All products 
Renting a server with an NVIDIA A10: what fits in 24 GB and how to tell a slice from a full card
The NVIDIA A10 is an Ampere-generation card with 24 GB of GDDR6 memory and a 150 W power draw. It takes a single PCIe 4.0 slot, so cloud platforms pack it densely and like to share it between virtual machines. The A10 is built for inference, graphics and video rather than training large models: it has Tensor Cores with BF16 and TF32 support, RT cores for ray tracing and hardware NVENC/NVDEC codecs. The table above holds only cloud plans billed by the hour, from a single card up to ×4 builds; the monthly price runs from roughly $330 to $6,200, with a median around $1,600.
What fits in 24 GB
Use the rule of thumb: in FP16 a model takes about 2 GB per billion parameters, in 8-bit quantization about 1 GB, in 4-bit 0.5–0.6 GB. On top of that you need headroom for the KV cache, which grows with context length and the number of concurrent requests.
- 7–8B parameters in FP16. The weights take 14–16 GB, leaving room for context and several parallel requests. This is the main reason people rent an A10: a chatbot, ticket classification, data extraction from documents.
- 13B in 8-bit. About 13–14 GB for the weights, with quality close to full precision.
- Around 30B in 4-bit. It fits, but only with a short context and one or two users: little memory is left for the KV cache.
- Fine-tuning. Full fine-tuning of even a 7B model will not fit on one card, but QLoRA for a 7B model runs fine. The A10 is not meant for training from scratch.
- Images, graphics, video. SDXL runs with memory to spare. RT cores make the A10 a good fit for remote graphics workstations, CAD and rendering with streamed output, and NVENC/NVDEC for transcoding video into several qualities at once.
A slice of the card or a full A10
Some cloud plans give you not the whole card but a share of it through NVIDIA virtual GPU software: 1/6, 1/3 or 1/2 with the matching part of the memory, that is 4, 8 or 12 GB. The A10 has no MIG, so the compute units are time-shared with neighbours while the memory is hard-assigned to each virtual machine. Such slices are most likely the rows at the top of the table. They are easy to spot:
- the video memory column shows less than 24 GB;
- the plan has few vCPUs and little RAM, a handful of cores instead of the dozens that come with a full card;
- on the server, nvidia-smi names the card with a profile suffix such as A10-4Q rather than just A10.
A slice is enough for a virtual desktop with 3D applications, light inference (embeddings, speech recognition, a 7B model in 4-bit on 8 GB) and a trial run before renting a full card. You need the whole A10 when the model takes more than 12 GB, when response latency must stay steady regardless of neighbours, or when you fine-tune.
Multiple cards, drivers and CUDA
The A10 has no NVLink: in ×2 and ×4 plans the cards exchange data over PCIe 4.0. For inference that is acceptable. vLLM and similar engines split a model across cards, so a ×2 plan (48 GB) holds a model of about 30B parameters in 8-bit and a ×4 plan (96 GB) holds a 70B model in 4-bit with room for context. Another mode often pays off better: run a separate copy of a 7–8B model on each card and balance requests between them. For training large models across several cards without NVLink, the A10 falls behind.
The A10 has compute capability 8.6. The architecture is supported from CUDA 11.1 onward, and CUDA 12 and current PyTorch builds need a driver from the 525 branch or newer. On card slices the virtual machine runs a vGPU guest driver that must match the branch of the host driver: a regular driver from the NVIDIA website may not work there, so use the provider's image and do not update the driver yourself. Right after delivery, check nvidia-smi: card model and profile, memory size, driver and CUDA versions.
Hourly billing and how the A10 compares with the L4 and A100
Every A10 plan in the catalog is billed by the hour. The median plan works out to roughly $2.20 per hour of uptime, the cheapest slice to about $0.45. Hourly A10 server rental pays off when you need the card for an evening, a batch job or a test week: delete the server and charges stop. The pitfalls are covered in our note on hourly billing. If a GPU server with an A10 stays busy around the clock for months, compare the total with a monthly physical machine in the dedicated GPU servers section. For data about people in the EU, also check the data centre country because of GDPR, and pick a location close to your users to keep latency low.
The closest alternative is the NVIDIA L4: the same 24 GB, Ada generation, about 72 W and FP8 support. The A10 has twice the memory bandwidth (600 GB/s versus 300), so in text generation, where speed is bound by memory, the A10 is usually faster while the L4 is more power-efficient. The A100, with 40 or 80 GB of fast memory, NVLink and MIG, is the choice when a model does not fit into 24 GB or you run full training; it costs noticeably more.