Can I put a Tesla into an ordinary PC bought on Allegro?

Sometimes, yes, but you risk no Above 4G in the BIOS, too weak a power supply and no room. The Dell/HP/Lenovo workstations on the list are predictable, which is why I recommend them.

Will a T5820/T7820/Z4/Z6 definitely handle a P40/V100?

That cannot be settled from the model name alone. Check the specific configuration: the PSU and available connectors, BIOS/firmware, MMIO space / Above 4G, the PCIe slot and mechanical space. Two workstations with the same name can differ in configuration.

How much do power and cooling really add?

When you buy the card from us, the power adapter is included, so in practice you only add cooling: a shroud with a fan, ~PLN 60–130. If you buy the card elsewhere, add the adapter as well (~PLN 30–100). Little in absolute terms, but without it the server does not work. (Indicative prices, as of September 2026.)

P40 or V100 to start with?

Budget and a lot of VRAM → P40. Faster inference, long context, a larger AI workload → V100. I explain the difference in architectures (Tensor Cores) in the text.

There is no single fixed threshold. A second card makes sense when one lacks VRAM or throughput for a specific model, context length, a larger batch or serving many users in parallel. With ~13B models a single P40 24GB gives plenty of headroom; with larger ones (of the order of 70B) the number of cards depends on the quantisation and how the model is run, including offloading.

Can it be cooled quietly?

It can be quieter than a server with roaring fans, but there is a trade-off between noise and temperature. The details are in our guide to cooling a passive card on a desk.

← Blog

How much does a cheap AI server cost in Poland - Dell/HP + Tesla + power and cooling (in PLN)

Agile Store · 9 September 2026 · 13 min read

topic:buying
topic:cooling
model:p40-24
model:v100-16
model:v100-32

Short answer: the used Tesla itself is only half the bill. For a working AI server you also need a workstation with a suitable BIOS and power supply (Dell Precision T5820/T7820, HP Z4/Z6, Lenovo ThinkStation P520), an adapter/cable for the CPU 8-pin power connector (these cards have a CPU input, not an ordinary PCIe one: you power them with a CPU 8-pin cable or NVIDIA's own adapter) and active cooling (a shroud with a fan: the cards are passive, designed for server-room airflow). These two items, power and cooling, are costs that almost nobody factors in up front, and without them you either cannot start the card or it overheats and starts throttling. (Our cards ship with the correct CPU 8-pin adapter included, so of these two the real extra expense is cooling.) Below is a realistic bill of materials in Polish zloty (PLN): all amounts are indicative, as of September 2026, because secondary-market prices change.

I am writing this from the perspective of someone who sells these cards and regularly sees the same scenario: a customer buys a Tesla P40 24GB for reasonable money, is pleased with the bargain, and then calls because "there is no way to connect it" and "why does it heat up to 90 degrees on the desk". Because a data-centre Tesla is not a GeForce. It has no fan of its own, no PCIe power connector, and it is designed to operate in a server with strong, forced airflow. You can build a cheap and useful local-AI machine from it, but you have to cost the whole thing, not just the card. In practice it will most often be a used workstation acting as an AI server, not a classic rack server.

What such a server is made of

Let us break it down into parts. Each item is a separate expense and a separate decision.

1. The base: a used workstation. Not every PC is suitable for working with a Tesla. You need a motherboard with Above 4G Decoding and correct handling of large MMIO regions in the BIOS: large accelerators such as the P40 or V100 can require a large PCIe address space, and allocating it incorrectly ends in a resource error or the GPU failing to initialise. The mere presence of an "Above 4G" option is not a guarantee of compatibility: the requirements depend on the specific card, firmware and platform, so verify the specific workstation model and its BIOS, not just the name of the computer. Add to that physical space for a dual-slot card approximately 267 mm long, a cooling duct and fan whose installed dimensions must be checked for the chosen design, and a power supply with spare capacity and the right connectors. Proven and inexpensive on the secondary market:

Dell Precision T5820 / T7820: a good BIOS, strong power supplies, plenty of space.
HP Z4 / Z6: solid, but they sometimes have proprietary power connectors (check before buying).
Lenovo ThinkStation P520: a decent single-socket base.

These are good starting points, but the model name alone does not guarantee compatibility. Configurations of used machines vary (PSU, BIOS/firmware, risers, PCIe slot population, available MMIO window, space, airflow), so verify the specification of the specific unit before buying. With that caveat, they are the favourites of people building budget AI servers.

2. The Tesla card. The heart of the matter: the P40 24GB when budget matters, the V100 16GB or V100 32GB when you need more power.

3. Power: the CPU 8-pin connector. Tesla cards (P40, V100 PCIe) have an 8-pin CPU-type power connector, not an ordinary PCIe one. NVIDIA allows two routes (P40 Product Brief, "Supported Auxiliary Power Connections"): 1× CPU 8-pin cable (EPS-keyed, matching the card's pinout) from a server power supply or, with an ordinary ATX power supply, a dedicated "CPU 8-pin → 2× PCIe 8-pin" adapter (two PCIe plugs from the power supply connected to the card through a dongle, fed from two separate PCIe cables, not one split lead). The CPU connector's pinout differs from PCIe, so do not plug in a random cable just because it fits the shape: choose the cable or adapter for the specific card model and your power supply. And the other way round: not every EPS/CPU 8-pin cable from any power supply is suitable: what matters is the pinout match and whether the power supply and cable can deliver the required power. Do not assume either that a detachable EPS/CPU cable from another (modular) power supply will be compatible: modular PSU cables do not have a universal pinout. That is why we sell our cards with the correct power adapter included: you get a reliable pinout and do not have to source it separately.

4. Active cooling. The P40 and V100 PCIe are passive (a heatsink without its own fan). In a server they are cooled by the chassis airflow; on a desk you have to provide the airflow yourself, most often with a 3D-printed shroud (a duct) with a 40–60 mm fan blowing through the heatsink. Without it, the card throttles or shuts down. I go into this in detail here: cooling a passive server card on a desk.

5. Optional: RAM and SSD. More RAM makes loading models and working with data easier; a fast NVMe SSD shortens the loading time of large GGUF files.

Hidden costs nobody counts

This is the heart of this post. The buyer looks at the price of the card and thinks that is the whole of it. But on top come:

Power adapter/cable (CPU 8-pin): about PLN 30–100 if you buy the card separately (check the cable or adapter's pinout and compatibility with the specific card and power supply; the EPS 8-pin label alone does not guarantee compatibility). When you buy the card from us, it is included: you do not pay separately and do not have to guess the pinout. (indicative, Allegro, as of September 2026)
Shroud + fan (cooling): about PLN 60–130 for a fan adapter (e.g. ~PLN 65) plus a fan (~PLN 40–60), or the cost of a 3D print + a fan. The item that decides temperature, throttling and stability, i.e. whether the card survives the summer. (indicative, Allegro, as of September 2026)
Possibly a stronger power supply: it depends on the unit, not the brand. The Dell T5820 has 425 W and 950 W PSU options; Dell Community reports indicate that the 425 W version does not support GPU power cables. HP QuickSpecs c05527757 states that the Z4 G4's 465 W option has no 6-pin graphics power cables. Lenovo P520 configurations with a 690 W or 900 W supply provide two 8-pin (6+2) GPU cables connected to the motherboard. Check the PSU label and GPU cables in the listing photos; if the supply is insufficient, budget for a higher-capacity unit compatible with that workstation and the required GPU cabling.
Electricity: every 100 W of continuous draw is about 876 kWh a year (100 W × 8,760 h); multiply by your tariff. A P40 or V100 PCIe under load draws up to 250 W. NVIDIA forum reports and the nvidia-pstated project indicate that the V100 has only the P0 state and that the P40 stays in P0 with a model held in VRAM unless software such as nvidia-pstated changes its state; this is community evidence rather than an NVIDIA specification. Measure the whole machine at the wall rather than guessing.
A display card: the P40 and V100 PCIe have no video outputs, and the Xeon W-2100 and W-2200 processors used in these single-socket workstations have no integrated graphics. Buy a base that includes a display card or budget for one. After initial setup, SSH permits remote operation if the particular machine boots without a display card; verify this before removing it. An additional card occupies an expansion slot.
Your time for assembly and BIOS configuration: free only in appearance.

The sum of these extra costs can add a noticeable percentage to the price of the card. Better to know this before buying than after. Before you buy anything, go through the checklist: what to watch for when buying a used data-centre card.

Budget variants: bill of materials

Below are two realistic builds. The items workstation, card, power and cooling are essential; RAM/SSD depends on the base (branded workstations often already have enough memory and a drive). All amounts are indicative (as of September 2026): secondary-market prices change, so treat them as an order of magnitude, not a certainty, and check current prices before deciding.

"Cheap build" variant: P40 24GB

Item	Example	Cost (PLN)
Workstation (used)	Dell T5820 / HP Z4 / Lenovo P520	~1,100–2,000
Card	Tesla P40 24GB	1,729 (1,406 net)
Power (CPU 8-pin)	included with our card (separately ~30–100)	0
Cooling	3D shroud + fan	~60–130
RAM/SSD (optional)	RAM upgrade + NVMe	0–400 (often in the base)
Total		~2,900–3,850 (with upgrades up to ~4,250)

Gross prices, indicative, as of September 2026 (base from Allegro, card from our range; secondary-market prices change). Note: a cheap T5820 often already comes with 32 GB of RAM and a 1 TB drive, in which case the RAM/SSD upgrade is not needed.

What for: a local LLM (Bielik, PLLuM), 24 GB of VRAM for 13B models with headroom, on a budget. A drawback I repeat honestly: the P40 (Pascal) has no Tensor Cores, and its FP16 performance is much weaker than the V100's, so in tasks that use FP16/Tensor Cores (including prefill of a long prompt) the V100 has a clear advantage.

"More powerful" variant: V100

Item	Example	Cost (PLN)
Workstation (used)	Dell T7820 / HP Z6	~2,500–3,500
Card	V100 16GB or V100 32GB	2,090 or 4,486
Power (CPU 8-pin)	included with our card (separately ~30–100)	0
Cooling	3D shroud + fan	~60–130
RAM/SSD (optional)	RAM upgrade + NVMe	0–400 (often in the base)
Total		~4,650–5,700 (V100 16GB; with RAM/SSD upgrades up to ~6,100) / ~7,050–8,100 (V100 32GB; with upgrades up to ~8,500)

Gross prices, indicative, as of 2026-09-19 (base from Allegro, cards from our range: V100 16GB PLN 1,699 net / PLN 2,090 gross, V100 32GB PLN 3,647 net / PLN 4,486 gross). For a single V100, consider a cheaper single-socket base such as the T5820 after checking its power, cooling, BIOS and card clearance. Compare a T7820 or Z6 if you need more CPU and RAM resources or plan to expand; a second GPU does not automatically require a dual-socket workstation.

What for: more compute power, HBM2 (high bandwidth) and first-generation Tensor Cores: in workloads that use them (FP16/Tensor Cores, the high bandwidth of HBM2) it can give a clear advantage over the P40, for example in prefill, long context and many parallel requests (served with llama.cpp, since current vLLM does not support Volta). 32 GB widens the range of models you can run on one card without splitting the model across GPUs.

If you need a rack solution with redundant power and the option to expand to several GPUs, see our ready-made Dell PowerEdge R740 AI servers.

What to watch out for when assembling

Above 4G Decoding / Memory Map IO above 4GB: enable it and check it before installing the card. With large GPUs the mere presence of this option does not guarantee correct allocation of MMIO space: it also depends on the BIOS, firmware and platform.
CPU connector ≠ PCIe: power the card with a CPU 8-pin cable compatible with the power supply or a dedicated CPU 8-pin → 2× PCIe adapter (our cards come with one included). This is not the place to experiment with a plug that fits "by force".
Cooling from the very first start: do not run the card "for a moment without a fan, just to check". A passive heatsink without airflow heats up in an instant.
A power supply with headroom: check the wattage and available GPU power cables. For NVIDIA's CPU 8-pin to dual PCIe adapter, we recommend two separate PCIe cables per card; do not assume that one cable with two plugs has sufficient current capacity. Dell Community reports indicate that the T5820's 425 W supply does not support GPU power cables. HP QuickSpecs c05527757 states that the Z4 G4's 465 W option has no 6-pin graphics power cables.
Drivers: R580 is the final driver branch for Pascal and Volta, supported through mid-2028, and CUDA 12.9 is the last toolkit that compiles for them. Pin branch 580 and choose CUDA 12.x application builds that include support for the specific GPU: sm_61 for P40 or sm_70 for V100. The CUDA version alone does not guarantee compatibility. Current vLLM requires compute capability 7.5 or higher and supports neither card. llama.cpp and Ollama support both with a compatible build and driver; also check the requirements of your chosen model and features.

Yes, we sell these cards (P40 24GB, V100 16GB, V100 32GB), so we have a commercial interest in selling them. All the more reason for us to show clearly the full cost of such a configuration (cooling, possibly a stronger power supply), rather than pretend that the card alone is enough. Every unit goes through a burn-in test, receives a certificate and ships with the correct CPU 8-pin power adapter included, and if you like, we will also advise on cooling for a specific base. Already have a workstation and do not know whether it will take a Tesla? Write to us: we will check together.

Sources

Card prices: from the shop's website (store.agilesoftware.pl/oferta/*, read on 2026-09-09; "indicative price", to order). Workstations/accessories: Allegro, indicative, as of September 2026 (the secondary market is volatile; ranges verified against listings on 2026-09-09).

Tesla P40 24GB: our listing (PLN 1,406 net / PLN 1,729 gross): https://store.agilesoftware.pl/en/products/p40-24
V100 16GB: our listing (PLN 1,699 net / PLN 2,090 gross): https://store.agilesoftware.pl/en/products/v100-16
V100 32GB: our listing (PLN 3,647 net / PLN 4,486 gross): https://store.agilesoftware.pl/en/products/v100-32
Used Dell Precision T5820 (~PLN 1,100–2,000) and T7820 (~PLN 2,500–3,500) workstations: Allegro (used listings, read September 2026): https://allegro.pl/listing?string=dell+precision+t5820
EPS 8-pin power adapter/cable for a Tesla (~PLN 30–100): Allegro: https://allegro.pl/listing?string=adapter+zasilania+eps+8pin
Third-party equivalent of the NVIDIA 030-0571-000 adapter: CPU 8-pin (to the card) → 2× PCIe 8-pin (from an ATX power supply): https://www.amazon.com/COMeap-2-Pack-Graphics-030-0571-000-Adapter/dp/B07M9X68DS
NVIDIA Tesla P40 Product Brief (PB-08338-001): "One CPU 8-pin auxiliary power connector"; the "Supported Auxiliary Power Connections" table (1× CPU 8-pin or a CPU→2× PCIe dongle); passive cooling; 250 W: https://images.nvidia.com/content/pdf/tesla/Tesla-P40-Product-Brief.pdf
Cooling: shroud adapter + fan for Tesla cards (~PLN 60–130): Allegro: https://allegro.pl/listing?string=adapter+chlodzenia+nvidia+tesla+wentylator
NVIDIA Tesla V100: technical data (HBM2, first-generation Tensor Cores, 250 W PCIe variant): https://images.nvidia.com/content/technologies/volta/pdf/tesla-volta-v100-datasheet-letter-fnl-web.pdf
NVIDIA Developer Forums: Tesla V100 PCIe powered by a CPU 8-pin connector (as with the P40): https://forums.developer.nvidia.com/t/nvidia-tesla-v100-power-connector/177834

FAQ

Can I put a Tesla into an ordinary PC bought on Allegro?: Sometimes, yes, but you risk no Above 4G in the BIOS, too weak a power supply and no room. The Dell/HP/Lenovo workstations on the list are predictable, which is why I recommend them.
Will a T5820/T7820/Z4/Z6 definitely handle a P40/V100?: That cannot be settled from the model name alone. Check the specific configuration: the PSU and available connectors, BIOS/firmware, MMIO space / Above 4G, the PCIe slot and mechanical space. Two workstations with the same name can differ in configuration.
How much do power and cooling really add?: When you buy the card from us, the power adapter is included, so in practice you only add cooling: a shroud with a fan, ~PLN 60–130. If you buy the card elsewhere, add the adapter as well (~PLN 30–100). Little in absolute terms, but without it the server does not work. (Indicative prices, as of September 2026.)
P40 or V100 to start with?: Budget and a lot of VRAM → P40. Faster inference, long context, a larger AI workload → V100. I explain the difference in architectures (Tensor Cores) in the text.
Do I need two cards?: There is no single fixed threshold. A second card makes sense when one lacks VRAM or throughput for a specific model, context length, a larger batch or serving many users in parallel. With ~13B models a single P40 24GB gives plenty of headroom; with larger ones (of the order of 70B) the number of cards depends on the quantisation and how the model is run, including offloading.
Can it be cooled quietly?: It can be quieter than a server with roaring fans, but there is a trade-off between noise and temperature. The details are in our guide to cooling a passive card on a desk.