  • Some titles benefit, but others may show:
    • Elden Ring and Zero Zone where older 8945H does better
  • BIOS update test
    • Updating BIOS on a related laptop improves FPS
    • Even after improvement, reviewer says results still don’t meet expectations

Core Locking Test (4P Only)

  • Limiting to 4 big cores improves performance
  • The implication: the “4P + 8E” design drives the problem
  • Proposed root causes:
    1. High inter-core delay between big and small cores
    2. Zen 5C small-core cache/frequency cuts hurt certain game patterns
    3. Scheduling strategy not mature—games may be allocated to E-cores and lose FPS

Conclusion for Gaming Laptops

  • The design may be better for CPU/throughput or scheduling-friendly cases
  • But it’s described as not ideal for gaming laptops that require consistent multi-core gaming performance

Battery Life / Laptop Practicality

Example Configurations

  • ASUS ROG / Chuang 16: 90Wh + RTX 4070
    • ~5h 13m
  • Lingyao 16 Air: 78Wh + core graphics
    • ~8h 12m

Summary

  • Battery life is described as close to previous-gen for this class
  • Positioned as an all-round + gaming laptop outcome

Overall Verdict & Forward-Looking Concerns

What Strix Point Delivers

  • Double-digit IPC improvements
  • Strong high-frequency single-core energy efficiency
  • Better multi-core performance and efficiency
  • Noticeably stronger iGPU core graphics

Major Caveats

  • High inter-core delay
  • Immature big/little scheduling
    • hurting multi-core gaming performance, especially when dGPU is involved
  • SoC PCIe channel limitations are mentioned as making it less suitable for heavy-duty gaming laptops with independent GPUs

Speculation on Future Products

  • Interest in:
    • a full Zen 5 all-big-core mobile chip focused on gaming
    • a Zen 5-based desktop Ryzen 9000-series
  • Tests are described as “in progress.”

Main Speakers / Sources

The subtitles do not clearly name individual speakers. It appears to be a single reviewer/host delivering the deep-dive analysis and benchmarks (no specific host name provided in the provided subtitles).

Original video