All products 
Renting an NVIDIA RTX 4000 Ada server: what fits in 20 GB and when one card is enough
The RTX 4000 Ada Generation is NVIDIA's professional workstation card on the Ada Lovelace architecture, with 20 GB of GDDR6 memory and ECC error correction. It has the same FP8-capable tensor cores, third-generation RT cores and AV1-capable NVENC encoder as the bigger cards in the line, at a modest power draw. Renting an RTX 4000 Ada server makes sense when you want a current architecture and certified drivers, and the 24–48 GB of the larger cards would be more than the job needs.
Two versions of the card, and what not to confuse it with
- RTX 4000 Ada Generation: a single-slot card with an active fan and a power draw of up to 130 W.
- RTX 4000 SFF Ada: a 70 W low-profile version with the same 20 GB of memory but lower clocks. It goes into compact servers and is noticeably slower than the full-size card on long compute jobs.
- Quadro RTX 4000: an older Turing card with 8 GB of memory. The name is similar, the class is not: if a plan does not say "Ada", ask the host which model it is.
Compared with the NVIDIA L4, the RTX 4000 Ada has slightly less memory (20 GB versus 24 GB), but it has its own cooling and drivers for graphics applications, while the L4 is a passively cooled data-centre accelerator built mainly for inference and video. The L40S on the same architecture offers 48 GB and far more compute, at several times the price.
What fits in 20 GB of video memory
Rule of thumb for language models: about 2 GB per billion parameters in FP16, about 1 GB in 8-bit quantization and 0.5–0.6 GB in 4-bit. On top of that comes the KV cache, which grows with context length and the number of concurrent requests.
- 7–8B parameters in FP16: they fit, but only with a short context, since the weights alone take 14–16 GB.
- 13B in 8-bit: comfortable, with room for context and several parallel requests.
- Around 30B in 4-bit: a tight fit with a small context; a 24–48 GB card is more convenient for models this size.
- Image generation. SDXL runs on one card with room to spare. Full-precision Flux does not fit in 20 GB, so people use FP8 or quantized versions, and Ada tensor cores handle FP8 in hardware.
- Rendering and video. Blender Cycles with OptiX, Unreal Engine, CAD packages, editing and encoding to H.264, HEVC and AV1 with NVENC.
For fine-tuning, 20 GB is enough for LoRA and QLoRA on models up to 7–8B parameters. Full training of even small language models needs cards with more memory.
One card, ×2 and ×4
The catalogue has plans with one, two and four cards. The RTX 4000 Ada has neither NVLink nor MIG: cards talk only over the PCIe 4.0 bus, and each one sees its own 20 GB. A ×4 server is not 80 GB of shared memory but four separate 20 GB pools. A large model can be split across cards (tensor or pipeline parallelism in vLLM, layer splitting in llama.cpp), but the traffic over PCIe slows generation down.
Multi-card plans pay off best on independent work: four copies of a model serve parallel requests, or a render job draws different frames on different cards. In cloud plans the card is usually passed through to the virtual machine whole; if a plan lists a share of a card, check how much memory you actually get.
Drivers and CUDA
The RTX 4000 Ada has compute capability 8.9. It needs CUDA 11.8 or newer, and FP8 needs CUDA 12 with recent libraries; the R535 driver branch and later supports both versions of the card. In the image, check the nvidia-smi output and the CUDA and PyTorch versions: framework builds for older CUDA may not know this architecture. For Blender, Unreal and CAD, where certification matters, the professional RTX Enterprise or Studio driver is used, so ask which driver the host installs and whether the image includes the OpenGL and Vulkan graphics stack.
Prices, hourly and monthly billing
As of 24 October 2026, an RTX 4000 Ada GPU server in the catalogue costs roughly $350 to $2,600 a month: single-card plans with 4 vCPUs at the bottom, four-card builds with dozens of cores and almost 200 GB of RAM at the top. A typical plan comes to around $750–800 a month. Almost all plans are billed by the hour, which suits a render on a deadline, model tests and one-off generation runs, where the server lives for a few hours and is then deleted. If the card is busy more than half of every day for the whole month, compare the hourly bill with the monthly price. How to count such plans is covered in the hourly-billed VPS selection. Most hosts take payment by card; if you process personal data of EU residents, also check the data-centre country, as it decides how GDPR applies and what latency your users in Europe or the US will see.