NVIDIA Datacenter GPUs Compared: A100 vs V100 vs P100 vs P40 (Tesla & Ampere)

A side-by-side of the four NVIDIA datacenter GPUs you can actually buy today — A100, V100, P100, and Tesla P40 — covering architecture, VRAM, Tensor Cores, power, and the specific AI, HPC, and VDI jobs each one wins. With guidance on which to pick for training vs inference.

Topics: NVIDIA, A100, V100, Tesla P40, Tesla P100, GPU Comparison

TL;DR — the one-line verdict for each

  • A100 — the modern AI workhorse. Ampere Tensor Cores, 40/80GB, MIG. Buy it for training and serious inference.
  • V100 — still-capable Volta Tensor Cores, 16/32GB. Great mid-tier for training smaller models and inference.
  • P40 — 24GB and INT8: an inference and VDI value pick, despite no Tensor Cores.
  • P100 — Pascal HBM2 with strong FP64: HPC and budget compute, not modern AI training.

---

Side-by-side

SpecA100V100P40P100
ArchitectureAmpereVoltaPascalPascal
VRAM40 / 80GB HBM2e16 / 32GB HBM224GB GDDR512 / 16GB HBM2
Tensor CoresYes (3rd gen)Yes (1st gen)NoNo
Best precisionTF32 / FP16 / INT8 / FP64FP16 / FP32 / FP64INT8 / FP32FP16 / FP64
MIG partitioningYesNoNoNo
Approx TDP (PCIe)~250W~250W~250W~250W
Sweet spotTraining + inferenceTraining + inferenceInference + VDIHPC / FP64 / budget

Specs per NVIDIA's official datasheets — A100, V100, P40, P100.

---

A100 vs V100 (the real training question)

If you're training transformers, the gap is large: NVIDIA cites the A100's TF32 Tensor Cores at up to 20x the V100 for some workloads, and the A100's 80GB option plus MIG make it far more flexible for big models and multi-tenant serving. The V100 is still a genuinely useful card for smaller models, classic CNNs, and inference — and at refurbished pricing it's often the better dollars-per-result for labs that don't need 80GB.

Pick A100 when: large models, multi-tenant inference (MIG), or you want headroom for years. Pick V100 when: budget matters, models fit in 16/32GB, and you want Tensor Cores without A100 pricing.

---

P40 vs V100 for inference (the VRAM upset)

The P40 has no Tensor Cores, so on paper it looks older than the V100 — but it carries 24GB and a strong INT8 inference engine. For serving larger inference models or VDI (many virtual desktops sharing a GPU), the P40's bigger memory can win. For training, the V100 is clearly ahead.

---

P100 — where it still fits

The P100 brought HBM2 to Pascal and has solid FP64 (double precision) — useful for scientific/HPC code that cares about FP64, and as a budget GPU-compute card. It is not the pick for modern deep-learning training (no Tensor Cores).

---

Which to buy, by workload

WorkloadBest value pick
LLM / transformer trainingA100 (80GB for big models)
Smaller model trainingV100
High-volume inferenceA100, or P40 for VRAM-on-a-budget
VDI / virtual workstationsP40 (24GB)
FP64 / HPCP100

---

FAQ

Are these new or refurbished? Most datacenter GPUs in the channel are off-lease/decommissioned. See Refurbished GPUs for AI for what to verify.

Do I need NVLink? Only for tight multi-GPU training. Single-card or loosely-coupled inference is fine on PCIe.

What about power/cooling? All ~250W, passively cooled — they need a GPU-capable server. See the Enterprise & AI GPU guide.

---

Pro Disk Network is an independent reseller of genuine NVIDIA datacenter GPUs (not affiliated with NVIDIA); cards are tested at full load before shipping. Overview: Enterprise & AI GPUs.

Sources: NVIDIA A100 / V100 / P40 / P100 datasheets (linked above).

Part of

AI / ML Infrastructure Hub

View all 36 pages →

Server CPUs and GPUs for AI/ML and HPC — Intel Xeon, AMD EPYC, NVIDIA A100/H100, AMD MI300, RTX 6000.