The Nvidia Tesla V100 32GB graphics accelerator card, a Volta architecture card originally released in 2017, is, against all odds, experiencing a resurgence in popularity and a rise in price on the second-hand market.

Although it’s an older graphics card and lacks the efficiencies of newer architectures, the Tesla V100 retains a key specification: it comes with 32 GB of HBM2 VRAM with an impressive bandwidth of 900 GB/s. This makes it a highly attractive option for inference, prompt processing, and token generation tasks without having to invest thousands of dollars in modern professional GPUs like the RTX 6000 Ada.

Currently, the market has a steady supply of these cards, available in both the SXM2 server format and the standard PCIe version. However, high demand from users looking to build home AI servers has caused their price to increase by about $200 compared to previous months.

Despite this increase, it is still considered an economical alternative. Assembling a system dedicated to local models for around $2,000 is a viable option using the V100, especially given the prohibitive cost of the latest generation of accelerators.

It’s not all advantages; this GPU has a high TDP of 250W, doesn’t support newer precision formats like FP8, and requires adaptive active cooling solutions or special adapters if using the SXM2 version on off-the-shelf motherboards. Even so, for those who prioritize memory capacity at a more or less reasonable cost, the V100 remains an interesting option.

In other Nvidia news, modders have successfully extracted and optimized the company’s latest DLSS 5 Neural Rendering model, delivering frame rate performance boosts of up to 40% in games that do not officially support it