NVIDIA Tesla P100 16GB HBM2 datacenter Black

Equipped with 16GB of HBM2 memory on a 4096-bit bus and 3584 CUDA cores, this Pascal-based accelerator delivers high double-precision throughput for HPC workloads. Its passive 250W design and PCIe Gen3 interface allow dense server integration, effectively replacing half a rack of CPU nodes in a single system. This refurbished card is best for research institutions running legacy CUDA simulations who need affordable, high-memory-bandwidth compute without scaling to newer architectures.

GPU Tesla P100
VRAM 16GB HBM2
TDP 250 W
Memory Bus 4096 bit
Class datacenter
NVIDIA Tesla P100 16GB HBM2 datacenter Black gpu
31 Overall Score
Also available in:

Snapshot

The 30-Second Version

A $117 ticket to 16GB of HBM2, if you don't mind bolting a fan to a server relic.

Download image

Pros & Cons

Pros

  • 16GB HBM2 on a 4096-bit bus, well above average for VRAM 82nd
  • $117 USD makes it one of the cheapest paths to 16GB of CUDA-accessible memory 78th
  • ECC support for error-sensitive workloads
  • Affordability is one of the best across all GPUs we track

Cons

  • Passive cooling demands server-grade airflow, not a desktop case
  • Gaming and raster performance are near dead last
  • 250W TDP with efficiency near the bottom, it's a power hog for what it delivers
  • No display outputs whatsoever, this is compute-only

What owners think

The Word on the Street

4.1/5 (20 reviews)
👍 Owners love the price-to-VRAM ratio, calling it a budget hack for running larger models locally.
👎 A common complaint is the cooling situation, multiple buyers report needing custom fan solutions or server chassis to keep temps sane.
🤔 Some owners appreciate ECC and the HBM2 bandwidth, but others point out that Pascal's age shows in compute throughput compared to newer cards.

The proof

Performance

What surprised us most is how lopsided this thing is. The memory subsystem is genuinely strong, 16GB of HBM2 on a 4096-bit interface ranks among the best, but raw compute sits at just the 14th percentile with 3584 CUDA cores on Pascal architecture. That means it's memory-rich and compute-poor, which is actually fine for loading larger models but painful for training speed. Efficiency is also near the bottom, so that 250W TDP buys you less work per watt than almost anything else we track. And the connectivity score is the worst we've seen, PCIe Gen3 x16 with zero display outputs means this card literally cannot drive a monitor.

Performance Percentiles

AI 64.9
Rt 16.1
Vram 81.8
Raster 15.7
Compute 14.1
Feature 40.2
Thermal 13.5
Efficiency 11.9
Connectivity 48.1
Affordability 77.8

Specifications

Full Specifications

Chipset

GPU Chipset Tesla P100
Brand NVIDIA
Architecture Pascal
Class datacenter

Memory

VRAM 16 GB
Memory Type HBM2
Memory Bus 4096
ECC Yes

Compute

CUDA Cores 3584

Power & Cooling

TDP 250
Cooler Passive
Fans 0

Connectivity

Interface PCI Express 3.0 x16
DisplayPort 4
Max Displays 4
Max Resolution 5120 x 2880

Physical

Length 267
Slot Width 2
Form Factor Full Height

Features

Ray Tracing No
Upscaling None
AV1 Encode No
VR Ready No
DirectX 12
RGB No

vs Competition

Cross-shop this against the MSI GeForce RTX 5060 SHADOW, PNY GeForce RTX 5060 Dual Fan OC, or Gigabyte GeForce RTX 5060 WINDFORCE OC if you want something that actually works in a normal PC with display outputs and modern efficiency. The XFX Radeon RX 7600 XT Speedster is another option if you're open to AMD. The P100 only wins if your single priority is maximum VRAM per dollar and you can handle the cooling situation yourself.

Spec NVIDIA Tesla P100 Gigabyte GeForce RTX 5070 Ti Eagle Ice ASUS GeForce RTX 5080 PRIME EVO OC MSI GeForce RTX 5070 Ti Ventus 3X OC ZOTAC GeForce RTX 5080 APOCALYPSE Sapphire Radeon RX 7900 XTX NITRO+
Chipset Tesla P100 GeForce RTX 5070 Ti GeForce RTX 5080 GeForce RTX 5070 Ti GeForce RTX 5080 Radeon RX 7900 XTX
Vram 16GB HBM2 16GB GDDR7 16GB GDDR7 16GB GDDR7 16GB GDDR7 24GB GDDR6
Boost Clock Mhz - 2542 2685 2482 2730 2680
Tdp W 250 300 360 300 360 339
Memory Bus Bits 4096 256 256 256 256 384
Ray Tracing false true true true true true
GPU Class datacenter consumer consumer consumer consumer consumer
Compare Compare Compare Compare Compare
Product AIRtVramRasterComputeFeatureThermalEfficiencyConnectivityAffordability
NVIDIA Tesla P100 64.916.181.815.714.140.213.511.948.177.8
Gigabyte GeForce RTX 5070 Ti Eagle Ice Compare 79.588.786.18988.195.991.77294.126.6
ASUS GeForce RTX 5080 PRIME EVO OC Compare 85.697.689.194.593.488.791.756.994.117.7
MSI GeForce RTX 5070 Ti Ventus 3X OC Compare 79.592.586.18988.195.991.77294.123.2
ZOTAC GeForce RTX 5080 APOCALYPSE Compare 85.697.689.194.593.495.991.756.994.120.3
Sapphire Radeon RX 7900 XTX NITRO+ Compare 95.196.39283.595.695.991.751.998.818.7

Price

Value & Pricing

At $117 USD, this is worth it if you specifically need cheap VRAM for inference experiments and you already have a way to cool a passive 250W card. For anyone else, the CAD 849 listing in Canada is a hard pass when modern alternatives exist at lower prices in that market.

From CA$849 2 offers across 1 retailer
Newegg 2 offers From CA$849

We started tracking prices for this product on Nov 22, 2026. The chart appears once we have more data.

Read more

Overview

The Tesla P100 is a 16GB HBM2 compute card you can grab for $117, and that's the whole pitch. It's not a graphics card in any practical sense: no display outputs, passive cooling built for server airflow, and a gaming score of 9.1/100 that puts it near the bottom of our rankings. But if you want cheap VRAM for CUDA experiments, this Pascal-era datacenter part still has a job to do.

Our scoring model gives it 50.6/100 for AI and LLM work, which is middling but respectable for something this old. The 4096-bit HBM2 bus feeding 16GB is the standout spec here, sitting well above average for VRAM capacity in our data. Just know what you're signing up for: a 250W passive card that needs a server chassis or a janky fan mod to not cook itself.

Common Questions

Q: Can I use this as a regular graphics card?

No. It has zero display outputs. This is a compute accelerator only, you'll need a separate GPU or integrated graphics for your monitor.

Q: Will it fit in a normal desktop case?

Physically yes at 267mm and 2 slots, but it's passively cooled. Without forced airflow from a server chassis or a DIY fan mount, it will thermal throttle or shut down under load.

Q: Is 16GB enough for running local LLMs?

For inference on quantized models up to around 13B parameters, yes. The HBM2 bandwidth helps with token generation speed, but don't expect fast training, the compute cores are dated.

Who Should Skip This

If you want a GPU that just works in your desktop for gaming, creative apps, or even hassle-free AI tinkering, this isn't it. Go grab the MSI GeForce RTX 5060 SHADOW or the Gigabyte GeForce RTX 5060 WINDFORCE OC instead. You'll get display outputs, active cooling, and modern efficiency without playing thermal engineer.

Verdict

Buy the Tesla P100 if you're a tinkerer who wants 16GB of HBM2 for under $120 and you've got a plan for cooling a passive server card. Skip it if you want plug-and-play anything. Our ranking puts it solidly mid-pack for AI work but near the bottom for everything else, and that's exactly what it is: a niche tool, not a general-purpose GPU.

Usage Scores

Overall (30.7)AI/LLM (52.5)Budget (29.8)Gaming 4k (13.3)Ray Tracing (11.4)Workstation (39.5)Gaming 1080p (10.6)Gaming 1440p (12.3)Content Creation (12.2)
  • Overall30.7
  • AI/LLM52.5
  • Budget29.8
  • Gaming 4k13.3
  • Ray Tracing11.4
  • Workstation39.5
  • Gaming 1080p10.6
  • Gaming 1440p12.3
  • Content Creation12.2

Similar Products