Tesla P40 Price, 24GB VRAM & Local LLMs
Price alert
We'll email you when the eBay (US) price reaches your target.
NVIDIA's Tesla P40 (2016) has 24GB of VRAM and still runs 30B-class MoE models at about 81 tokens/s in llama.cpp, though vLLM won't run on it. It takes a CPU 8-pin power cable, so have the right adapter ready.
eBay
Tesla P40 price history
Current$398+0.51%
eBay
Price updated
$490$453$417$380$343
Low $380High $469
Sep 23Oct 19Nov 14
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Tesla P40 VRAM & full specs
- Memory24 GB · GDDR5
- Bandwidth347 GB/s
- INT847 TOPS
- Bus width
- 384-bit
- Architecture
- Pascal (GP102)
- Cores
- 3,840
- Ecosystem
- CUDA
- Interface
- PCIe 3.0 x16
- NVLink
- Not supported
- Power
- 250 W
- Power connector
- 8-pin CPU (EPS-12V)
- Slots
- 2
- Length
- 267 mm
- Type
- Data center
- Launch date
- Sep 13, 2016
Best Tesla P40 deals
Powered by VRAMGlass real-time price tracking — handpicking the best-value hardware deals across the web
Picked from
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Picked from
As an Amazon Associate I earn from qualifying purchases. The price you pay stays the same.
What LLMs can a single Tesla P40 run?
One P40 is best with 30B-class MoE at 4-bit, about 80 tokens/s with over 110K of context. 27B to 31B dense gets only 14 to 16; 8-bit adds little.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 E4B8B | 3.8 GB | Runs | 62.6 | 128K |
 Qwen3.5 9B9.7B | 4.1 GB | Runs | 57.8 | 256K |
| Gemma 4 12B12B | 4.7 GB | Runs | 50.2 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 88.1 | 256K |
| Qwen3.6 27B27.8B | 11.8 GB | Runs | 21.4 | 192K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 25.5 | 221K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 100.1 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 20.6 | 143K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 112.8 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 5 GB | Runs | 48.8 | 128K |
| Qwen3.5 9B9.7B | 5.7 GB | Runs | 42.9 | 256K |
| Gemma 4 12B12B | 7.1 GB | Runs | 34.5 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 66.8 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 15.3 | 117K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 15.5 | 122K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 80.7 | 115K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 13.7 | 66K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 83.7 | 123K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 8.2 GB | Runs | 30.8 | 128K |
| Qwen3.5 9B9.7B | 9.5 GB | Runs | 26.6 | 256K |
| Gemma 4 12B12B | 12.7 GB | Runs | 20 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Needs RAM offload | 22.5 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Too big | — | — |
| Qwen3.8 27B27.8B | 29 GB | Too big | — | — |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Needs RAM offload | 21.1 | 198K |
| Gemma 4 31B30.7B | 32.6 GB | Too big | — | — |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Needs RAM offload | 29.7 | 256K |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Needs RAM offloadMoE only: expert weights sit in system RAM, so speed is estimated from 70 GB/s RAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
What can the Tesla P40 run with Strata?
With Strata, the Tesla P40 can run models far larger than its VRAM. Strata keeps most of the model in RAM and on the SSD and only the most-used experts on the GPU. Benchmarks come from Strata’s author and community, with multi-GPU setups marked by card count and today’s prices added by VRAMGlass.
Where Strata puts the model
Qwen3.8 Flash Next Q2_0 · 66.4 GB- VRAM 24 GB: the parts every token uses, plus the most-used experts
- RAM about 40 GB: all the experts; the CPU computes the ones the GPU doesn’t hold
- SSD 28.8 GB: the lookup table, a few rows read per token
IQ2_XS · 68 GBMeasured
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
What LLMs can multiple Tesla P40 cards run?
Four P40s fit gpt-oss 120B at 4-bit, but without vLLM support, llama.cpp splitting is slower. Two only manage 70B-class at 2-bit, under 10 tokens/s.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 4 cards | $1,592 | 96 GB | 61.4 | 1,000 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 8 cards | $3,184 | 192 GB | 46.1 | 2,000 W |
 Hy3295B (21B active) | 107.2 GB | 8 cards | $3,184 | 192 GB | 18.8 | 2,000 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 8 cards | $3,184 | 192 GB | 35.3 | 2,000 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 8 cards | $3,184 | 192 GB | 48.9 | 2,000 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 8 cards | $3,184 | 192 GB | 31.5 | 2,000 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 8 cards | $3,184 | 192 GB | 30.4 | 2,000 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Mainly running MoE models
Qwen3.6 35B A3B and GLM 4.7 Flash at 4-bit run at about 81 to 84 tokens/s, with more than 115K of context
- You only use llama.cpp or Ollama and can live with about 16 tokens/s on 27B dense models
- You have a server with strong airflow, or are willing to add your own shroud and fans
Not a good fit
- You need vLLM or FP16 compute
vLLM has never supported this generation, and its FP16 throughput is only 1/64 of FP32
- You want to keep up with new drivers and CUDA versions
CUDA 13 no longer supports it, and neither do drivers after 580
- A regular case without directed airflow
It is a 250W passively cooled card with no fan
Before you buy
Power is CPU 8-pin, not GPU 8-pin
At the end of the P40 is a CPU 8-pin (EPS) socket, whose pinout differs from the PCIe 8-pin common on graphics cards, so don't force a GPU cable into it. NVIDIA states that you can connect a CPU cable from the power supply directly, or use an adapter that combines two PCIe 8-pin into one CPU 8-pin; make sure you buy the right adapter.
Passive cooling means adding your own fans
The P40 has no fan and is designed to rely on server airflow; its heatsink accepts air from either direction. In a desktop you need to fit your own shroud and a high-speed fan (3D-printed shrouds are common), and some people use two 40mm fans per card. The card also has no video output, so your build needs integrated graphics or another graphics card for the display.
Enable Above 4G Decoding on the motherboard
The P40 needs a large chunk of address space (32 GB), and if Above 4G Decoding isn't enabled in the BIOS, the card may not be detected. Older platforms and some consumer motherboards don't have this option, so check your motherboard manual before buying.
Check whether ECC is on
The P40 ships with ECC enabled by default, in which case nvidia-smi shows only 22912 MiB of total memory, about 1.6 GB short of 24GB. If you don't need ECC, turn it off with nvidia-smi -e 0, and after a reboot you get the full 24GB.
Software comes down to llama.cpp and its relatives
This generation has no Tensor Cores and extremely low FP16 throughput, so frameworks that rely on FP16 compute run very slowly on it. llama.cpp automatically avoids FP16 and computes in FP32 and INT8 instead, so GGUF quantized models still run. Ollama still supports it too, but needs a 570 or later driver, and 580 is the last branch that supports the card. When downloading llama.cpp prebuilt packages, pick the CUDA 12 build; the CUDA 13 build no longer supports this card.
Compared with similar GPUs
Tesla P40 vs GeForce RTX 3090, Instinct MI50 32GB, and GeForce RTX 2080 Ti 22GB: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | Tesla P40 | GeForce RTX 3090 | Instinct MI50 32GB | GeForce RTX 2080 Ti 22GB |
|---|---|---|---|---|
| Memory | 24 GB | 24 GBSame | 32 GB33% more | 22 GB8% less |
| Bandwidth | 347 GB/s | 936 GB/s170% higher | 1,024 GB/s195% higher | 616 GB/s78% higher |
| BF16 / FP16 compute | — | 142 TFLOPS | 27 TFLOPS | 108 TFLOPS |
| INT8 compute | 47 TOPS | 285 TOPS506% higher | 53 TOPS13% higher | 215 TOPS358% higher |
| Qwen3.8 27B speed (2-bit) | 25.5 t/s | 64.8 t/s154% faster | 70.3 t/s176% faster | 44 t/s73% faster |
| Largest dense model (4-bit) | 37B | 37BSame | 51B38% more | 34B8% less |
| Power | 250 W | 350 W40% higher | 300 W20% higher | 250 WSame |
| GPU interconnect | Not supported | NVLink 112.5 GB/s | — | NVLink 100 GB/s |
| Interface | PCIe 3.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 3.0 x16 |
| Released | 2016-09 | 2020-09 | 2018-11 | 2018-09 |
| eBay median |
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Tesla P40 rental price per hour
On Compshare, the Tesla P40 rents for $0.06 per hour on demand, about $14/month at 8 hours a day. Buying one (eBay median asking price $398) pays off in about 23 years, 11.5 years or 3.8 years at 4, 8 or 24 hours a day.
| Option | Price | 4 h/day | 8 h/day | 24 h/day | View |
|---|---|---|---|---|---|
 CompshareRentUpdated 29 min. ago | $0.06/hr | $7/mo | $14/mo | $41/mo | Rent |
 eBayUpdated 3 hr. ago | $398 | Payback ~23 yrsPower $5/mo | Payback ~11.5 yrsPower $11/mo | Payback ~3.8 yrsPower $33/mo | Buy |
Break-even uses the US residential average ($0.18/kWh) and 250 W rated power, excluding resale value and the rest of the build.
Rental links are referral links: we may receive a commission or credit if you sign up through them. As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Amazon.jp
Xianyu