A100 40GB vs RTX 5090
Both run on the same nodes, the same templates and the same tiers here — so the choice comes down to memory, bandwidth and price. Every figure below is from the catalogue or the vendor datasheet.
- A100 40GB
- $0.09/h40 GB · Ampere · 32 available
- RTX 5090
- $0.09/h32 GB · Blackwell · 128 available
- Price gap
- NoneOn spot, per GPU-hour
- Over 100 h
- $0.00Difference for the same GPU-hours
What actually separates them.
Four differences that change a decision, computed from the numbers rather than asserted.
Bigger models fit on the A100 40GB
40 GB against 32 GB. At FP16 that is about 16B parameters on one GPU, against 12B.
Token generation is faster on the RTX 5090
1,792 GB/s against 1,555 GB/s. Inference reads the whole model from memory for every token, so bandwidth, not FLOPS, usually sets the ceiling.
More memory per dollar on the A100 40GB
444 GB per dollar per hour against 356 GB. The usual tie-breaker when both cards fit the job.
Every figure we hold on both cards.
A row is marked only when both values exist and one is genuinely better. A100 40GB leads on 4 rows, RTX 5090 on 5.
| A100 40GB | RTX 5090 | |
|---|---|---|
| GPU memoryDecides which models fit at all | 40 GB HBM2 | 32 GB GDDR7 |
| Memory bandwidthSets token throughput on inference | 1,555 GB/s | 1,792 GB/s |
| FP32 | 20 TFLOPS | 105 TFLOPS |
| 8-bit float (FP8)Halves the memory a model needs, where the stack supports it | Not supported | Supported |
| Architecture | Ampere (2020) | Blackwell (2025) |
| Board powerLower is cheaper to run at scale — for us, not for your bill | 250 W | 575 W |
| GPU-to-GPU linkOnly matters for multi-GPU training | PCIe only | PCIe only |
| Node sizes | 1×, 2×, 4×, 8× | 1×, 2×, 4× |
| Spot price | $0.09/h | $0.09/h |
| On-demand price | $0.17/h | $0.19/h |
| Memory per dollarGPU memory per $1 of spot time, per hour | 444 GB | 356 GB |
| 30-day reclaim rateShare of spot instances reclaimed | <5% | 10–15% |
| Available now | 32 GPUs | 128 GPUs |
Specifications come from each vendor's datasheet — linked on the A100 40GB page and the RTX 5090 page. Prices are ours, per GPU-hour, billed per minute.
The same models, on one GPU of each.
Weights only, with 20% of the memory kept free for activations and the KV cache. A cross means one GPU is not enough — the model page has the multi-GPU arithmetic.
| Model | Precision | A100 40GB 40 GB | RTX 5090 32 GB |
|---|---|---|---|
| Llama 3.1 8B8B parameters | FP16 · 16 GB | Fits | Fits |
| INT4 · 4 GB | Fits | Fits | |
| Qwen2.5 32B32B parameters | FP16 · 64 GB | Needs 2× node | Needs 4× node |
| INT4 · 16 GB | Fits | Fits | |
| Llama 3.3 70B70B parameters | FP16 · 140 GB | Needs 8× node | Needs more than one node |
| INT4 · 35 GB | Needs 2× node | Needs 2× node | |
| Mixtral 8x22B141B parameters | FP16 · 282 GB | Needs more than one node | Needs more than one node |
| INT4 · 71 GB | Needs 4× node | Needs 4× node |
The same 100 GPU-hours on each.
Compute only, billed per minute. Storage is $0.08 per GB-month on either card.
| A100 40GB | RTX 5090 | |
|---|---|---|
| One hour, spot | $0.09 | $0.09 |
| One day, spot | $2.16 | $2.16 |
| 100 GPU-hours, spot | $9.00 | $9.00 |
| 100 GPU-hours, on-demand | $17.00 | $19.00 |
| A month, reserved | $65.70 | $65.70 |
Take the A100 40GB when…
- Fine-tuning with LoRA and QLoRA — Comfortable for parameter-efficient training on 13B–34B models without sharding.
- Inference servers — Holds 30B-class models quantised, or several smaller models side by side.
- Partitioned serving (MIG) — The GPU can be split into isolated instances, each with its own memory and compute slice.
Take the RTX 5090 when…
- Fine-tuning small models — LoRA on 7B–13B models, and full fine-tuning below 3B.
- Image and video generation — Diffusion pipelines run entirely in memory at this capacity.
- Anything that can checkpoint — On the spot tier this model is 53% below its own on-demand price, with a 2-minute notice before a reclaim.
A100 40GB or RTX 5090?
Which is cheaper, the A100 40GB or the RTX 5090?
The RTX 5090, at $0.09 per GPU-hour on spot against $0.09 — 0% less. On-demand: $0.19 against $0.17.
Which one should I take for inference?
Take the one that holds the model with room for the KV cache, then the one with more memory bandwidth. Here that is the A100 40GB on capacity (40 GB) and the RTX 5090 on bandwidth.
Can I run both on the same job?
Not inside one instance — a node holds GPUs of a single model. You can run two instances in parallel, on the same persistent volume in turn, or split the work: a big card for the model that needs capacity, a cheap one for the parts that do not.
Other comparisons: the full list · all 47 models
Try both for less than the price of deciding.
An hour on each costs $0.18 on spot. Nothing renews, and your disk moves between them.
Billed per minute from the moment the instance is reachable. Minimum credit $40, no subscription.