Sony's Mark Cerny Has Worked on "Big Chunks of RDNA 5" with AMD

Original Article ↗ Hacker News Discussion ↗

Mark Cerny, RDNA5, and AMD Collaboration

  • Commenters note Cerny appears to work as a consultant rather than a Sony employee, influencing both PlayStation SoCs and AMD GPUs.
  • The article’s “RDNA5” branding is questioned: Cerny himself is quoted as saying “RDNA 5, or whatever AMD ends up calling it,” suggesting a name in flux.

RDNA, CDNA, and UDNA Convergence

  • Several posts argue that AMD’s public roadmap ends RDNA at 4, with a shift to “UDNA1,” a unified architecture with CDNA (HPC/datacenter).
  • There’s disagreement on how similar RDNA and CDNA already are: some claim broad commonality, others detail substantial differences (wavefront width, execution model, latency, feature sets).
  • UDNA is seen as both an architectural and organizational consolidation, potentially merging teams and long-term strategy.

Console Custom Silicon and Semi-Custom Paths

  • Sony could theoretically request an RDNA4-derived design instead of adopting early UDNA, resulting in a “RDNA5” that remains semi-custom and never appears as a retail GPU.
  • Past semi-custom work (e.g., console/APU overlap, Steam Deck–style chips) is cited as precedent, with console learnings later feeding into APUs.

Generational Leap: PS4 → PS5

  • Many feel the visual jump from PS4 Pro to PS5 is small relative to the FLOPS increase, especially compared with earlier eras (PS1→PS3).
  • Explanations offered:
    • More pixels (4K vs 1080p) consume much of the extra compute.
    • Hardware progress has slowed (Dennard scaling and Moore’s law weakening, rising node costs).
    • Huge performance gains now often go into higher FPS and faster loading instead of visibly new effects.
    • The biggest PS5 leap is widely credited to NVMe SSD + hardware decompression, not raw GPU power.

Engines, Bloat, and Unreal

  • Some argue engine practices, especially Unreal’s default pipeline optimized for highly dynamic scenes, waste potential for many game types.
  • Others respond that fully dynamic lighting and environments dramatically improve workflows and design freedom, even if they cost performance.

Anti-Aliasing and Image Quality Tradeoffs

  • TAA is debated:
    • One side says it’s an efficient, necessary replacement for supersampling/MSAA and has improved significantly.
    • Critics argue temporal and upscaling techniques sacrifice clarity, introduce ghosting/blur, and inflate FPS metrics while degrading real image quality.
    • There’s agreement that objective metrics for temporal artifacts are poor, making evaluation hard.

Performance vs Fidelity and Player Preferences

  • Multiple comments say most players choose performance modes when presented with a choice; a cited Sony stat claims ~75% pick performance.
  • Others question how representative this is across genres, noting fast competitive games may skew the data.
  • Some users report being FPS-tolerant (e.g., 25–30fps is acceptable if stable); others insist modern displays make low framerates intolerably blurry.

Optimization Culture and Rising Costs

  • Several posts lament that modern games are less optimized, with studios relying on hardware advances and middleware.
  • A few argue the bottleneck has shifted:
    • Asset production (high-res models/textures) dominates cost, leading to teams with far more artists than programmers.
    • AAA engines increasingly optimize for artist workflows rather than peak runtime performance.
  • Others counter that even in earlier eras, many console games already used C and higher-level tooling; the “all hand-tuned asm” narrative is overstated.

Storage, Streaming, and New Techniques

  • PS5’s SSD and streaming capabilities are highlighted as enabling design changes (fewer fake loading corridors, highly detailed continuous worlds).
  • Examples mentioned include Cyberpunk’s serialization bottlenecks on PS4 and newer techniques like Nanite:
    • Supporters say Nanite shines in extremely complex scenes and is optional.
    • Critics say it adds overhead and can hurt performance in simpler content.

Hardware Progress, GPUs, and AI/Crypto

  • Some posters attribute weaker generational jumps to fundamental tech limits (SRAM/IO scaling stalling, expensive shrinks).
  • Others note that GPU vendors now prioritize datacenter/AI features (high VRAM, interconnects), potentially slowing pure gaming advances; there’s disagreement on how much this affects consoles.
  • A side thread argues that software has also grown more capable at leveraging parallelism, while another stresses that fully saturating modern multi-core + GPU + NPU systems remains rare.

APIs and Vulkan-on-PlayStation Debate

  • One question asks why Sony doesn’t support Vulkan on PS5.
  • Defenders of Sony’s proprietary APIs (GNM/GNMX) say consoles benefit from ultra-low-level, hardware-specific interfaces and avoid Khronos politics.
  • Pro-Vulkan voices argue standards reduce developer burden and avoid NIH/lock-in; they criticize Vulkan’s extension “spaghetti” but still see it as the best collaborative option.
  • There’s nuanced discussion of Vulkan’s strengths (barrier model, SPIR-V) and weaknesses (complex extensions, OpenGL legacy).

AMD Software Stack and ROCm

  • A commenter reports that recent ROCm releases now “just work” with tools like llama.cpp on AMD GPUs, contrasting with years of painful setup.
  • Others note llama.cpp can bypass ROCm entirely via Vulkan, but ROCm compatibility is treated as a useful barometer of AMD’s software maturity.

Miscellaneous

  • Some express excitement that UDNA and Cerny’s work could improve AMD’s datacenter competitiveness against Nvidia, with the caveat that poor drivers/support could again damage trust.
  • There’s skepticism about current-gen consoles’ limited exclusive library, but anticipation that titles like GTA VI may finally showcase the hardware.