Low-power professional card for small-form-factor workstations and dense inference nodes.
Weekly median price per GPU per hour
By provider shows one card per company, placed where its best offer ranked. By configuration lists every offer.
What you can rent comes first: in stock, then waitlist, then not reported, then out of stock. Priced offers rank ahead of quote-only, and on-demand ahead of other billing types.
Within a group, five factors set the order:
We match the provider, its country, GPU model, form factor, billing type, availability and instance name. Matching is partial and case-insensitive. Everyday words work too, so "interruptible" finds spot and "sold out" finds out of stock.
Heads up: A provider's own page may quote a different figure, for example on tax, region or a promotion. A monthly-only plan shows a derived hourly rate. Verify before provisioning. More on how we price.
One RTX 4000 SFF Ada has 20 GB of VRAM. In practice, that's enough memory for roughly 22B parameters at 4-bit or 4B at 16-bit, assuming a 32K context. Below are some open-weight LLMs, with the estimated memory and GPUs each one needs.
| Model | Memory (INT4 / FP4) | RTX 4000 SFF Adas needed | Cost /hr | Cost /mo |
|---|---|---|---|---|
Qwen3.8-27B
27B
|
18 GB
|
1
|
–
|
–
|
Gemma 4 31B
30.7B
|
22 GB
|
2
|
–
|
–
|
GPT-OSS-120B
117B · 5.1B active
|
67 GB
|
4
|
–
|
–
|
GLM-5.3-Flash
320B · 18B active
|
178 GB
|
10
|
–
|
–
|
MiniMax-M3
428B · 23B active
|
239 GB
|
14
|
–
|
–
|
DeepSeek V4.1 Flash
552B · 16B active
|
306 GB
|
17
|
–
|
–
|
Kimi K3
2.8T · 104B active
|
1,544 GB
|
86
|
–
|
–
|
Estimates based on the median on-demand rate. Memory is weights plus FP8 KV cache at 32K context per request (FP16/BF16 in the 16-bit column). GPU counts assume 90% of advertised VRAM is usable. No guarantee of runtime support, usable performance, or that a matching quantized build exists. How we estimate costs.
Nvidia RTX 4000 SFF Ada · Per GPU
| Compute · dense | |
|---|---|
| FP8 | 153.4 TFLOPS 306.8 with sparsity |
| FP16 / BF16 | 76.7 TFLOPS 153.4 with sparsity |
| FP32 | 19.2 TFLOPS |
| Precision support | FP8FP16BF16TF32FP32INT4INT8 |
| Memory | |
|---|---|
| Capacity | 20 GB GDDR6 |
| Bandwidth | 280 GB/s |
| Bus width | 160-bit |
| ECC | Yes |
| Silicon | |
|---|---|
| Architecture | Ada Lovelace |
| Process | TSMC 4N |
| Transistors | 35.8 billion |
| Shader cores | 6,144 CUDA cores |
| Matrix cores | 192 Tensor cores |
| Compute units | 48 SMs |
| Fabric and host | |
|---|---|
| Host interface | PCIe 4.0 x16 |
| Power | |
|---|---|
| Board power | 70 W |
| Cooling | Active |
As of September 15, 2026, we track 3 configs from 3 providers. Prices are per GPU per hour.
| Billing type | Configs | Cheapest |
|---|---|---|
On-demand | 3 | $0.43 |
No median for on-demand: we only show this when at least 3 providers list the GPU on that billing type.
The cheapest verified in-stock estimate is $360 per month on-demand.
GetDeploying currently tracks RTX 4000 SFF Ada configs from 3 providers. Koyeb has verified in-stock on-demand configs. See the full price comparison above for every provider and config.
One RTX 4000 SFF Ada runs models up to roughly 22B parameters at 4-bit quantization or 4B at 16-bit, assuming a 32K context. Larger models run across multiple GPUs: the model table above shows the estimated memory and GPU count for popular open-weight LLMs.
20GB GDDR6 with ECC inside a 70W board power envelope, so it runs in small-form-factor workstations and dense chassis that cannot supply or cool a full-height card. Ada Lovelace FP8 Tensor Cores cover small-model inference, AI development, rendering and viewport work.
The 70W envelope costs throughput: single-precision performance is roughly 30% below the full-height RTX 4000 Ada, which shares the same 20GB and pin count. There is no NVLink to pool memory across cards. For continuous cloud inference at a similar power draw, the L4 offers 24GB and higher memory bandwidth.
Last updated
