Your Cheap Tesla Won't Boot

Five things that go wrong after the card arrives, and what each costs to fix | Prices live as of November 3, 2026

Our guide to used server GPUs covers which card to buy. This page is about the week after it arrives, when a $150 P100 is sitting in your machine doing nothing and you are trying to work out whether you have been sold a brick.

Almost always you have not. These cards were built to be slotted into a server by someone who already knew all of this, so none of it is written down where a first-time buyer would find it.

Black screen on first boot: Above 4G Decoding

This is the one that convinces people their card is faulty. You install the GPU, power on, and get nothing: no POST, no beep, sometimes not even the motherboard splash.

Datacenter GPUs expose their entire framebuffer as one memory region. A P40 wants 24GB of address space mapped at once, and a desktop BIOS will not do that by default because consumer cards never ask for it. The setting that permits it is called Above 4G Decoding, sometimes Above 4GB MMIO or 64-bit capable devices depending on the vendor. It lives in the PCIe or Advanced section, and it is off by default on most consumer boards.

Enabling it is not always painless. On some boards, particularly older ones, turning Above 4G Decoding on causes its own set of problems: blank output, unresponsive USB, or a machine that boots with one card and refuses with two. If that happens you are into firmware-version territory, and the answer is usually a BIOS update rather than a different GPU.

Server and workstation boards tend to have it enabled already, which is one quiet argument for buying an older Xeon or Threadripper platform if you plan to run several of these cards.

The power connector is not the one you think

Most used datacenter cards, the P40, P100, M40 and K80 among them, take a single 8-pin EPS connector. That is the CPU-style plug, not the PCIe 8-pin your power supply has hanging off the GPU cables.

The two look similar, they are keyed differently, and the pinouts are not the same. The adapter you want goes from dual PCIe 8-pin (from your PSU) to one CPU 8-pin (into the card). NVIDIA's own part number for it is 030-0571-000, and third-party copies sell for around $10.

Budget for the power draw too. A P40 or M40 pulls up to 250W on its own, on top of whatever your system already uses. A 600W supply is the realistic floor for one card in an otherwise modest build, and more if your CPU is hungry.

No fan, and no way to add one without a shroud

Every Tesla card is passively cooled. Not quietly cooled, not lightly cooled: there is no fan at all. The heatsink is a duct designed for a rack chassis blowing high-pressure air straight through it.

A desktop case does not do that. Case fans move a lot of air at very low pressure, which is the wrong shape of airflow entirely, so simply pointing one at the card accomplishes less than you would expect. Left alone, a P40 will heat up until it throttles, and throttling is not subtle. One builder documenting a living-room P40 measured 3.6 tokens per second at 93°C before cooling, and a steady 10 tokens per second after. Nearly a third of the card's speed was sitting behind a $25 part.

The fix is a shroud: a printed plastic duct that clamps to the end of the card and seals a fan against the heatsink opening, so the air has nowhere to go except through the fins. Shrouds are sold ready-made on eBay and Amazon for roughly $25 including a fan as of September 2026, and the STL files are free on Printables if you have a printer.

• A 40mm blower is the loudest option and the most effective. Fine in another room, unpleasant beside a desk.
• Twin 40mm fans in a shroud trade some airflow for much less noise, and the ariezlabs build shows this is enough for sustained inference.
• A ducted 120mm fan is quietest if you have the case depth, because a large slow fan moving air through a sealed duct beats a small screaming one.

Whatever you fit, watch the first real workload with nvidia-smi -l 1. Under 80°C sustained is healthy. A P40 begins throttling near 90°C and protects itself around 95°C.

There is no display output

No Tesla card has a video port. Neither does an A40 or an MI210. They were never intended to draw a desktop, and no driver setting changes that.

So you need something else to produce a picture: integrated graphics on the CPU, or a cheap second GPU in another slot. This catches people who assumed they could drop a P40 into a machine whose only graphics card they had just removed. Check your CPU actually has an iGPU before you plan around it, as many Xeons and several Ryzen parts do not.

The P40's FP16 is a trap, and the P100's is not

This is the one that costs performance rather than a weekend, and it is the least known of the five.

The P40 and the P100 are both Pascal, both from 2016, and they behave completely differently in half precision. The P40 runs FP16 at 1/64th of its FP32 rate: 11.76 TFLOPS drops to roughly 184 GFLOPS. The P100 does the opposite, running FP16 at double its FP32 rate for about 18.7 TFLOPS.

That is a hundredfold difference in half-precision throughput between two cards people treat as near-equivalents. It rarely bites during plain quantized inference in llama.cpp, which spends its time on integer kernels and is bound by memory bandwidth anyway. It bites hard the moment you run something that assumes FP16 is the fast path: diffusion models, most fine-tuning, and a good deal of PyTorch written since 2020.

Neither card supports bf16 at all, which is a separate problem. Bf16 is the default in much of the current training and inference ecosystem, and Pascal predates it. Software will usually fall back to FP32 rather than fail, but you are then paying for the fallback.

The practical reading: buy the $345 P40 for capacity, because 24GB in one pool is what it is for. Buy the $150 P100 for speed, because it has 732 GB/s of HBM2 and real half precision. If you intend to fine-tune anything, the 16GB card is more useful than the 24GB one, which inverts the ranking our own $/GB table would give you.

What it all actually costs

Added up, the accessories are a rounding error against a $1,500 RTX 3090 and a real fraction of a $150 P100. The card price below is live; the accessory figures were observed in September 2026 and we do not track them:

That is still cheap for 24GB, which is why people keep buying these cards. But $380 and an afternoon of BIOS settings is a different proposition from $345 and a spare slot, and it is worth knowing which one you are signing up for. The $191 M40 carries every one of these costs for older silicon, so the accessory bill is a larger share of a worse card.

If none of this appeals, that is the honest argument for a $1,500 RTX 3090: same 24GB, its own fan, a display output, ordinary PCIe power, and nothing to configure. You are paying 4.3 times as much for the card and nothing at all for the afternoon.

Before you buy, not after

• Does the motherboard have Above 4G Decoding? Check the manual, not the marketing page.
• Do you have onboard graphics or a spare slot and card for display?
• Is there physical room? These are full-length dual-slot cards, and the shroud adds 40 to 60mm past the end.
• Does the PSU have two spare PCIe 8-pins for the adapter, and headroom for another 250W?
• Does your actual workload need FP16? If so, the P100 over the P40, or neither.

Our guide to buying used GPUs on eBay covers seller checks and spotting a card pulled from a failed server. For which card to buy in the first place, start with used server GPUs for local AI or the live price table.