Memory accounting
The first calculator counts one copy of all weights at a uniform precision. It does not include runtime memory. Lower precision is a storage assumption, not a promise of software compatibility or unchanged model quality.
GPU PLANNING · FORMULAS, NOT BENCHMARKS
Put numbers behind
your GPU requirements.
Start with model-weight memory and measured electricity use. Then validate the complete workload before choosing hardware.
These are planning estimates, not GPU fit tests, performance benchmarks or quotations. The example inputs are illustrative, not recommended specifications or current electricity tariffs.
Model-weight memory calculator
How much storage would the weights alone occupy if every parameter used the selected bit width?
Weights only. Not total VRAM.
Example · 7 billion parameters at 16-bit14.00 GB / 13.04 GiB
Ideal weight-storage lower bound, before runtime overhead.
Formula: bytes = parameters in billions × 1,000,000,000 × bits ÷ 8. GB = bytes ÷ 109; GiB = bytes ÷ 230.
Actual use also includes quantization metadata, any higher-precision layers, KV cache, activations and runtime workspaces. Training adds gradients and optimizer state. This estimate cannot confirm that a model fits one GPU or can be split across several GPUs.
Measured-power energy estimator
Project one system's electricity use from its average wall power while working and during the remaining hours.
System energy. Not full ownership cost.
Example · 500 W working / 100 W remaining · 8 h/day · 30 days168.00 kWh
5.60 kWh/day · 1,680.00 currency units at 10 per kWh.
Formula: daily kWh = [working W × working hours + remaining-hours W × (24 − working hours)] ÷ 1,000. Period kWh = daily kWh × days. Energy cost = period kWh × price per kWh.
GPU TDP and PSU nameplate capacity are not average wall power. Cooling, UPS losses or network equipment outside your measurement are excluded, as are demand charges, fixed fees, taxes, hardware and support. A flat rate cannot reproduce a time-of-use bill.
BEFORE YOU BUY
Turn an estimate into
an evaluation brief.
Keep these details with your test results. A calculation is a starting point; a reproducible workload is the evidence.
Discuss a configuration- Define the workload.Record the exact model revision, storage/compute precision, inference or training task, context length and expected concurrency.
- Record the complete platform.List GPU count and memory, interconnect, CPU, system RAM, storage, driver, framework and serving-engine versions.
- Test the real operating point.Use representative inputs and output lengths. Measure peak memory, latency, throughput and average wall power together; retain the command, settings and raw results.
- Review deployment and commercial terms.Check chassis, power, cooling and software compatibility, plus condition, warranty, delivery and support in the written quotation.
Methodology, sources & limits
Prepared by GPU WORLD · Reviewed . These tools apply arithmetic to your inputs; they do not query a hardware benchmark database. No GPU performance measurements are presented on this page.
Energy accounting
Power multiplied by time gives energy. The second calculator adds two daily operating states, converts Wh to kWh, then applies your rate. Use measured averages representative of each state. For multi-state schedules or changing rates, calculate each period separately.
The calculator code processes inputs in this browser without sending them to a server or storing calculation history. Normal site privacy settings still apply. Manufacturer and technical references do not establish product availability, reseller status or performance for a quoted configuration.