{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-08-14-008",
    "source_story_id": "tmp-story-tempo-moe-load-balancing",
    "edition_id": "mp-2026-08-14-morning-0036",
    "edition_url": "https://themachinepress.com/edition/2026-08-14",
    "position": 8,
    "story_type": "dispatch",
    "section": "chips-infrastructure",
    "editorial_classification": "editorial",
    "headline": "The Slowest GPU Was Solving Two Different Problems",
    "slug": "the-slowest-gpu-was-solving-two-different-problems",
    "dek": "TEMPO models expert serving as memory-bound below one token threshold and compute-bound above it, then dispatches for the actual batch regime.",
    "summary": "TEMPO models expert serving as memory-bound below one token threshold and compute-bound above it, then dispatches for the actual batch regime.",
    "body_text": "Measurements on two data-center GPU generations show that balancing token counts alone misses a regime change: small expert loads pay mainly for weight streaming, while larger loads pay for padded matrix-multiplication tiles. TEMPO solves a fixed-charge makespan problem outside the critical path and integrates with SGLang. On Qwen3-235B in the predicted win region, throughput rose four to six percent and p99 latency fell about 15.6 percent; a communication-bound DeepSeek-V3 case did not improve, matching the authors' phase diagram rather than a universal-win claim.",
    "why_it_matters": "TEMPO models expert serving as memory-bound below one token threshold and compute-bound above it, then dispatches for the actual batch regime.",
    "limitations": [
      "On Qwen3-235B in the predicted win region, throughput rose four to six percent and p99 latency fell about 15.6 percent; a communication-bound DeepSeek-V3 case did not improve, matching the authors' phase diagram rather than a universal-win claim."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-14-008/the-slowest-gpu-was-solving-two-different-problems",
    "json_url": "https://themachinepress.com/story/mp-2026-08-14-008.json",
    "first_published_at": "2026-08-14T09:00:00.000-04:00",
    "modified_at": "2026-08-14T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-08-14-008-001",
        "text": "TEMPO models expert serving as memory-bound below one token threshold and compute-bound above it, then dispatches for the actual batch regime.",
        "source_ids": [
          "source-2026-08-14-008"
        ],
        "qualification": "On Qwen3-235B in the predicted win region, throughput rose four to six percent and p99 latency fell about 15.6 percent; a communication-bound DeepSeek-V3 case did not improve, matching the authors' phase diagram rather than a universal-win claim."
      }
    ],
    "source_ids": [
      "source-2026-08-14-008"
    ],
    "tags": [
      "mixture of experts",
      "GPU serving",
      "load balancing",
      "latency"
    ],
    "image_url": "https://themachinepress.com/issues/2026-08-14/tempo-data-center-file-image.webp",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-08-14-008",
      "title": "arXiv preprint 2608.13057",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2608.13057",
      "canonical_url": "https://arxiv.org/abs/2608.13057",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-08-13T06:21:11.000-04:00",
      "accessed_at": "2026-08-14T08:27:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-08-14-008-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "The Slowest GPU Was Solving Two Different Problems",
    "publisher": "The Machine Press",
    "published_at": "2026-08-14T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-14-008/the-slowest-gpu-was-solving-two-different-problems"
  }
}