HARDWARE & WORKLOAD SIZING
GPU benchmarks
start with your workload.
Compare hardware memory, estimate model weight storage and prepare a repeatable performance evaluation.
Compare hardware specifications.
Memory and bandwidth are useful starting points. They are not application speed measurements.
GPU memory in GB.
| GPU | Memory | Memory bandwidth |
|---|---|---|
| H100 SXM | 80 GB | 3.35 TB/s |
| A100 SXM | 80 GB | 2,039 GB/s |
| RTX 4090 | 24 GB | 1,008 GB/s |
Sources: NVIDIA H100 SXM, NVIDIA A100 SXM, NVIDIA RTX 4090. Specifications checked 17 September 2026. Exact GPU variants matter.
How much space do model weights need?
A transparent estimate: parameters × bytes per parameter. Runtime memory requires additional headroom.
7 billion parameters × 2 bytes = 14 GB (decimal).
Excludes activations, KV cache, gradients, optimizer state, quantization metadata and framework overhead. This is a sizing calculation, not a GPU benchmark or a guarantee that a model will fit.
Measure what your application needs.
Use one repeatable workload and record the conditions behind every result.
| Workload | Measure | Keep consistent |
|---|---|---|
| LLM inference | Time to first token, tokens per second and tail latency | Model, precision, context, batch size and serving engine |
| Training | Time per step, throughput and peak memory | Dataset, optimizer, precision, batch size and checkpoint policy |
| Rendering | Time per frame and output quality | Scene, resolution, samples and renderer version |
Xanpath application benchmark results are not published here. These tools help you prepare a test plan; ask us about evaluating your actual workload.
Turn a hardware shortlist into a test plan.
Bring your workload. We’ll start with the questions that matter.
