hardware · 2026-11-19 · Tier 1

Models design and tune the silicon: OpenAI x AMD PPA, Exa on Inferentia, Meta REA, NVIDIA Daviator

Models design and tune the silicon (lab posts, 2026-11-18)

TL;DR. Four lab posts on the same day show AI agents working one layer below the model: on the chip, the kernel and the training recipe. OpenAI says its models found up to 14% PPA (power, performance and area, the three numbers every chip design trades against) on AMD flagship GPU IP in under five months, on designs AMD engineers had tuned for years; AMD has folded the changes into a design sent to production, and the same approach improved Jalapeño, OpenAI's first custom inference chip, with gains verified on real silicon (GPT-6 Astra was the model named). Exa re-embedded 30B documents on AWS Inferentia2/Trainium2 and reached about 1.5x throughput with an agentic development loop: weight-layout fixes, compiler autocast (gated at 0.999 cosine similarity), drop-in NKI attention, and a custom weight-stationary fused SwiGLU MLP kernel that removed HBM spill. Meta says agents now handle over 60% of "technique propagation" (porting a proven ranking-model improvement to other ads models), with 3x engineering productivity, a median 4x cut in training compute on select combinations and an estimated $45M+ compute savings this year. NVIDIA's Daviator batches k-point work in VASP with mixed FP32/FP64 precision for up to 22x faster time-to-solution and 75% lower cost than CPU.

The agent moved down the stack
Same loop at every layer: propose a change, verify it on real hardware, keep it only if it pays.
flowchart LR
  A["Agent<br/><small>proposes a change</small>"] --> C["Chip layout<br/><small>OpenAI x AMD, 14% PPA</small>"]
  A --> K["Kernel<br/><small>Exa NKI, 1.5x</small>"]
  A --> R["Recipe port<br/><small>Meta REA, 4x less compute</small>"]
  C --> V{"Verify on hardware<br/><small>silicon, profiler, A/B</small>"}
  K --> V
  R --> V
  V -->|pays| P["Production<br/><small>shipped design or model</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class A core
  class C,K,R input
  class V loop
  class P exit
  linkStyle 6 stroke:#2f9e44,stroke-width:2px
Purple is the agent, blue the layer it edits, amber the hardware check, green what ships.

Why it matters

  • PPA gains compound across every token served. A 14% PPA improvement on GPU IP is roughly a generation's worth of tuning; if it reaches production silicon it lowers cost per token for everyone buying AMD parts, not just OpenAI.
  • The verifier is the bottleneck, again. Every success here has a hard checker: real silicon, a 0.999 cosine gate, production A/B. This repeats the CompoundingKernel (11-18) finding that trustworthy timing runs per hour, not model strength, limited a 1,000-agent kernel search.
  • Skeptic's note (X feed): a fabless design is only as good as the fab capacity, HBM allocation and packaging behind it, which is why OpenAI's Broadcom deal matters more than the GDSII file.

How it relates to prior wiki pages

  • Extends the 11-11 to 11-18 gpu-kernels thread ("orchestration becomes the programming model") from kernels to physical design.
  • Pairs with AscendGYM (same day), which prices agent kernel search in tokens.
  • OpenAI's own data (research acceleration post) says agents succeed 79% of the time monitoring training runs but only 25% writing inference and serving code: the layer these posts celebrate is the one agents are weakest at in general, which makes the verifier-heavy setups above the exception, not the rule.

Sources: OpenAI chip design · The Decoder · Exa on AWS · Meta REA · NVIDIA Daviator · raw: labs 11-18 23:28, labs 11-19 11:12