> ## Documentation Index > Fetch the complete documentation index at: https://docs.iotools.cloud/llms.txt > Use this file to discover all available pages before exploring further. # GPU VRAM Calculator > Estimate the GPU VRAM needed to run or train a transformer / large language model. Enter parameter count, precision (fp16, int8, int4…), context length and batch size to break down weights, KV cache, activations, gradients and optimizer state, and check which GPUs it fits on. [Try GPU VRAM Calculator in your browser →](https://iotools.cloud/tool/gpu-vram-calculator/) ## OpenAPI ````yaml https://api.iotools.cloud/v1/openapi post /v1/tool/gpu-vram-calculator openapi: 3.1.0 info: title: iotools.cloud API version: 1.0.0 description: >- Run any iotools.cloud tool over HTTP. Authenticate with `Authorization: Bearer iot_live_…`. **Only `POST /v1/tool/{slug}` costs credits.** Every GET here — the catalog, a tool's schema, your balance — is free. A tool call is charged its own weight or your plan's per-call minimum, whichever is larger; `x-iotools-credit-cost` on each operation is quoted at the free-tier minimum, and `GET /v1/tools/list` returns the exact figure for your key. `GET /v1/me/credits` reports what you have left, and `GET /v1/me/usage` reports what it went on. Failures are RFC 9457 problem documents — branch on `code`. servers: - url: https://api.iotools.cloud description: Production security: - bearerAuth: [] tags: - name: Catalog description: Find a tool and read its contract. Free. - name: Converters description: >- Convert between formats, encodings, and units — Base64, CSV and JSON, timestamps, and more. Fast, free, and processed right in your browser. - name: Formatters description: >- Format, beautify, minify, and validate code and data — JSON, HTML, CSS, SQL, and regex. Clean up messy input in one click, with nothing to upload. - name: Generators description: >- Generate exactly what you need and on demand — passwords, UUIDs, QR codes, hashes, random numbers, and more. Secure, instant, and free. - name: Calculators description: >- Crunch the numbers fast — from everyday math to specialized conversions and unit work. Free online calculators that run entirely in your browser. - name: Editors description: >- Edit and transform text, code, and images with quick, focused editors that run entirely in your browser — nothing to install, and no sign-up needed. - name: Utilities description: >- Everyday developer and web utilities — DNS and IP lookups, redirect and certificate checkers, and other quick diagnostics. Free and privacy-friendly. - name: Account description: Your key's allowance, limits and usage history. paths: /v1/tool/gpu-vram-calculator: post: tags: - Calculators summary: GPU VRAM Calculator description: >- Estimate the GPU VRAM needed to run or train a transformer / large language model. Enter parameter count, precision (fp16, int8, int4…), context length and batch size to break down weights, KV cache, activations, gradients and optimizer state, and check which GPUs it fits on. [Try GPU VRAM Calculator in your browser →](https://iotools.cloud/tool/gpu-vram-calculator/) operationId: run_gpu_vram_calculator requestBody: required: false content: application/json: schema: type: object properties: modelPreset: enum: - custom - gpt2-124m - llama-3-2-1b - llama-3-2-3b - mistral-7b - llama-3-8b - llama-2-13b - mixtral-8x7b - llama-3-70b - llama-3-1-405b description: Model Preset params: type: number description: Parameters (B) minimum: 0.001 maximum: 2000 hiddenDim: type: number description: Hidden Dim minimum: 64 maximum: 65536 layers: type: number description: Layers minimum: 1 maximum: 256 mode: enum: - inference - training description: Workload precision: enum: - fp32 - fp16 - int8 - int4 description: Precision batchSize: type: number description: Batch Size minimum: 1 maximum: 4096 seqLen: type: number description: Sequence Length minimum: 1 maximum: 1048576 optimizer: enum: - adam - sgd-momentum - sgd description: Optimizer mixedPrecision: type: boolean description: Mixed precision (fp32 master weights + grads) gradCheckpoint: type: boolean description: Gradient checkpointing (recompute activations) required: [] examples: mistral_7b_fp16_inference_4k_context: summary: Mistral 7B, fp16 inference, 4K context value: modelPreset: mistral-7b mode: inference precision: fp16 batchSize: '1' seqLen: '4096' 70b_model_int4_quantized_inference_4k_context: summary: 70B model, int4 quantized inference, 4K context value: modelPreset: custom params: '70' hiddenDim: '8192' layers: '80' mode: inference precision: int4 batchSize: '1' seqLen: '4096' 7b_model_mixed_precision_adam_training_2k_context: summary: 7B model, mixed-precision Adam training, 2K context value: modelPreset: custom params: '7' hiddenDim: '4096' layers: '32' mode: training precision: fp16 batchSize: '1' seqLen: '2048' optimizer: adam mixedPrecision: 'true' gradCheckpoint: '' responses: '200': description: Tool output content: application/json: examples: mistral_7b_fp16_inference_4k_context: summary: Mistral 7B, fp16 inference, 4K context value: tool: gpu-vram-calculator tool_version: 1.0.1 outputs: breakdown: - component: Model weights memory: 13.49 GB share: 78.4% - component: KV cache memory: 2.00 GB share: 11.6% - component: Activations memory: 0.16 GB share: 0.9% - component: Framework overhead memory: 1.56 GB share: 9.1% - component: Total VRAM memory: 17.21 GB share: 100% gpuFit: - gpu: RTX 4060 Ti 16GB vram: 16 GB fit: Over (108%) gpus-needed: 2 GPUs - gpu: RTX 4090 24GB vram: 24 GB fit: Fits (72%) gpus-needed: 1 GPU - gpu: RTX 5090 32GB vram: 32 GB fit: Fits (54%) gpus-needed: 1 GPU - gpu: A100 40GB vram: 40 GB fit: Fits (43%) gpus-needed: 1 GPU - gpu: A100 / H100 80GB vram: 80 GB fit: Fits (22%) gpus-needed: 1 GPU - gpu: H200 141GB vram: 141 GB fit: Fits (12%) gpus-needed: 1 GPU credits_used: 3 credits_remaining: null 70b_model_int4_quantized_inference_4k_context: summary: 70B model, int4 quantized inference, 4K context value: tool: gpu-vram-calculator tool_version: 1.0.1 outputs: breakdown: - component: Model weights memory: 32.60 GB share: 69.4% - component: KV cache memory: 10.00 GB share: 21.3% - component: Activations memory: 0.08 GB share: 0.2% - component: Framework overhead memory: 4.27 GB share: 9.1% - component: Total VRAM memory: 46.94 GB share: 100% gpuFit: - gpu: RTX 4060 Ti 16GB vram: 16 GB fit: Over (293%) gpus-needed: 3 GPUs - gpu: RTX 4090 24GB vram: 24 GB fit: Over (196%) gpus-needed: 2 GPUs - gpu: RTX 5090 32GB vram: 32 GB fit: Over (147%) gpus-needed: 2 GPUs - gpu: A100 40GB vram: 40 GB fit: Over (117%) gpus-needed: 2 GPUs - gpu: A100 / H100 80GB vram: 80 GB fit: Fits (59%) gpus-needed: 1 GPU - gpu: H200 141GB vram: 141 GB fit: Fits (33%) gpus-needed: 1 GPU credits_used: 3 credits_remaining: null 7b_model_mixed_precision_adam_training_2k_context: summary: 7B model, mixed-precision Adam training, 2K context value: tool: gpu-vram-calculator tool_version: 1.0.1 outputs: breakdown: - component: Model weights memory: 13.04 GB share: 10.6% - component: Gradients memory: 13.04 GB share: 10.6% - component: Optimizer state (Adam / AdamW) memory: 52.15 GB share: 42.2% - component: fp32 master weights memory: 26.08 GB share: 21.1% - component: Activations memory: 8.00 GB share: 6.5% - component: Framework overhead memory: 11.23 GB share: 9.1% - component: Total VRAM memory: 123.54 GB share: 100% gpuFit: - gpu: RTX 4060 Ti 16GB vram: 16 GB fit: Over (772%) gpus-needed: 8 GPUs - gpu: RTX 4090 24GB vram: 24 GB fit: Over (515%) gpus-needed: 6 GPUs - gpu: RTX 5090 32GB vram: 32 GB fit: Over (386%) gpus-needed: 4 GPUs - gpu: A100 40GB vram: 40 GB fit: Over (309%) gpus-needed: 4 GPUs - gpu: A100 / H100 80GB vram: 80 GB fit: Over (154%) gpus-needed: 2 GPUs - gpu: H200 141GB vram: 141 GB fit: Tight (88%) gpus-needed: 1 GPU credits_used: 3 credits_remaining: null schema: type: object required: - tool - tool_version - outputs - credits_used - credits_remaining properties: outputs: type: object properties: breakdown: type: array description: Memory Breakdown items: type: object properties: component: type: string description: Component memory: type: string description: Memory share: type: string description: Share x-iotools-columns: - component - memory - share gpuFit: type: array description: GPU Fit Check items: type: object properties: gpu: type: string description: GPU vram: type: string description: VRAM fit: type: string description: Fit gpus-needed: type: string description: GPUs Needed x-iotools-columns: - gpu - vram - fit - gpus-needed tool: type: string description: The tool's slug, echoing the {slug} in the request path. tool_version: type: string description: Output-contract version for this tool. credits_used: type: integer description: >- Credits this call consumed, after any settlement refund. 0 when metering is disabled. credits_remaining: type: - integer - 'null' description: >- Credits left in the current monthly allowance, or null when metering is disabled. request_id: type: string description: Correlation id, also sent as x-request-id. '400': $ref: '#/components/responses/ValidationError' '401': $ref: '#/components/responses/Unauthorized' '402': $ref: '#/components/responses/InsufficientCredits' '403': $ref: '#/components/responses/ToolNotAllowed' '404': $ref: '#/components/responses/ToolNotFound' '413': $ref: '#/components/responses/PayloadTooLarge' '429': $ref: '#/components/responses/RateLimited' '500': $ref: '#/components/responses/ToolFailed' '503': $ref: '#/components/responses/ToolDisabled' security: - bearerAuth: [] components: responses: ValidationError: description: Invalid request. content: application/problem+json: schema: type: object description: >- RFC 9457 problem document, served as application/problem+json. Branch on `code`; `title` is human prose and may be reworded without notice. Some failures add extension members — `fields` on validation errors, `retry_after` on 429s, `credits_used`/`credits_remaining` on billing-adjacent failures — documented on the responses that carry them. required: - type - title - status - code properties: type: type: string format: uri description: Stable documentation URI for this failure. examples: - https://iotools.cloud/docs/errors/validation_error title: type: string description: Short human-readable summary of the failure. examples: - Invalid request status: type: integer description: HTTP status code, matching the response's own status. code: type: string description: >- Stable machine-readable error code — branch on this, not `title`. enum: - validation_error detail: type: string description: Human explanation of this occurrence. request_id: type: string description: Correlation id, also sent as x-request-id. fields: type: - object - string additionalProperties: type: string description: >- What failed: a map of field name → message. Absent when the body itself is malformed; a single string when the failure isn't tied to one field. example: type: https://iotools.cloud/docs/errors/validation_error title: Invalid request status: 400 code: validation_error detail: One or more inputs are invalid — see `fields`. request_id: e4042b29-8f1e-4c7a-9b52-6f0d1a3c7e11 fields: inputString: Required Unauthorized: description: Missing or invalid API key. content: application/problem+json: schema: type: object description: >- RFC 9457 problem document, served as application/problem+json. Branch on `code`; `title` is human prose and may be reworded without notice. Some failures add extension members — `fields` on validation errors, `retry_after` on 429s, `credits_used`/`credits_remaining` on billing-adjacent failures — documented on the responses that carry them. required: - type - title - status - code properties: type: type: string format: uri description: Stable documentation URI for this failure. examples: - https://iotools.cloud/docs/errors/invalid_api_key title: type: string description: Short human-readable summary of the failure. examples: - Invalid API key status: type: integer description: HTTP status code, matching the response's own status. code: type: string description: >- Stable machine-readable error code — branch on this, not `title`. enum: - invalid_api_key detail: type: string description: Human explanation of this occurrence. request_id: type: string description: Correlation id, also sent as x-request-id. example: type: https://iotools.cloud/docs/errors/invalid_api_key title: Invalid API key status: 401 code: invalid_api_key detail: 'Provide ''Authorization: Bearer ''.' request_id: e4042b29-8f1e-4c7a-9b52-6f0d1a3c7e11 InsufficientCredits: description: Monthly credit allowance exhausted. content: application/problem+json: schema: type: object description: >- RFC 9457 problem document, served as application/problem+json. Branch on `code`; `title` is human prose and may be reworded without notice. Some failures add extension members — `fields` on validation errors, `retry_after` on 429s, `credits_used`/`credits_remaining` on billing-adjacent failures — documented on the responses that carry them. required: - type - title - status - code properties: type: type: string format: uri description: Stable documentation URI for this failure. examples: - https://iotools.cloud/docs/errors/insufficient_credits title: type: string description: Short human-readable summary of the failure. examples: - Insufficient credits status: type: integer description: HTTP status code, matching the response's own status. code: type: string description: >- Stable machine-readable error code — branch on this, not `title`. enum: - insufficient_credits detail: type: string description: Human explanation of this occurrence. request_id: type: string description: Correlation id, also sent as x-request-id. credits_used: type: integer description: Always 0 — a refused call charges nothing. credits_remaining: type: integer description: Credits left in the allowance — fewer than this call costs. example: type: https://iotools.cloud/docs/errors/insufficient_credits title: Insufficient credits status: 402 code: insufficient_credits detail: This call costs 1 credit and 0 remain in this month's allowance. request_id: e4042b29-8f1e-4c7a-9b52-6f0d1a3c7e11 credits_used: 0 credits_remaining: 0 ToolNotAllowed: description: Tool exists but has no API surface. content: application/problem+json: schema: type: object description: >- RFC 9457 problem document, served as application/problem+json. Branch on `code`; `title` is human prose and may be reworded without notice. Some failures add extension members — `fields` on validation errors, `retry_after` on 429s, `credits_used`/`credits_remaining` on billing-adjacent failures — documented on the responses that carry them. required: - type - title - status - code properties: type: type: string format: uri description: Stable documentation URI for this failure. examples: - https://iotools.cloud/docs/errors/tool_not_allowed title: type: string description: Short human-readable summary of the failure. examples: - Tool not available over the API status: type: integer description: HTTP status code, matching the response's own status. code: type: string description: >- Stable machine-readable error code — branch on this, not `title`. enum: - tool_not_allowed detail: type: string description: Human explanation of this occurrence. request_id: type: string description: Correlation id, also sent as x-request-id. example: type: https://iotools.cloud/docs/errors/tool_not_allowed title: Tool not available over the API status: 403 code: tool_not_allowed detail: >- "Background Remover" is available on iotools.cloud but has no API endpoint. request_id: e4042b29-8f1e-4c7a-9b52-6f0d1a3c7e11 ToolNotFound: description: No such tool. content: application/problem+json: schema: type: object description: >- RFC 9457 problem document, served as application/problem+json. Branch on `code`; `title` is human prose and may be reworded without notice. Some failures add extension members — `fields` on validation errors, `retry_after` on 429s, `credits_used`/`credits_remaining` on billing-adjacent failures — documented on the responses that carry them. required: - type - title - status - code properties: type: type: string format: uri description: Stable documentation URI for this failure. examples: - https://iotools.cloud/docs/errors/tool_not_found title: type: string description: Short human-readable summary of the failure. examples: - Tool not found status: type: integer description: HTTP status code, matching the response's own status. code: type: string description: >- Stable machine-readable error code — branch on this, not `title`. enum: - tool_not_found detail: type: string description: Human explanation of this occurrence. request_id: type: string description: Correlation id, also sent as x-request-id. example: type: https://iotools.cloud/docs/errors/tool_not_found title: Tool not found status: 404 code: tool_not_found detail: No tool with that slug. See GET /v1/tools/list. request_id: e4042b29-8f1e-4c7a-9b52-6f0d1a3c7e11 PayloadTooLarge: description: Body too large. content: application/problem+json: schema: type: object description: >- RFC 9457 problem document, served as application/problem+json. Branch on `code`; `title` is human prose and may be reworded without notice. Some failures add extension members — `fields` on validation errors, `retry_after` on 429s, `credits_used`/`credits_remaining` on billing-adjacent failures — documented on the responses that carry them. required: - type - title - status - code properties: type: type: string format: uri description: Stable documentation URI for this failure. examples: - https://iotools.cloud/docs/errors/payload_too_large title: type: string description: Short human-readable summary of the failure. examples: - Payload too large status: type: integer description: HTTP status code, matching the response's own status. code: type: string description: >- Stable machine-readable error code — branch on this, not `title`. enum: - payload_too_large detail: type: string description: Human explanation of this occurrence. request_id: type: string description: Correlation id, also sent as x-request-id. example: type: https://iotools.cloud/docs/errors/payload_too_large title: Payload too large status: 413 code: payload_too_large detail: Request body exceeds this tool's size limit. request_id: e4042b29-8f1e-4c7a-9b52-6f0d1a3c7e11 RateLimited: description: Per-minute rate limit exceeded. headers: Retry-After: description: Seconds to wait before retrying (RFC 9110 delta-seconds). schema: type: integer content: application/problem+json: schema: type: object description: >- RFC 9457 problem document, served as application/problem+json. Branch on `code`; `title` is human prose and may be reworded without notice. Some failures add extension members — `fields` on validation errors, `retry_after` on 429s, `credits_used`/`credits_remaining` on billing-adjacent failures — documented on the responses that carry them. required: - type - title - status - code properties: type: type: string format: uri description: Stable documentation URI for this failure. examples: - https://iotools.cloud/docs/errors/rate_limited title: type: string description: Short human-readable summary of the failure. examples: - Rate limit exceeded status: type: integer description: HTTP status code, matching the response's own status. code: type: string description: >- Stable machine-readable error code — branch on this, not `title`. enum: - rate_limited detail: type: string description: Human explanation of this occurrence. request_id: type: string description: Correlation id, also sent as x-request-id. retry_after: type: integer description: >- Seconds until the window resets — the same value as the `Retry-After` header. example: type: https://iotools.cloud/docs/errors/rate_limited title: Rate limit exceeded status: 429 code: rate_limited detail: Too many requests. Retry in 30s. request_id: e4042b29-8f1e-4c7a-9b52-6f0d1a3c7e11 retry_after: 30 ToolFailed: description: Tool failed to run. content: application/problem+json: schema: type: object description: >- RFC 9457 problem document, served as application/problem+json. Branch on `code`; `title` is human prose and may be reworded without notice. Some failures add extension members — `fields` on validation errors, `retry_after` on 429s, `credits_used`/`credits_remaining` on billing-adjacent failures — documented on the responses that carry them. required: - type - title - status - code properties: type: type: string format: uri description: Stable documentation URI for this failure. examples: - https://iotools.cloud/docs/errors/processing_error - https://iotools.cloud/docs/errors/internal_error title: type: string description: Short human-readable summary of the failure. examples: - Tool failed to run - Internal error status: type: integer description: HTTP status code, matching the response's own status. code: type: string description: >- Stable machine-readable error code — branch on this, not `title`. enum: - processing_error - internal_error detail: type: string description: Human explanation of this occurrence. request_id: type: string description: Correlation id, also sent as x-request-id. credits_used: type: integer description: >- Always 0 on `processing_error` — a failed run is refunded, floor included. credits_remaining: type: integer description: >- Credits left after the refund. Absent when metering is disabled. example: type: https://iotools.cloud/docs/errors/processing_error title: Tool failed to run status: 500 code: processing_error detail: The tool failed to run. Please try again. request_id: e4042b29-8f1e-4c7a-9b52-6f0d1a3c7e11 credits_used: 0 ToolDisabled: description: Tool temporarily disabled. content: application/problem+json: schema: type: object description: >- RFC 9457 problem document, served as application/problem+json. Branch on `code`; `title` is human prose and may be reworded without notice. Some failures add extension members — `fields` on validation errors, `retry_after` on 429s, `credits_used`/`credits_remaining` on billing-adjacent failures — documented on the responses that carry them. required: - type - title - status - code properties: type: type: string format: uri description: Stable documentation URI for this failure. examples: - https://iotools.cloud/docs/errors/tool_disabled - https://iotools.cloud/docs/errors/api_unconfigured title: type: string description: Short human-readable summary of the failure. examples: - Tool temporarily disabled - API not configured status: type: integer description: HTTP status code, matching the response's own status. code: type: string description: >- Stable machine-readable error code — branch on this, not `title`. enum: - tool_disabled - api_unconfigured detail: type: string description: Human explanation of this occurrence. request_id: type: string description: Correlation id, also sent as x-request-id. example: type: https://iotools.cloud/docs/errors/tool_disabled title: Tool temporarily disabled status: 503 code: tool_disabled detail: This tool is temporarily unavailable. Try again shortly. request_id: e4042b29-8f1e-4c7a-9b52-6f0d1a3c7e11 securitySchemes: bearerAuth: type: http scheme: bearer bearerFormat: iot_live_… ````