NVIDIA Tesla P40 24GB GDDR5 datacenter
Equipped with 24GB of GDDR5 memory and the Pascal-based Tesla P40 chip, this card delivers 47 TOPS of INT8 inference performance for high-throughput deep learning deployment. Its passive cooling design and 250W TDP suit dense server environments, while ECC memory ensures data integrity for sustained compute loads. This GPU is best for data centers running large-scale AI inference workloads where replacing over 140 CPU servers with a single 8-GPU node is a priority.
No current retailer price
Snapshot
The 30-Second Version
The Tesla P40 is a 24GB AI monster that costs less than a mid-range gaming card. Just don't expect it to play Minesweeper, let alone anything modern.
Pros & Cons
Pros
- 24GB of VRAM for under $400 is a steal for AI tinkerers 91st
- 91st percentile AI performance in our database 77th
- ECC memory means fewer silent errors on long training runs 70th
- Passive cooling means zero fan noise, if you can keep it cool
Cons
- Gaming performance is genuinely terrible, 10.4/100 at 1080p
- No active cooling, so you'll need a server chassis or a fan mod
- Only 2 display outputs and ancient connectivity options
- Efficiency is rough at 250W for the performance you get
What owners think
The Word on the Street
How owner sentiment changed over time
ExclusiveBased on when customers actually wrote their reviews - so you can see whether early praise held up.
6 written reviews, Feb 2018 – Jul 2024: 2 happy, 1 mixed, 3 unhappy
Across all 9 store ratings: 3.8★. The chart counts the 6 written reviews.
Each dot is one written review, in the month it was written. Happy = 4-5★, mixed = 3★, unhappy = 1-2★.
The proof
Performance
What surprised us most is how well this old dog holds up for AI. Our database puts it in the 91st percentile for AI workloads, which is wild for a card this old. The 24GB of VRAM is the real star here, letting you load models that would choke cards with half the memory. But the raster performance is a different story, landing in the 26th percentile. This thing has no active cooling and a 250W TDP, so you'll need serious airflow in your chassis or it'll throttle itself into oblivion.
Specifications
Full Specifications
Chipset
| GPU Chipset | Tesla P40 |
| Brand | NVIDIA |
| Architecture | Pascal |
| Class | datacenter |
Memory
| VRAM | 24 GB |
| Memory Type | GDDR5 |
| Memory Bus | 384 |
| ECC | Yes |
Compute
| Base Clock | 1303 |
| Boost Clock | 1300 |
Performance
| PassMark G3D | 11595 |
| GPU Compute | 4082 |
Power & Cooling
| TDP | 250 |
| Power Connectors | None |
| Cooler | Passive |
| Fans | 0 |
Connectivity
| Interface | PCI Express 3.0 x16 |
| DisplayPort | 1 |
| HDMI | 1 |
| Max Resolution | 7680 x 4320 |
Physical
| Length | 269 |
| Slot Width | 2 |
| Weight | 1.4 kg / 3.1 lbs |
Features
| Ray Tracing | No |
| AV1 Encode | No |
| DirectX | 12 |
vs Competition
Against modern cards like the Gigabyte AORUS RTX 5080 or ASUS TUF RTX 5070 Ti, the P40 gets absolutely demolished in gaming and raster workloads. Those cards are in a completely different universe for anything involving pixels on a screen. But here's the thing: none of them have 24GB of VRAM at this price point. The RTX 5070 Ti has 16GB, and the 5080 has 16GB too. For loading big AI models, the P40's memory advantage is real. If you're doing AI work and don't care about gaming, the P40 is the budget king. If you want to do both, save up for a used RTX 3090 instead.
| Spec | NVIDIA Tesla P40 | ASUS GeForce RTX 5090 TUF Gaming | Gigabyte GeForce RTX 5070 Ti Eagle Ice | MSI GeForce RTX 5070 Ti Ventus 3X OC | ZOTAC GeForce RTX 5080 APOCALYPSE | PNY GeForce RTX 5080 ARGB OC |
|---|---|---|---|---|---|---|
| Chipset | Tesla P40 | GeForce RTX 5090 | GeForce RTX 5070 Ti | GeForce RTX 5070 Ti | GeForce RTX 5080 | GeForce RTX 5080 |
| Vram | 24GB GDDR5 | 32GB GDDR7 | 16GB GDDR7 | 16GB GDDR7 | 16GB GDDR7 | 16GB GDDR7 |
| Boost Clock Mhz | 1300 | 2407 | 2542 | 2482 | 2730 | 2620 |
| Tdp W | 250 | 575 | 300 | 300 | 360 | 360 |
| Memory Bus Bits | 384 | 512 | 256 | 256 | 256 | 256 |
| Ray Tracing | false | true | true | true | true | true |
| GPU Class | datacenter | consumer | consumer | consumer | consumer | consumer |
| Compare | Compare | Compare | Compare | Compare |
| Product | AI | Rt | Vram | Raster | Feature | Thermal | Efficiency | Connectivity | Social Proof | Affordability |
|---|---|---|---|---|---|---|---|---|---|---|
| NVIDIA Tesla P40 | 90.8 | 23.9 | 76.6 | 25.4 | 23.2 | 20 | 15.7 | 21 | 8.4 | 69.8 |
| ASUS GeForce RTX 5090 TUF Gaming Compare | 97.1 | 90.5 | 97.7 | 98.8 | 96 | 72.3 | 24.2 | 99.3 | 98.7 | 2.3 |
| Gigabyte GeForce RTX 5070 Ti Eagle Ice Compare | 79.6 | 88.7 | 86 | 89 | 96 | 91.6 | 72.2 | 94.1 | 93.3 | 26.8 |
| MSI GeForce RTX 5070 Ti Ventus 3X OC Compare | 79.6 | 92.4 | 86 | 89 | 96 | 91.6 | 72.2 | 94.1 | 83.8 | 23.4 |
| ZOTAC GeForce RTX 5080 APOCALYPSE Compare | 85.7 | 97.5 | 89 | 94.5 | 96 | 91.6 | 57.1 | 94.1 | 98.7 | 20.4 |
| PNY GeForce RTX 5080 ARGB OC Compare | 85.7 | 97.5 | 89 | 94.5 | 96 | 91.6 | 57.1 | 94.1 | 74.7 | 14.5 |
Price
Value & Pricing
At USD 368 to USD 535 across vendors, the P40 is a weird value proposition. If you can snag one at the low end of that spread, it's a fantastic budget AI card. But at the top end, you're getting into territory where newer used cards start to look tempting. The Amazon listing at the lower end is the one to watch. For pure AI inference on a budget, this is hard to beat. For anything else, it's a waste of money.
Read more
Overview
The NVIDIA Tesla P40 is a datacenter workhorse from 2016 that still punches way above its weight for AI inference, thanks to 24GB of GDDR5 and a Pascal chip that was built for exactly this kind of grunt work. The one thing to know: this is not a gaming card, and if you try to treat it like one, you'll be sorely disappointed. It's a specialized tool for running large language models and other AI workloads on a budget, and in that lane it's still surprisingly relevant.
Common Questions
Q: Can I use the Tesla P40 for gaming?
No. Seriously, don't. It scores 10.4/100 in our 1080p gaming tests. It has no active cooling and wasn't designed for real-time rendering. Get a used RTX 3060 or RX 6600 instead.
Q: Do I need special cooling for this card?
Yes, absolutely. The P40 is passive-cooled and expects server-grade airflow. In a normal desktop case, it'll overheat and throttle. You'll need to either mount a fan directly on it or put it in a chassis with serious front-to-back airflow.
Q: Is 24GB of VRAM enough for modern AI models?
For most 7B to 13B parameter models, yes. You can run quantized versions of Llama 2, Mistral, and similar models comfortably. For 70B models, you'll need to quantize aggressively or run multiple P40s. It's not going to match a modern 48GB card, but at this price, it's a great entry point.
Who Should Skip This
If you're looking for a gaming GPU, this isn't it. Go get a used RTX 3060 Ti or RX 6700 XT instead. If you want something that works out of the box without tinkering with cooling and drivers, this isn't it either. The P40 is for people who enjoy the challenge of making datacenter hardware work in a homelab, and who specifically need massive VRAM for AI workloads on a budget.
Verdict
The Tesla P40 is a one-trick pony, but it's a really good trick. If you're building a budget AI inference rig and need maximum VRAM per dollar, this is the card to get. Just know what you're signing up for: no active cooling, terrible gaming performance, and a card that's nearly a decade old. For the right person, it's an absolute bargain. For everyone else, it's a paperweight.
Usage Scores
- Overall29.8
- AI/LLM61.1
- Budget24.9
- Gaming 4k11.8
- Ray Tracing11.5
- Workstation45.6
- Gaming 1080p9
- Gaming 1440p10.8
- Content Creation12.8