Compare GPUs
Specifications with their sources, measured runs where the conditions match, current offering prices and the memory fit of your workload. Accelerators are never ranked by peak FLOPS.
Specifications
vendor datasheet (RA-BIT v2 spec registry) and Kovara registry, per field| Specification | NVIDIA A100 40GB — provider does not specify PCIe/SXMGPU0089 |
|---|---|
| Memory capacity | 40GBKovara registry · verified 2026-09-25 |
| Memory type | HBM2Kovara registry · verified 2026-09-25 |
| Memory bandwidth | 1.56TB/sKovara registry · verified 2026-09-25 |
| Board power (TDP) | —Not published |
| Form factor | —Not published |
| NVLink | YesKovara registry · verified 2026-09-25 |
| MIG partitioning | YesKovara registry · verified 2026-09-25 |
Supported precisions
dense tensor peak| Dense peak | NVIDIA A100 40GB — provider does not specify PCIe/SXM |
|---|---|
| FP32 | — |
| TF32 | — |
| BF16 | — |
| FP16 | — |
| FP8 | — |
| FP4 |
Current Kovara offerings
active, public, not archived| Accelerator | Comparable price range | Live offerings | Providers | Latest check | |
|---|---|---|---|---|---|
| NVIDIA A100 40GB — …Ie/SXM | $0.69–$3.09USD / GPU-hour | 17 | 5 | 2026-11-17 | Offerings → |
Measured evidence
MeasuredRun a benchmarkNo published runs share the same conditions across these accelerators yet. Per accelerator: NVIDIA A100 40GB — provider does not specify PCIe/SXM: 0 runs.
Your workload
memory fit and decode rangeName a published model (or a parameter count), the weight precision, the context and how many sequences must run at once. RA-BIT v2 then sizes memory and an analytical decode range for each accelerator.
Share this view: permalink