{
  "schemaVersion": 1,
  "id": "atlas-lesson-chip-workload-contract",
  "title": "Workload contract and architecture selection",
  "category": "learning",
  "format": "Atlas lesson",
  "description": "Turns desired service behavior into measurable hardware and software requirements.",
  "canonicalUrl": "https://semiaifoundry.com/learn/atlas/chip-workload-contract/",
  "publishedDate": null,
  "revisionDate": "2026-10-07",
  "author": null,
  "topic": "Custom silicon ecosystem",
  "series": null,
  "license": null,
  "rightsUrl": "https://semiaifoundry.com/ai-atlas/licenses.html",
  "version": "1.3.0",
  "status": "Published educational lesson",
  "evidenceScope": "Workload mix values are synthetic. Public co-design accounts do not reveal a provider’s complete internal traces, IP contracts or acceptance thresholds.",
  "contentCoverage": "published-lesson",
  "artifactUrl": "https://semiaifoundry.com/ai-atlas/lesson-data/chip-workload-contract-5b9ec61ef1e7641d.json",
  "interactiveUrl": "https://semiaifoundry.com/ai-atlas/#chip-workload-contract",
  "related": [],
  "supersedes": [],
  "supersededBy": [],
  "contentHash": "32566e9f8893a06146760f57b889b24bf0471ae00cd032ee7889a83853d47579",
  "contentLength": 5551,
  "jsonUrl": "https://semiaifoundry.com/agents/publications/atlas-lesson-chip-workload-contract.json",
  "markdownUrl": "https://semiaifoundry.com/agents/publications/atlas-lesson-chip-workload-contract.md",
  "sections": [
    {
      "id": "section-1",
      "title": "Workload contract and architecture selection",
      "start": 41,
      "end": 428
    },
    {
      "id": "section-2",
      "title": "The mechanism",
      "start": 428,
      "end": 1718
    },
    {
      "id": "section-3",
      "title": "Science and assumptions",
      "start": 1718,
      "end": 1746
    },
    {
      "id": "section-4",
      "title": "Workload-weighted demand",
      "start": 1746,
      "end": 2516
    },
    {
      "id": "section-5",
      "title": "Engineering example",
      "start": 2516,
      "end": 2740
    },
    {
      "id": "section-6",
      "title": "Distinctions and tradeoffs",
      "start": 2740,
      "end": 3199
    },
    {
      "id": "section-7",
      "title": "Specifications and interfaces",
      "start": 3199,
      "end": 3478
    },
    {
      "id": "section-8",
      "title": "Interfaces",
      "start": 3478,
      "end": 3684
    },
    {
      "id": "section-9",
      "title": "Check your understanding",
      "start": 3684,
      "end": 3713
    },
    {
      "id": "section-10",
      "title": "Which requirement most directly prevents a misleading inference performance target?",
      "start": 3713,
      "end": 4163
    },
    {
      "id": "section-11",
      "title": "Evidence and scope",
      "start": 4163,
      "end": 4511
    },
    {
      "id": "section-12",
      "title": "Sources",
      "start": 4511,
      "end": 4986
    },
    {
      "id": "section-13",
      "title": "Continue exploring",
      "start": 4986,
      "end": 5009
    },
    {
      "id": "section-14",
      "title": "Foundations",
      "start": 5009,
      "end": 5223
    },
    {
      "id": "section-15",
      "title": "Related lessons",
      "start": 5223,
      "end": 5551
    }
  ],
  "citations": [
    {
      "title": "Explore in Atlas",
      "url": "https://semiaifoundry.com/ai-atlas/#chip-workload-contract"
    },
    {
      "title": "Inside the hardware and co-design of MTIA",
      "url": "https://engineering.fb.com/2024/08/22/ml-applications/meta-mtia-hardware-co-design/"
    },
    {
      "title": "Chiplet Actuary: A Quantitative Cost Model and Multi-Chiplet Architecture Exploration",
      "url": "https://arxiv.org/abs/2203.12268v4"
    },
    {
      "title": "Published lesson data",
      "url": "https://semiaifoundry.com/ai-atlas/lesson-data/chip-workload-contract-5b9ec61ef1e7641d.json"
    },
    {
      "title": "Atlas releases and licensing",
      "url": "https://semiaifoundry.com/ai-atlas/licenses.html"
    },
    {
      "title": "Inference serving and request scheduling",
      "url": "https://semiaifoundry.com/learn/atlas/serving/"
    },
    {
      "title": "Quantization and numerical representation",
      "url": "https://semiaifoundry.com/learn/atlas/quantization/"
    },
    {
      "title": "Custom inference silicon: Maia and MTIA",
      "url": "https://semiaifoundry.com/learn/atlas/inference-custom-silicon/"
    },
    {
      "title": "RTL and functional verification",
      "url": "https://semiaifoundry.com/learn/atlas/chip-rtl-verification/"
    },
    {
      "title": "Custom-chip volume economics",
      "url": "https://semiaifoundry.com/learn/atlas/chip-volume-economics/"
    }
  ],
  "recordHash": "8f7c2afec941420c4cf926f744e8efa7fa28cfc52b5fcd7b88a9a29c550bf530",
  "contentFormat": "text/markdown",
  "content": "Custom silicon ecosystem · Atlas v1.3.0\n\n# Workload contract and architecture selection\n\nTurns desired service behavior into measurable hardware and software requirements.\n\n[Explore in Atlas](https://semiaifoundry.com/ai-atlas/#chip-workload-contract)Design abstractions to manufactured package; not one physical scale\n\nLearning objectiveTranslate model behavior and service objectives into a reviewable architecture contract.\n\n## The mechanism\n\nA custom-chip project begins by defining the work it must perform and the conditions under which success will be measured. A workload contract records operator shapes, numerical quality, latency targets, throughput, state capacity and expected model changes. Architecture exploration then asks how much arithmetic, memory, communication and control are needed together. Existing IP can supply processors, interfaces or other blocks, but its licensing and integration constraints matter. Meta’s MTIA co-design account is one public example of workload-driven development. A chip that excels on one benchmark can still fail the contract if the production graph falls back to a slow unsupported path.\n\n1.  Collect production-like traces, including unusual sequence lengths, operator shapes and tail-latency cases.\n2.  Define acceptable numerical quality and measurement conditions so the performance target has a stable denominator.\n3.  Build resource models for arithmetic, memory traffic, resident state and inter-device communication.\n4.  Prototype the compiler and critical kernels against architectural alternatives before fixing the hardware interfaces.\n5.  Record unsupported paths and change scenarios, then use these as explicit risks in the architecture decision.\n\n## Science and assumptions\n\n### Workload-weighted demand\n\n$$\nD_{\\mathrm{resource}}=\\sum_k\\lambda_k q_k\n$$\n\nAggregate resource demand sums each workload class’s request rate times its per-request demand.\n\nChoose one resource and unit before summing; bytes and operations cannot be added together.\n\nSymbols and units\n\nD\\_resource\n\nAggregate resource demand rate\n\nbytes/s or operations/s\n\nλₖ\n\nArrival rate of workload class k\n\nrequests/s\n\nqₖ\n\nResource demand per class-k request\n\nbytes/request or operations/request\n\nk\n\nWorkload-class index\n\ninteger index\n\nλk is requests/s for workload class k and qk is bytes or operations per request. Dresource is bytes/s or operations/s.\n\nAssumptions Linear aggregation before queueing and reuse effects; correlated bursts require an additional capacity margin.\n\n## Engineering example\n\nSynthetic mix: 80 requests/s needing 1 GB each and 20 needing 4 GB each creates 160 GB/s of un-reused traffic. Designing from the common 1 GB request alone underestimates the mix by a factor of two.\n\n## Distinctions and tradeoffs\n\nA chip specification is not just a target peak FLOPS number; it is a contract for complete workloads and their operating conditions.\n\nSpecialization improves the frequent path while narrowing flexibility; programmable fallbacks and memory headroom cost area but can extend useful life.\n\nAI models evolve on a different schedule from silicon, so a workload contract must expose assumptions that may fail before a product ships.\n\n## Specifications and interfaces\n\nAcceptance unit\n\nUseful workload at fixed quality\n\nFor example accepted requests/s under a stated latency target.\n\nRequirements horizon\n\nPresent mix plus explicit change scenarios\n\nFuture operator needs are assumptions, not known measurements.\n\n### Interfaces\n\nService and model teams Input\n\nRepresentative traces and acceptance criteria\n\nArchitecture and software teams Output\n\nResource budgets, supported operations and executable model prototypes\n\n## Check your understanding\n\n### Which requirement most directly prevents a misleading inference performance target?\n\n1.  A quality threshold and latency condition attached to throughput\n2.  The largest possible transistor count\n3.  A product name shared with a benchmark winner\n\nAnswer and reasoning\n\nA quality threshold and latency condition attached to throughput\n\nThroughput is only useful when the outputs satisfy the task and arrive under the required service conditions.\n\n## Evidence and scope\n\nWorkload mix values are synthetic. Public co-design accounts do not reveal a provider’s complete internal traces, IP contracts or acceptance thresholds.\n\nUse the manufacturing investigation to trace compatible inputs, assembly and operating constraints; no current standalone calculator models this lesson’s full mechanism.\n\n## Sources\n\n-   [Inside the hardware and co-design of MTIA](https://engineering.fb.com/2024/08/22/ml-applications/meta-mtia-hardware-co-design/)\n-   [Chiplet Actuary: A Quantitative Cost Model and Multi-Chiplet Architecture Exploration](https://arxiv.org/abs/2203.12268v4)\n\n[Published lesson data](https://semiaifoundry.com/ai-atlas/lesson-data/chip-workload-contract-5b9ec61ef1e7641d.json) · [Atlas releases and licensing](https://semiaifoundry.com/ai-atlas/licenses.html)\n\n## Continue exploring\n\n### Foundations\n\n-   [Inference serving and request scheduling](https://semiaifoundry.com/learn/atlas/serving/)\n-   [Quantization and numerical representation](https://semiaifoundry.com/learn/atlas/quantization/)\n\n### Related lessons\n\n-   [Custom inference silicon: Maia and MTIA](https://semiaifoundry.com/learn/atlas/inference-custom-silicon/)\n-   [RTL and functional verification](https://semiaifoundry.com/learn/atlas/chip-rtl-verification/)\n-   [Custom-chip volume economics](https://semiaifoundry.com/learn/atlas/chip-volume-economics/)"
}