Tesla P40

NVIDIAdatacenter

Overview
Architecture	Pascal (2016)
Launch date	Sep 2016 (datasheet footer dated Sep16)
Compute capability	6.1
Memory
VRAM	24 GB GDDR5
Memory bandwidth	346 GB/s
ECC	Yes
Compute — vector
FP32	12 TFLOPS
Compute — matrix / tensor
INT8	47 TOPS
Cores & clocks
Streaming Multiprocessors	30
Shader cores	3,840
Board & system
TDP	250 W
Form factor	Full-Height, Dual-Slot (4.4" H x 10.5" L)
Cooling	Passive
PCIe	PCIe 3.0 x16

Compatible CUDA Toolkit versions

PyTorch on the Tesla P40

Check against your driver and Python version →

A plain pip install torch won't work on this card. It installs PyTorch 2.14.0's CUDA 13.0 build, which has no code for compute capability 6.1, so it fails at runtime with no kernel image is available for execution on the device. Install the CUDA 12.6 build instead:

pip install torch==2.14.0 --index-url https://download.pytorch.org/whl/cu126

PyTorch	CUDA builds that run on this card	Plain `pip install torch`
2.14.0	12.6	doesn't (CUDA 13.0 build)
2.13.0	12.6	doesn't (CUDA 13.0 build)
2.12.0	12.6	doesn't (CUDA 13.0 build)
2.11.0	12.6	doesn't (CUDA 13.0 build)
2.10.0	12.6	doesn't (CUDA 12.8 build)
2.9.0	12.6	doesn't (CUDA 12.8 build)
2.8.0	12.6	doesn't (CUDA 12.8 build)
2.7.0	11.8, 12.6	works
2.6.0	11.8, 12.4, 12.6	works
2.5.0	11.8, 12.1, 12.4	works
2.4.0	11.8, 12.1, 12.4	works
2.3.0	11.8, 12.1	works
2.2.0	11.8, 12.1	works
2.1.0	11.8, 12.1	works
2.0.0	11.7, 11.8	works
1.13.0	11.6, 11.7	works
1.12.0	10.2, 11.3, 11.6	—
1.11.0	10.2, 11.3	—

From each build's compiled architecture list (Linux x86_64 wheels). Your driver must also support the CUDA version — see compatible CUDA versions.

Libraries on the Tesla P40

Whether each library's own requirements cover this card. Each row links to the version and source it was checked against on library support.

Library	Supported	Details
vLLM 0.30.0	No	needs compute capability 7.5+
FlashAttention-2 2.8.3.post1	No	no build for compute capability 6.1
FlashAttention-3 main (2026-09-27)	No	no build for compute capability 6.1
bitsandbytes 0.50.2	Partly	no LLM.int8() (needs 7.5+)
TensorRT 11.3.0	No	needs compute capability 7.5+
llama.cpp master (2026-09-27)	Yes

Native low-precision formats

Whether the card's matrix hardware runs these formats natively. Without native support a format can still work, but through slower emulation or conversion, not at the card's rated speed.

BF16	FP8	FP4
No	No	No

Source: NVIDIA TensorRT 11.3.0 support matrix (hardware precision table); pre-Turing architectures predate all three formats.

Added 2026-09 — the most-discussed used budget card for local LLM inference (24GB for a fraction of an RTX 3090's price), and compute capability 6.1 means it lost CUDA support in 13.0 along with the rest of Pascal; CUDA 12.x is the last toolkit that can target it. Figures (architecture, 12 TFLOPS FP32, 47 TOPS INT8, 24GB, 346 GB/s, PCIe 3.0 x16, dual-slot full-height form factor, 250W max power, ECC) sourced directly from NVIDIA's official Tesla P40 datasheet (images.nvidia.com/ content/pdf/tesla/184427-Tesla-P40-Datasheet-NV-Final-Letter-Web.pdf, footer "Sep16"), read via pdftotext. NVIDIA marks both the FP32 and INT8 figures "With Boost Clock Enabled". The datasheet does not list a memory type; GDDR5 is the P40's documented memory (the P40 is a GP102 part, not the HBM2 GP100 used by the P100). shaderCoreCount (3840) and computeUnitCount (30 SMs) are the full GP102 die at 128 FP32 cores per SM — consistent with the datasheet's 12 TFLOPS: 3840 cores x 2 FLOP x ~1.53 GHz boost ≈ 11.8 TFLOPS. Not on the datasheet itself. fp16TFLOPSVector is deliberately omitted, and this matters more than any other number for this card: unlike the P100 (compute capability 6.0, fast packed FP16), GP102-based cards like the P40 execute FP16 at a tiny fraction of their FP32 rate. In practice, LLM inference on a P40 runs in FP32 or via integer-quantized kernels, not FP16. No Tensor Cores (introduced with Volta), so all matrix fields except int8TOPS are correctly omitted; int8TOPS is Pascal's DP4A instruction throughput on the CUDA cores, not Tensor-Core throughput.

← All GPUs · Compare GPUs →