API Overview

AccelLab — Accelerator Emulation Lab API lets you emulate co-processor behavior before hardware exists. Model latency distributions, queueing effects, and throughput ceilings for GPU, NPU, FPGA, and Quantum accelerator profiles — all from a single API.

Whether you're planning silicon investments, benchmarking deployment targets, or stress-testing inference pipelines, AccelLab gives you production-grade simulation with explainable, reproducible results and zero hardware required.

Key Capabilities

Latency Simulation — p50 / p95 / p99 distributions with M/M/1 queueing model
Throughput Ceilings — ops/s estimates under realistic concurrency and saturation
Profile Comparison — head-to-head verdicts with delta breakdowns
Compatibility Checks — validate workloads against latency and throughput requirements
Scenario Packs — 15+ pre-built workloads (LLM inference, CV pipelines, quantum circuits, stress tests)
Deterministic Seeding — reproducible results across runs
Result Caching — identical requests return cached results instantly

Authentication

Authentication is handled automatically by RapidAPI. Subscribe to AccelLab, and all required headers (x-rapidapi-key, x-rapidapi-host) are injected by the platform. No manual key management is needed.

Pricing & Tier Breakdown

Feature	BASIC (Free)	PRO ($29/mo)	ULTRA ($149/mo)	MEGA ($499/mo)
Monthly Quota	250 requests	10,000 requests	100,000 requests	500,000 requests
Overage	$0.001/request	$0.001/request	$0.001/request	$0.001/request
All Endpoints	Yes	Yes	Yes	Yes
All Profiles & Packs	Yes	Yes	Yes	Yes
Result Caching	Yes	Yes	Yes	Yes
Best For	Evaluation & prototyping	Active development & CI pipelines	Production workloads & teams	Enterprise-scale & high-volume

Tip: Every tier has access to all endpoints, profiles, and packs. Tiers differ only in quota volume. Start with Basic (free) to evaluate, then scale as needed.

Endpoints

1. GET /health

Health Check — Returns service status, version, and timestamp. No authentication required.

Example Response:

{
  "status": "ok",
  "version": "1.0.0",
  "timestamp": "2026-02-20T21:00:00.000Z"
}

2. GET /profiles

List Accelerator Profiles — Returns all available accelerator profiles with summary metadata. Use these IDs in simulation and comparison requests.

Example Response:

[
  {
    "id": "gpu-server-a",
    "name": "GPU-Server-A",
    "category": "GPU",
    "description": "High-end GPU server for training and inference",
    "version": "1.0.0",
    "updated_at": "2025-06-26T00:00:00+00:00"
  },
  {
    "id": "npu-edge-1",
    "name": "NPU-Edge-1",
    "category": "NPU",
    "description": "Edge NPU optimized for low-power inference",
    "version": "1.0.0",
    "updated_at": "2025-06-26T00:00:00+00:00"
  },
  {
    "id": "fpga-streaming",
    "name": "FPGA-Streaming",
    "category": "FPGA",
    "description": "Streaming-optimized FPGA for signal processing",
    "version": "1.0.0",
    "updated_at": "2025-06-26T00:00:00+00:00"
  },
  {
    "id": "quantum-sim-1",
    "name": "Quantum-Sim-1",
    "category": "Quantum",
    "description": "Quantum circuit simulator",
    "version": "1.0.0",
    "updated_at": "2025-06-26T00:00:00+00:00"
  }
]

Available Profile IDs: gpu-server-a, gpu-server-b, npu-edge-1, npu-datacenter, fpga-streaming, fpga-batch, quantum-sim-1, quantum-hybrid

3. GET /profiles/{id}

Get Profile Detail — Returns the full profile including internal model parameters (base service times, throughput ceiling, concurrency limits, tail amplification).

Example: GET /profiles/gpu-server-a

{
  "id": "gpu-server-a",
  "name": "GPU-Server-A",
  "category": "GPU",
  "description": "High-end GPU server for training and inference",
  "version": "1.0.0",
  "model_json": {
    "base_service_ms": {
      "matmul": 2.1,
      "conv2d": 3.4,
      "attention": 4.8,
      "elementwise": 0.3,
      "reduce": 0.8,
      "fft": 1.9
    },
    "throughput_ceiling_ops_s": 15000,
    "memory_bandwidth_gb_s": 900,
    "max_concurrency": 64,
    "tail_amplification_factor": 1.4
  }
}

4. GET /packs

List Scenario Packs — Returns all pre-built workload scenario packs. Packs contain realistic operation mixes for common ML/AI use cases.

Example Response:

[
  {
    "name": "llm-inference-basic",
    "version": "1.0.0",
    "description": "Basic LLM inference workload with attention and matmul ops",
    "tags": ["llm", "inference", "basic"],
    "updated_at": "2025-06-26T00:00:00+00:00"
  },
  {
    "name": "cv-realtime",
    "version": "1.0.0",
    "description": "Real-time computer vision pipeline with conv2d and reduce ops",
    "tags": ["cv", "realtime", "edge"],
    "updated_at": "2025-06-26T00:00:00+00:00"
  },
  {
    "name": "quantum-variational",
    "version": "1.0.0",
    "description": "Variational quantum eigensolver simulation workload",
    "tags": ["quantum", "vqe", "optimization"],
    "updated_at": "2025-06-26T00:00:00+00:00"
  }
]

Available Packs: llm-inference-basic, llm-inference-heavy, cv-batch-training, cv-realtime, edge-anomaly-detection, latency-sensitive-api, mixed-precision-training, multi-model-ensemble, nlp-tokenization, quantum-sampling, quantum-variational, recommendation-engine, signal-processing, stress-test-large-batch, stress-test-max-qps

5. GET /packs/{name}

Get Pack Detail — Returns the full pack definition including the operation list and an example usage payload you can send directly to /simulate.

Example: GET /packs/llm-inference-basic

{
  "name": "llm-inference-basic",
  "version": "1.0.0",
  "description": "Basic LLM inference workload with attention and matmul ops",
  "tags": ["llm", "inference", "basic"],
  "pack_json": {
    "ops": [
      { "op": "attention", "shape": [1, 512, 768], "dtype": "fp16", "count": 12 },
      { "op": "matmul", "shape": [768, 768], "dtype": "fp16", "count": 12 },
      { "op": "elementwise", "shape": [1, 512, 768], "dtype": "fp16", "count": 24 },
      { "op": "reduce", "shape": [1, 512, 768], "dtype": "fp32", "count": 6 }
    ],
    "concurrency": 4,
    "qps_target": 200
  },
  "example_usage": {
    "profile_id": "gpu-server-a",
    "workload": {
      "ops": [
        { "op": "attention", "shape": [1, 512, 768], "dtype": "fp16", "count": 12 },
        { "op": "matmul", "shape": [768, 768], "dtype": "fp16", "count": 12 }
      ],
      "concurrency": 4,
      "qps_target": 200
    }
  }
}

6. POST /simulate

Run Simulation — The core engine. Submit a workload against an accelerator profile and receive latency percentiles (p50/p95/p99), throughput estimates, risk signals, and an explainable trace showing exactly how each number was computed.

Example Request:

{
  "profile_id": "gpu-server-a",
  "workload": {
    "ops": [
      { "op": "matmul", "shape": [1024, 1024], "dtype": "fp16", "count": 10, "batch": 1 },
      { "op": "attention", "shape": [1, 512, 768], "dtype": "fp16", "count": 12 },
      { "op": "elementwise", "shape": [1, 512, 768], "dtype": "fp32", "count": 8 }
    ],
    "concurrency": 4,
    "qps_target": 500,
    "warm_cache": true
  },
  "mode": "standard",
  "seed": "demo-run-1"
}

Example Response:

{
  "run_id": "run_a1b2c3d4e5f6",
  "profile_id": "gpu-server-a",
  "cached": false,
  "summary": {
    "p50_ms": 1.87,
    "p95_ms": 5.23,
    "p99_ms": 8.41,
    "throughput_ops_s": 938,
    "estimated_cost_index": 0.12
  },
  "signals": {
    "queueing_risk": "low",
    "tail_risk": "low",
    "saturation": false
  },
  "trace": {
    "steps": [
      "Base service time for matmul@fp16 on GPU-Server-A: 2.1ms (batch factor: 1.00 → 2.10ms)",
      "Base service time for attention@fp16 on GPU-Server-A: 4.8ms (batch factor: 1.00 → 4.80ms)",
      "Base service time for elementwise@fp32 on GPU-Server-A: 0.3ms (batch factor: 1.00 → 0.30ms)",
      "Concurrency factor (4 workers): utilization ρ=0.145",
      "M/M/1 queueing delay: 0.13ms (ρ < 0.7, stable)",
      "Jitter applied (seed=demo-run-1): ±0.02ms",
      "Tail inflation at p99: 1.0x base (saturation headroom ok)"
    ],
    "notes": [
      "Cache warm: reduced first-op penalty",
      "Profile GPU-Server-A supports matmul, conv2d, attention, elementwise, reduce, fft"
    ]
  }
}

Supported Operations

Op	Description	Supported dtypes
`matmul`	Matrix multiplication	fp16, fp32, fp64, int8
`conv2d`	2D convolution	fp16, fp32, fp64, int8
`attention`	Self/cross-attention	fp16, fp32, fp64, int8
`elementwise`	Element-wise ops (ReLU, add, etc.)	fp16, fp32, fp64, int8
`reduce`	Reduction (sum, mean, max)	fp16, fp32, fp64, int8
`fft`	Fast Fourier Transform	fp16, fp32, fp64, int8

7. POST /compare

Compare Two Profiles or Runs — Submit two workload/profile combinations (or existing run IDs) and get a head-to-head verdict with delta breakdowns and plain-English explanations.

Example Request:

{
  "a": {
    "profile_id": "gpu-server-a",
    "workload": {
      "ops": [
        { "op": "matmul", "shape": [1024, 1024], "dtype": "fp16", "count": 10 },
        { "op": "attention", "shape": [1, 512, 768], "dtype": "fp16", "count": 12 }
      ],
      "concurrency": 4,
      "qps_target": 500
    }
  },
  "b": {
    "profile_id": "npu-datacenter",
    "workload": {
      "ops": [
        { "op": "matmul", "shape": [1024, 1024], "dtype": "fp16", "count": 10 },
        { "op": "attention", "shape": [1, 512, 768], "dtype": "fp16", "count": 12 }
      ],
      "concurrency": 4,
      "qps_target": 500
    }
  }
}

Example Response:

{
  "request_id": "c3d4e5f6-a1b2-...",
  "verdict": "a_better",
  "deltas": {
    "p50_ms": -0.85,
    "p95_ms": -1.92,
    "p99_ms": -3.14,
    "throughput_ops_s": 220
  },
  "explain": [
    "A is 0.85ms faster at p50",
    "A has 3.14ms lower p99 tail latency",
    "A has 220 ops/s higher throughput"
  ],
  "results": {
    "a": { "p50_ms": 1.87, "p95_ms": 5.23, "p99_ms": 8.41, "throughput_ops_s": 938, "estimated_cost_index": 0.12 },
    "b": { "p50_ms": 2.72, "p95_ms": 7.15, "p99_ms": 11.55, "throughput_ops_s": 718, "estimated_cost_index": 0.18 }
  }
}

8. POST /compatibility-check

Check Compatibility — Validate whether an accelerator profile can meet your latency, throughput, and cost requirements. Returns pass/fail with specific gaps and actionable suggestions.

Example Request:

{
  "profile_id": "npu-edge-1",
  "workload": {
    "ops": [
      { "op": "conv2d", "shape": [1, 224, 224, 3], "dtype": "int8", "count": 5 },
      { "op": "reduce", "shape": [1, 1000], "dtype": "fp32", "count": 1 }
    ],
    "concurrency": 1,
    "qps_target": 30
  },
  "requirements": {
    "max_p95_ms": 50,
    "max_p99_ms": 100,
    "min_throughput_ops_s": 20,
    "max_cost_index": 0.5
  }
}

Example Response:

{
  "request_id": "d4e5f6a1-b2c3-...",
  "profile_id": "npu-edge-1",
  "compatible": true,
  "gaps": [],
  "suggestions": ["Profile meets all requirements with headroom"],
  "simulation_summary": {
    "p50_ms": 8.2,
    "p95_ms": 22.1,
    "p99_ms": 38.5,
    "throughput_ops_s": 30,
    "estimated_cost_index": 0.08
  }
}

Response Headers

All authenticated responses include standard rate-limit headers:

Header	Description
`X-RateLimit-Limit`	Maximum requests allowed per window
`X-RateLimit-Remaining`	Requests remaining in the current window
`X-RateLimit-Reset`	Seconds until the window resets
`X-Request-Id`	Unique request identifier for debugging/support

Error Responses

All errors follow a consistent JSON structure:

{
  "error": {
    "code": "VALIDATION_ERROR",
    "message": "Required: profile_id, workload.ops[]",
    "request_id": "a1b2c3d4-..."
  }
}

Status	Code	Description
400	VALIDATION_ERROR	Missing or invalid request parameters
403	FORBIDDEN	Unauthorized request (missing subscription)
404	NOT_FOUND	Profile or pack not found
405	METHOD_NOT_ALLOWED	Wrong HTTP method (e.g., GET on a POST endpoint)
429	RATE_LIMITED	Rate limit or quota exceeded
500	INTERNAL_ERROR	Server-side error

Use Cases

Hardware Evaluation & Silicon Planning

Compare GPU, NPU, FPGA, and Quantum profiles against your target workloads before committing to hardware purchases or silicon tape-out. Use /compare to get quantified head-to-head verdicts.

ML Pipeline Optimization

Test your inference or training workloads across different accelerator profiles to identify the optimal deployment target. Use scenario packs like llm-inference-heavy or mixed-precision-training for realistic modeling.

CI/CD Integration Testing

Embed /compatibility-check into your CI pipeline to ensure new model architectures still meet your latency SLAs on the target hardware before deploying to production.

Edge Deployment Planning

Validate that edge-optimized profiles (npu-edge-1) can handle your workload under real-world constraints using packs like edge-anomaly-detection or cv-realtime.

Capacity Planning & Stress Testing

Use stress packs (stress-test-max-qps, stress-test-large-batch) to find saturation points and plan capacity. The signals field warns when you're approaching dangerous utilization levels.

Quantum Readiness Assessment

Explore quantum accelerator profiles (quantum-sim-1, quantum-hybrid) with quantum-specific packs to understand what quantum advantage might look like for your workloads.

Full Documentation

Complete API documentation, interactive playground, and accelerator profile catalog available at:

https://accellab.org

Contact

Name	AccelLab
Email	accellab@fastmail.com
Website	https://accellab.org