NVIDIA
GPUDatacenterReleased Nov 10, 2015

NVIDIA Tesla M40

12 GB GDDR5 · 384-bit · 3,072 cores

The Tesla M40 was NVIDIA's dedicated deep-learning training accelerator, announced in November 2015 as the training counterpart to the inference-oriented Tesla M4. A full GM200 Maxwell GPU provides 3,072 CUDA cores and 12 GB of GDDR5 at 288 GB/s on a passively cooled 250 W dual-slot card built for server airflow. NVIDIA claimed it cut a typical AlexNet training run from ten days on CPUs to 1.2 days. A 24 GB version followed in 2016.

General

Manufacturer
NVIDIA
Model
Tesla M40
Segment
Datacenter
GPU chip
GM200
Architecture
Maxwell
Release date
Nov 10, 2015
Price
Search on Amazon

Memory

Memory size
12 GB
Memory type
GDDR5
Bus width
384-bit
Bandwidth
288 GB/s
Memory clock
1,500 MHz

Clocks & silicon

GPU clock
948 MHz
Boost clock
1,114 MHz
Shader units
3,072 CUDA Cores
Tensor / AI cores
—
RT cores
—
Transistors
8 billion
Technology (nm)
28 nm
TSMC 28 nm
TDP / TBP
250 W

Interface & APIs

Bus interface
PCIe 3.0 x16
SLI / multi-GPU
—
DirectX
N/A (compute-only)
OpenGL
N/A (compute-only)
Vulkan
N/A (compute-only)
Other APIs
CUDA 5.2, OpenCL 3.0, OpenACC

AI support

Precisions
FP8FP16BF16INT8INT4TF32FP64
AI accelerator
None (no matrix units)
Peak AI TOPS
—
Notes
No matrix or tensor hardware. NVIDIA rates the card at 7 TFLOPS single precision and 0.2 TFLOPS double precision with GPU Boost; training runs on the 3,072 CUDA cores in single precision, as Maxwell predates the doubled half-precision rate Pascal introduced.

Related NVIDIA GPUs

All NVIDIA GPUs →

Datacenter gpus from NVIDIA announced around the same time as the NVIDIA Tesla M40. Add any of them to your compare list to see the specs side by side.

NVIDIAgpu

NVIDIA Tesla M4

4 GB GDDR5 · 128-bit · 1,024 cores

Released
Nov 10, 2015
NVIDIAgpu

NVIDIA Tesla M6

8 GB GDDR5 · 256-bit · 1,536 cores

Released
Aug 30, 2015
NVIDIAgpu

NVIDIA Tesla M60

8 GB GDDR5 · 256-bit · 2,048 cores

Released
Aug 30, 2015