Thermo SampleManager LIMS AI Assistant Sizing and Value Calculator
This calculator sizes a private AI assistant for Thermo SampleManager LIMS: how many GPUs the busiest hour needs, how much index space your documents take, and whether the time saved pays for the build. It uses simple, visible formulas you can check.
Your numbers
Daily active users as a percent of named users. 20 to 50 is common after the first quarter.
Includes retrieved context. Retrieval-heavy answers are often 2,000 to 6,000 tokens.
Aggregate batched throughput for your model and GPU under vLLM. Measure on your hardware; 500 to 3,000 is a wide realistic range.
Amortized hardware, power, cooling and support. Not a cloud price.
Your results
Planning estimate only. Throughput depends on model, quantization, context length and hardware; measure on your own GPUs. Time saved is an assumption you should validate with a pilot.
Net monthly value vs named users
Sweeps Named users from 20 to 600, holding your other inputs fixed.
Validate these numbers with a pilot
Optional: Netray can run a small read-only pilot on Thermo SampleManager LIMS and replace the assumptions with measured throughput and time saved. The calculator works without it.
Optional. The tool and every export work without this. No spam.
How to use it
Start with named users and the share who use the assistant on a typical day. Set questions per active user per day from a pilot or from comparable tools. Average tokens per question should include the retrieved context sent to the model, not only the typed question. Then enter the sustained tokens per second you expect per GPU, the number of documents and pages to index, minutes saved per question, and costs. The primary results are GPUs needed and net monthly value, with payback in months.
- Use a pilot to replace guesses
- Keep tokens per question realistic for retrieval
- Measure GPU throughput with your own model
- Include only costs you will actually bear
Formulas and assumptions
Daily active users = named users x active share. Questions per day = active users x questions each. Tokens per day = questions x tokens. Peak tokens per second = questions x busiest-hour share x tokens / 3,600. GPUs = ceil(peak tokens per second / tokens per second per GPU), minimum one. Index size = documents x pages x 3 chunks x (4 KB vector at 1,024 dimensions + 2 KB text). Hours saved = questions x 22 x minutes / 60. Payback = one-time cost / (monthly value - GPU cost).
Reading the results
For most deployments the GPU count is small, one or two, because peak load is bursty and answers are short compared with training workloads. The index is usually modest, a few gigabytes to a few tens of gigabytes. The sensitive input is minutes saved per question: halve it and see whether payback still works. If it does not, the project may still make sense for risk or compliance reasons, but you should say so honestly. In this context: Regulated settings: documented intended use, risk assessment and testing; assistant advisory only, no result entry, review or authorization.
Common mistakes
Sizing on average load instead of the busiest hour, ignoring retrieved context in the token count, and assuming vendor tokens-per-second figures apply to your model and prompt lengths are the usual errors. Another is counting time saved on questions people would never have asked before. Count only replaced effort, and add a safety factor for growth and for a redundant GPU.
- Average instead of peak load
- No redundancy GPU
- Optimistic minutes saved
- Forgetting re-indexing and evaluation time
Frequently Asked Questions
What tokens per second per GPU should I use?
It depends on model size, quantization, context length and batching. A range of 500 to 3,000 aggregate tokens per second on a single data-center GPU is realistic for mid-size open-weight models under vLLM. Benchmark your own setup before committing.
Why does the calculator size for the busiest hour?
Users ask questions in bursts, such as the start of a shift or before a meeting. Sizing on the daily average would leave the system slow when it matters, so the model uses a configurable busiest-hour share of daily questions.
Does the index size include the model?
No. Index size covers the vectors and text of your documents only. Model weights are separate: a 14B model at 4-bit is roughly 8 to 10 GB, and a 70B model is several times that.
Is the value figure a guaranteed saving?
No. It multiplies questions by minutes saved and hourly cost, all of which are assumptions. Validate minutes saved with a pilot and treat the result as a range, not a promise.
If you want these numbers checked against your real usage, Netray can run a short pilot on Thermo SampleManager LIMS and replace the assumptions with measurements.
Related Tools
Thermo SampleManager LIMS Private AI Layer Blueprint Generator
Generate a shareable architecture plan for a private, on-prem AI assistant on Thermo SampleManager LIMS: components, security controls, 3-phase rollout and open questions for IT.
On-Prem AIOn-Prem GPU Sizing Calculator for ERP AI
Work out how much GPU memory and how many cards you need to serve an open-weight model to your ERP users on-prem.
On-Prem AIAir-Gapped LLM Deployment Checklist
A practical control checklist for deploying and maintaining large language models in a fully air-gapped environment, from initial staging through ongoing patching and drift detection.
On-Prem AIOn-Prem AI Security Hardening Checklist
A practical control checklist for securing self-hosted language models, covering model provenance, network isolation, data governance, host hardening, and audit readiness.
More On-Prem AI tools
See all- On-Prem LLM Total Cost of Ownership Calculator
- GPU Sizing Calculator for LLM Inference
- On-Prem vs Cloud AI Cost Comparison Calculator
- Self-Hosted LLM Hardware Estimator
- Air-Gapped AI Readiness Assessment
- LLM API vs Self-Hosted Cost Calculator
- RAG Infrastructure Sizing Calculator
- Open-Weight Model Selector
- On-Prem AI ROI Calculator
Keep exploring
Free tools
- Thermo SampleManager LIMS Private AI Layer Blueprint GeneratorGenerate a shareable architecture plan for a private, on-prem AI assistant on Thermo SampleManager LIMS: components, security controls, 3-phase rollout and open questions for IT.
- LabWare LIMS AI Assistant Sizing and Value CalculatorEstimate GPUs, tokens per day, index size, hours saved and payback for a private AI assistant on LabWare LIMS.
- SAP S/4HANA On-Prem AI Assistant Sizing and Value CalculatorEstimate peak concurrent requests, GPUs, tokens per day, index size, hours saved and payback for a private AI assistant on SAP S/4HANA.