Stock items leave our warehouse in Germany within 24 hours
NVIDIA accelerators for AI inference, training and rendering: Tesla P40, T4 and V100, L4, A40, A100 and RTX A6000. Every card is checked and tested before it ships.
No products found
Adjust the filters or ask us directly. We can source most hardware within a few days.
NVIDIA data-centre GPUs differ from PC graphics cards less in the chip than in everything around it: cooling, memory, power connector, drivers and licensing. If you're buying refurbished, the architecture is the thing to know. It decides which number formats a card handles quickly and how long current software will keep supporting it.
Up to the Turing generation, data-centre cards carried the Tesla name; from Ampere onwards it was dropped. The letter before the number gives the architecture: P for Pascal, V for Volta, T for Turing, A for Ampere, L for Ada Lovelace, H for Hopper. The number roughly sets the class: the A100 is the flagship of its generation, the L4 a frugal low-profile card. The RTX A series, such as the RTX A6000, is the actively cooled workstation line formerly known as Quadro.
| Architecture | Cards in our catalogue | Memory | Strengths and limits |
|---|---|---|---|
| Pascal (2016) | Tesla P40 | 24 GB GDDR5 | lots of memory for LLM inference; no tensor cores, very slow FP16 |
| Volta (2017) | Tesla V100 16 GB | 16 GB HBM2 | first tensor cores, strong FP64 for simulation |
| Turing (2018) | Tesla T4 | 16 GB GDDR6 | 70 W, low profile, INT8 inference and video transcoding |
| Ampere (2020) | A100 40 GB, A100 80 GB, A40, RTX A6000 | 40 or 80 GB HBM2/HBM2e, 48 GB GDDR6 | TF32 and BF16, MIG partitioning (A100), vGPU and rendering (A40) |
| Ada Lovelace (2022) | L4 |
| 24 GB GDDR6 |
| FP8, AV1 encoder, 72 W in a low-profile card |
In practice, the compute capability that CUDA software queries tells the generations apart: 6.1 on the P40, 7.0 on the V100, 7.5 on the T4, 8.0 on the A100, 8.6 on the A40 and RTX A6000, 8.9 on the L4. Many AI frameworks require a minimum version, and older architectures gradually drop out of support in new CUDA releases. So check what your software needs before you buy.
CUDA runs on both. In a server, though, four differences decide:
Servers with the space, power and airflow for several cards are under GPU servers, such as the Dell R740 with 2× Tesla V100, or systems from Supermicro with room for up to eight cards. We take decommissioned cards through our buyback service.