Evidence / Cloud GPU comparisons
Compare GPUs. Check the evidence.
Choose a GPU that fits your workload, then compare the full rental configuration. These prices use the same catalog as our homepage, GPU pages and provider pages.
Which provider is cheapest? It depends on GPU variant, memory, region, instance size and pricing type. Here, “lowest listed” means the lowest matching rate among the 9 tracked providers, not the whole market. A price is not a capacity guarantee.
Source publication dates in this view: Sep 24, 2026–Sep 26, 2026. Publication is date-level; the time each price was observed is not supplied.
Catalog generated: . This is not a source or deployment timestamp.
Compare listed rental prices
Default: H100 80GB and A100 80GB, one GPU per instance, on-demand, all listed regions. Prices exclude unverified extras such as storage, egress, taxes and support. Multi-GPU rows show the entire instance bill.
Links marked affiliate may earn us money or compute credits. Commercial arrangements do not change inclusion, ordering or recommendations. Affiliate disclosure · Methodology and coverage.
1 listed offers · On-demand · 2 GPU per instance · ordered by per-GPU hourly price.
Different GPUs are workload alternatives, not performance equivalents. Match VRAM, variant, region and instance size before comparing rates. “Unspecified” means the source name does not identify PCIe or SXM.
| Provider / GPU / variant | VRAM per GPU | Pricing | Per GPU / hr | Full instance / hr | Region | Source / published date |
|---|---|---|---|---|---|---|
| DataCrunch RTX 6000 Ada · Unspecified Instance: 2RTX6000ADA.20V Link to offer | 48 GB | On-demand | $1.14 | $2.29 2 GPUs | FIN-01 | Check provider priceSource snapshot (ZIP)Published |
Manufacturer specifications, with conditions
Reviewed September 26, 2026 against NVIDIA documentation. These are hardware specifications and theoretical peaks, not NVGPU measurements or third-party benchmark results. Sparse Tensor Core peaks require supported structured sparsity; they are not expected dense training throughput.
| Reference variant | Memory | Memory bandwidth | FP16 Tensor Core peak | Conditions and source |
|---|---|---|---|---|
| H200 SXM | 141 GB HBM3e | 4.8 TB/s | Not compared here | HGX reference configuration. Confirm the rental SKU and reported memory; memory capacity alone is not a throughput result. NVIDIA source |
| B200 SXM | 180 GB HBM3e | Up to 8 TB/s | Not compared here | HGX reference configuration; do not apply it to GB200 or other Blackwell products. Verify framework support and the exact rental configuration. NVIDIA source |
| H100 SXM | 80 GB HBM3 | 3.35 TB/s | 1,979 TFLOPS with sparsity | SXM reference; do not apply its bandwidth or peak throughput to an unspecified or PCIe rental. NVIDIA source |
| H100 NVL (per GPU) | 94 GB HBM3 | 3.9 TB/s | 1,671 TFLOPS with sparsity | NVL is distinct from H100 80GB. A two-GPU NVL pair has 188 GB combined memory; that is not per-GPU VRAM. NVIDIA source |
| A100 80GB PCIe / SXM | 80 GB HBM2e | 1,935 / 2,039 GB/s | 312 dense / 624 with sparsity | PCIe and SXM have different bandwidth and host configurations. Both peak figures are manufacturer specifications. NVIDIA source |
| A100 40GB PCIe / SXM | 40 GB HBM2 | 1,555 GB/s | 312 dense / 624 with sparsity | 40GB is a separate capacity; matching arithmetic throughput does not mean a workload will fit. NVIDIA source |
| L40S | 48 GB GDDR6 ECC | 864 GB/s | 733 TFLOPS with sparsity | No NVLink. Check PCIe communication overhead for workloads split across GPUs. NVIDIA source |
| GeForce RTX 4090 | 24 GB GDDR6X | Not compared here | Not compared here | No NVLink. Memory capacity and software support constrain suitable workloads; no training-speed result is implied. NVIDIA source |
NVGPU has not run application benchmarks for these comparisons. To assess a third-party result, require the model and revision, precision, batch size, sequence lengths, GPU count and variant, software versions, interconnect, test date and methodology.
How should I choose between H100, A100, L40S and RTX 4090?
Start with measured memory use. An RTX 4090 reference card has 24 GB; L40S has 48 GB; A100 comes in 40 GB and 80 GB variants. H100 80GB and H100 NVL 94GB are distinct. We infer that more memory can enable larger models, batches or KV caches, but capacity alone does not predict speed.
For multi-GPU training, check the actual rental topology. NVIDIA lists NVLink for the H100 and A100 reference systems, while L40S and RTX 4090 do not have NVLink. The GPU name does not establish that a cloud instance connects every GPU that way. Ask for topology and benchmark communication overhead.
For inference, measure tokens per second and latency at your intended concurrency and context lengths. For training, compare time and total cost to the same quality target. Compute cost equals the whole-instance hourly rate multiplied by billed hours, including setup and idle time. A lower hourly rate can still produce a higher job cost.
Plan an AI training comparison · Cloud versus local costs · Report a correction