In our measurements the V100 generated text 2.6–3.1× faster than the P40 with the same model and quant, but the first question is whether the bottleneck of your RAG is generation, or retrieval and context processing. Data from one model, one backend and one methodology for three cards.
Read more →The Tesla P40 generates tokens decently, but processing a long prompt (prefill) is slow on it, because the P40 (GP102) has no Tensor Cores and does not offer high FP16 throughput.
Read more →Two cards will let you run a model that does not fit on one, but not because the VRAM adds up into one big pool. llama.cpp and Ollama split the model into layers and distribute them between the cards.
Read more →Whisper (and preferably faster-whisper) transcribes Polish very well, and locally on a GPU it does so quickly and without sending recordings to an external service. In our test the V100 32GB reached ~15× realtime, and the P40/A2 ~6×.
Read more →What hardware for local RAG over company documents: the split between embeddings (P4/A2) and the LLM (V100/P40), the stack, GDPR/ACL/OCR and an honest comparison of your own hardware vs the cloud.
Read more →A used 2U server + Tesla for local AI: which base (R730/DL380 vs R740/Supermicro), power, and why cooling in a rack is not free: fan kit, riser, electricity, noise, colocation.
Read more →How much power used accelerators (P4, A2, P40, P100, V100) really draw and which has the best ratio of tokens per watt, from our own power-draw measurements.
Read more →The used Tesla itself is only half the bill. For a working AI server you also need a workstation with a suitable BIOS, power for the CPU 8-pin connector (not PCIe) and active cooling for a passive card. I set out the full bill of materials in PLN for the P40 24GB and the V100, with the hidden costs nobody talks about, and with the power adapter that comes included with our cards.
Read more →For ~PLN 2,000 you have a choice: a lot of VRAM for less (a used data-centre V100/P100/A2 16 GB card) or the convenience of a consumer RTX 3060/4060 Ti. I lay it out honestly: measured LLM and SDXL benchmarks, power draw, cooling, the lack of display output and the hidden cost of the older CUDA/PyTorch ecosystem, and I say when a Tesla makes no sense.
Read more →How we check used data-centre cards before sale: six qualification stages per unit and a certificate assigned to the unit, which you can verify yourself.
Read more →Most Teslas encode video in hardware, but not every one is suited to it. The full NVENC/NVDEC matrix, measured transcode streams for the whole family and an anti-guide: what to buy for Plex/Jellyfin, and what not to.
Read more →The Tesla P4 is an inexpensive, energy-efficient transcoder for Plex/Jellyfin. Step-by-step NVENC configuration, Docker pitfalls and our benchmark with its methodology: how many 1080p and 4K streams the encoder really handles, and what the figure does not guarantee.
Read more →You can run PLLuM locally on a used Tesla like any GGUF model, through Ollama or llama.cpp. This article covers how much VRAM you need, which card (P40/P100/V100) to choose and how to get it running.
Read more →For a local LLM in quantised form, choose the P40 (24 GB, it wins in GGUF). Choose the P100 with its trade-offs in mind for fast FP16 (2:1), high HBM2 bandwidth and strong FP64, at the cost of 8 GB less memory. The biggest trap: the P40's FP16 is around 1/64 of FP32, so native FP16/ExLlama runs terribly. With our own head-to-head benchmark on Bielik-11B.
Read more →We measured Bielik and PLLuM on used Teslas: real tokens/s with prefill time, not just generation alone. The figures, the measurement conditions and a template for your own verification.
Read more →As CUDA 13 drops Maxwell, Pascal and Volta, the A2 (Ampere) and T4 (Turing) from our range remain in official CUDA Toolkit 13.x support. We separate the CUDA Toolkit, the driver and the AI libraries, show that the P40/V100 still work on CUDA 12.x and, with an application → card table, suggest when paying extra for an A2/T4 makes sense and when more VRAM at a lower price is better.
Read more →Your Tesla card gives no picture or disappears from nvidia-smi? It is usually not a fault. A structured diagnosis: no video output, Above 4G and PCI/MMIO/BAR, the driver, CPU 8-pin power (pinout), passive cooling and a 2–4× P40 rig.
Read more →If you are looking for an inexpensive second-hand card for AI, you have probably already come across alarming headlines: "NVIDIA ends CUDA support for Pascal and Volta". It sounds like a death sentence for the Tesla P40 or V100, the most popular used cards for local models. The short answer: the cards do not stop working, but how long the newest software versions will support them is changing.
Read more →For several months I have been generating images on cards that once stood in server rooms as compute accelerators, not as graphics cards in the traditional sense. No video output, model-specific power connections and passive cooling that needs forced airflow: these cards are hardware you have to tame before they produce their first render in A1111 or ComfyUI.
Read more →The short answer before I go into detail: if you are planning real fine-tuning, even with LoRA, and compute throughput on large batches matters, go for the V100, ideally the 32GB version. If you are building an economical server on which fine-tuning is an addition to inference, and low power consumption and a modern stack (BF16, TF32) are the priority, the A2 will be the better choice.
Read more →The short answer: NVIDIA server cards (P40, P100, V100, P4, A2) have no fan of their own. You have to provide the airflow for them, whether in a server or in a desktop. Without it the card will overheat and throttle its clocks, and under sustained load it can shut itself down as a safety measure.
Read more →If you conduct research, manage a grant budget or buy equipment for a student research group, sooner or later you hit the same wall: a new data-centre accelerator costs as much as a year's stipend for several PhD students. The answer I arrived at myself after several purchases for my own testing needs is unglamorous: buy used data-centre hardware, choose the card for the specific task rather than for the marketing, and accept that this hardware will not last you a decade.
Read more →Buying a second-hand compute card is different from buying a used gaming GPU. The biggest risk is not that the card is "old": a Tesla P40 from 2016 can still do solid work for the next few years. The real risk is ending up with a unit that is faulty or worn out: a card that has spent years in a data centre under constant 24/7 load, sometimes at the edge of its thermal limits.
Read more →The short answer: yes, the P40 is very well suited to local LLMs, provided you stick to quantised models (GGUF) in llama.cpp or Ollama and do not try to run something that expects native FP16. Ignore this caveat and you will be convinced you have bought a faulty card. Take it into account and you get 24 GB of VRAM for working with 13B and 34B models for a fraction of the price of new hardware.
Read more →In short: a quantised 7B model will fit in 8 GB, 13B needs 16 GB to run comfortably, 34B only just fits in 24 GB, and for 70B you realistically need around 40-48 GB of VRAM, that is, two 24 GB cards or one card with a large amount of memory. Below I explain where these figures come from and why the context can upset all of it.
Read more →The short answer before I go into detail: if you are setting up a local language model for the first time and do not know where to start, in most cases the Tesla P40 with 24 GB of VRAM is the best place to start. Not because it is the fastest, but because it gives you a lot of memory on a single card for reasonable money. Below: why VRAM comes first, what bandwidth adds, and when to pick something else.
Read more →Server cards deliver data-centre performance for a fraction of the price of new hardware, but they have quirks you will not find in the specification. A practical guide to the V100, P40, P100, P4 and A2.
Read more →