All products 
NVIDIA A16 for rent: four GPUs on one board and what your machine actually gets
The NVIDIA A16 is built differently from most server cards. One PCIe board carries four separate Ampere-generation GPUs, each with its own 16 GB of GDDR6, for 64 GB per board at about 250 W. The card was designed for virtual desktops and video: one board is shared among many users, and it has enough hardware codecs for dozens of video streams. Each of the four GPUs, however, is modest in compute, well below an A10 or an L4. Renting an A16 server makes sense when the number of users and streams matters more than the speed of one large model.
What "A16" means in a plan
The main trap: "A16" in a plan name almost never means the whole 64 GB board. A virtual machine usually gets one of the four GPUs with its 16 GB, or a vGPU slice of one GPU with part of its memory. That is why video memory in the catalog starts at 2 GB: these are small vGPU profiles meant for a desktop, not for neural networks.
- vGPU slice (2 to 8 GB). A share of one GPU; the compute units are time-shared with neighbours. Good for a remote desktop, viewing drawings and light graphics.
- One full GPU (16 GB). The most a single model or application can get on an A16. The option for small-model inference and video encoding.
- Several GPUs (32, 64 or 128 GB). Two, four or eight separate processors, that is, up to two boards. There is no shared memory pool: the A16 has no NVLink and no MIG partitioning. Splitting a model across GPUs over PCIe works, but slowly. These plans pay off when you run several independent copies of a service or serve many workstations.
AI workloads: what fits into 16 GB on one GPU
Memory rule of thumb: about 2 GB per billion parameters in FP16, about 1 GB in 8-bit, 0.5–0.6 GB in 4-bit, plus the context cache (KV cache), which grows with conversation length. On one A16 GPU this means:
- 3–4B models run in FP16 with room for a long context;
- 7–8B models are comfortable in 8-bit or 4-bit; in FP16 the weights take almost all the memory and leave no room for context;
- 13–14B models fit only in 4-bit and with a short context;
- OCR, embeddings for document search, classifiers and small speech-recognition models are a comfortable load;
- Stable Diffusion image generation works, but noticeably slower than on RTX-class cards.
Memory bandwidth per GPU is low, so token generation is unhurried even on a small model. Training on an A16 makes little sense: QLoRA for a 7B model technically fits into 16 GB but takes many times longer than on an A100. For a chatbot that has to answer fast, or a model above 14B, pick a card with more memory and compute.
Virtual desktops, CAD and video
The A16's strength is VDI. With NVIDIA vGPU one board serves many remote workstations: office applications with hardware-accelerated graphics, CAD model and drawing viewers, light 3D. A GPU server with an A16 fits a team that needs remote workstations when buying a powerful computer for everyone is too expensive. For heavy rendering and complex scenes one small GPU is not enough.
For video, what matters is that each of the four GPUs has its own NVENC encoders and NVDEC decoders. A server with several GPUs transcodes many streams at once to H.264 and HEVC while barely loading the CPU. AV1 decoding is supported, but the Ampere generation has no hardware AV1 encoding.
Drivers, vGPU licence, billing and prices
- CUDA. The A16 has compute capability 8.6: you need CUDA 11.1 or newer and a driver that supports it. Current PyTorch builds and CUDA 12 containers work; BF16 and TF32 are supported, FP8 is not.
- Driver inside vGPU. A vGPU machine uses a guest driver whose version must match the driver on the provider's host. The standard driver from the NVIDIA website may not work, so ask which image or package to use.
- vGPU licence. NVIDIA vGPU software is licensed separately. The provider may or may not include the licence in the price: check before paying, and ask whether CUDA compute is allowed in your profile, since that depends on the licence type.
- Hourly or monthly. Only a small share of A16 plans in the catalog bill by the hour; the rest are monthly. Desktops and transcoding usually run non-stop, so a monthly plan is cheaper for them. Hourly is handy for a test: run your model or video pipeline, check nvidia-smi output, then decide. More in the hourly billing VPS selection.
All A16 plans in the catalog are cloud or VPS plans. Prices run from roughly $85 to $5,500 a month, with the middle of the range around $690. At the bottom are vGPU slices with a couple of vCPUs; at the top, servers with several GPUs, dozens of vCPUs and hundreds of gigabytes of RAM. If the desktops or data belong to EU users, check the data centre country for GDPR and for latency to your team or audience in the EU or US. Compare not just the total price but how much video memory and how many full GPUs you get for it.