NVIDIA Tesla P40
vs
NVIDIA Tesla P100 PCIe 16 GB

vs
NVIDIA Tesla P40 vs NVIDIA Tesla P100 PCIe 16 GB graphics card comparison

GPU Comparison Result

NVIDIA Tesla P40 vs Tesla P100 PCIe 16 GB: Which is Better for AI, Rendering, and Computation

NVIDIA Tesla P40 and Tesla P100 belong to the Pascal generation but were designed for different tasks. The P40 is aimed at inference, virtualization, and single-precision calculations. The P100 is a specialized HPC accelerator featuring fast HBM2 memory, high FP64 performance, and accelerated FP16 computations.

Therefore, the larger number of CUDA cores in the P40 does not inherently make it a more powerful card in all scenarios.

Key Differences

Feature Tesla P40 Tesla P100 PCIe 16 GB
CUDA Cores 3840 3584
Video Memory 24 GB GDDR5 16 GB HBM2
Bandwidth 346 GB/s 732 GB/s
FP32 about 12 TFLOPS about 9.3 TFLOPS
FP64 severely limited about 4.7 TFLOPS
FP16 no noticeable speedup about 18.7 TFLOPS
INT8 up to 47 TOPS not a primary mode
TDP 250 W 250 W

Where the Tesla P40 Excels

The Tesla P40 offers 3840 CUDA cores, 24 GB of memory, and higher peak FP32 performance.

The main practical advantage is the volume of video memory. The additional 8 GB allows for larger models, scenes, and datasets to be accommodated. This is beneficial for rendering, image processing, and running neural networks that do not fit into 16 GB.

The P40 also supports INT8 acceleration up to 47 TOPS. Therefore, it is better suited for inference tasks where a pre-trained model processes a large number of requests.

Strengths of the P40:

  • 24 GB of video memory;
  • high FP32 speed;
  • INT8 acceleration;
  • rendering large scenes;
  • inference without Tensor Cores.

Why the P100 is Faster in HPC

The Tesla P100 uses HBM2 memory with a bandwidth of 732 GB/s-more than twice that of the P40. This is particularly important in tasks where the GPU is constantly transferring large amounts of data between memory and computing units.

High bandwidth is beneficial in linear algebra, simulations, engineering calculations, and scientific applications. In algorithms limited by memory speed, the P100 may outperform the P40 despite its lower peak FP32 performance.

The main advantage of the P100 is its full double-precision speed. It delivers about 4.7 TFLOPS of FP64, whereas the P40's performance in this mode is severely curtailed. In scientific and engineering programs, this difference can be measured in multiples rather than percentages.

The P100 is significantly preferable for:

  • physical and climate simulations;
  • computational chemistry;
  • engineering analysis;
  • financial modeling;
  • scientific calculations in FP64;
  • algorithms sensitive to memory speed.

What to Choose for Neural Networks

Both cards were released before Tensor Cores, so they significantly lag behind modern accelerators in current AI workloads.

The P100 is better suited for model training due to its fast HBM2 and high FP16 speed. The P40 is more logical to use for inference, especially when 24 GB of memory or INT8 is important.

However, the outcome greatly depends on the software. Modern frameworks are increasingly optimized for Tensor Cores, BF16, and newer architectures, so the potential of Pascal is not realized in all tasks.

Rendering and CUDA

In rendering and general CUDA computations, the P40 often appears more attractive. It excels where performance is defined by the number of FP32 operations and the available memory.

The P100 can pull ahead if the application is constrained by data transfer rates. Therefore, the final decision depends on the specific workload: the P40 is stronger in compute-intensive FP32 tasks, while the P100 is better in memory-intensive algorithms.

Both cards use passive cooling and require a strong directed airflow. In a typical case without separate cooling, they overheat quickly. There are also no video outputs, making these accelerators impractical for a gaming computer.

Is It Worth Buying Them Today?

Both models are outdated and do not support Tensor Cores. They should primarily be considered on the secondary market and only for specific tasks.

The Tesla P40 is interesting for its 24 GB of video memory, inexpensive inference, and rendering. The Tesla P100 retains value where FP64, FP16, and high HBM2 bandwidth are required.

When purchasing, it is also necessary to consider the condition of the card, compatibility with drivers, power requirements, and the need for separate cooling.

Conclusion

The Tesla P40 is better suited for inference, rendering, and tasks that require 24 GB of video memory. Its advantages include high FP32 speed, INT8 support, and greater memory capacity.

The Tesla P100 PCIe 16 GB is significantly stronger in HPC, FP64, FP16, and workloads that depend on memory bandwidth. HBM2 and full double-precision units make it a specialized accelerator for scientific and engineering calculations.

P40 is the choice for memory capacity and FP32/INT8. P100 is a tool for tasks where calculation accuracy and data transfer speed are more critical.

Advantages

  • Higher Boost Clock: 1531MHz (1531MHz vs 1329MHz)
  • Larger Memory Size: 24GB (24GB vs 16GB)
  • More Shading Units: 3840 (3840 vs 3584)
  • Newer Launch Date: September 2016 (September 2016 vs June 2016)
  • Higher Bandwidth: 732.2 GB/s (694.3 GB/s vs 732.2 GB/s)

Basic

NVIDIA
Label Name
NVIDIA
September 2016
Launch Date
June 2016
Professional
Platform
Professional
Tesla P40
Model Name
Tesla P100 PCIe 16 GB
Tesla Pascal
Generation
Tesla
1303MHz
Base Clock
1190MHz
1531MHz
Boost Clock
1329MHz
PCIe 3.0 x16
Bus Interface
PCIe 3.0 x16
11,800 million
Transistors
15,300 million
240
TMUs
?
Texture Mapping Units (TMUs) serve as components of the GPU, which are capable of rotating, scaling, and distorting binary images, and then placing them as textures onto any plane of a given 3D model. This process is called texture mapping.
224
TSMC
Foundry
TSMC
16 nm
Process Size
16 nm
Pascal
Architecture
Pascal

Memory Specifications

24GB
Memory Size
16GB
GDDR5X
Memory Type
HBM2
384bit
Memory Bus
?
The memory bus width refers to the number of bits of data that the video memory can transfer within a single clock cycle. The larger the bus width, the greater the amount of data that can be transmitted instantaneously, making it one of the crucial parameters of video memory. The memory bandwidth is calculated as: Memory Bandwidth = Memory Frequency x Memory Bus Width / 8. Therefore, when the memory frequencies are similar, the memory bus width will determine the size of the memory bandwidth.
4096bit
1808MHz
Memory Clock
715MHz
694.3 GB/s
Bandwidth
?
Memory bandwidth refers to the data transfer rate between the graphics chip and the video memory. It is measured in bytes per second, and the formula to calculate it is: memory bandwidth = working frequency × memory bus width / 8 bits.
732.2 GB/s

Display and Media

No outputs
Outputs
No outputs

Theoretical Performance

147.0 GPixel/s
Pixel Rate
?
Pixel fill rate refers to the number of pixels a graphics processing unit (GPU) can render per second, measured in MPixels/s (million pixels per second) or GPixels/s (billion pixels per second). It is the most commonly used metric to evaluate the pixel processing performance of a graphics card.
127.6 GPixel/s
367.4 GTexel/s
Texture Rate
?
Texture fill rate refers to the number of texture map elements (texels) that a GPU can map to pixels in a single second.
297.7 GTexel/s
183.7 GFLOPS
FP16 (half)
?
An important metric for measuring GPU performance is floating-point computing capability. Half-precision floating-point numbers (16-bit) are used for applications like machine learning, where lower precision is acceptable. Single-precision floating-point numbers (32-bit) are used for common multimedia and graphics processing tasks, while double-precision floating-point numbers (64-bit) are required for scientific computing that demands a wide numeric range and high accuracy.
19.05 TFLOPS
367.4 GFLOPS
FP64 (double)
?
An important metric for measuring GPU performance is floating-point computing capability. Double-precision floating-point numbers (64-bit) are required for scientific computing that demands a wide numeric range and high accuracy, while single-precision floating-point numbers (32-bit) are used for common multimedia and graphics processing tasks. Half-precision floating-point numbers (16-bit) are used for applications like machine learning, where lower precision is acceptable.
4.763 TFLOPS
11.995 TFLOPS
FP32 (float)
?
An important metric for measuring GPU performance is floating-point computing capability. Single-precision floating-point numbers (32-bit) are used for common multimedia and graphics processing tasks, while double-precision floating-point numbers (64-bit) are required for scientific computing that demands a wide numeric range and high accuracy. Half-precision floating-point numbers (16-bit) are used for applications like machine learning, where lower precision is acceptable.
9.335 TFLOPS

Miscellaneous

30
SM Count
?
Multiple Streaming Processors (SPs), along with other resources, form a Streaming Multiprocessor (SM), which is also referred to as a GPU's major core. These additional resources include components such as warp schedulers, registers, and shared memory. The SM can be considered the heart of the GPU, similar to a CPU core, with registers and shared memory being scarce resources within the SM.
56
3840
Shading Units
?
The most fundamental processing unit is the Streaming Processor (SP), where specific instructions and tasks are executed. GPUs perform parallel computing, which means multiple SPs work simultaneously to process tasks.
3584
48 KB (per SM)
L1 Cache
24 KB (per SM)
3MB
L2 Cache
4MB
250W
TDP
250W
1.3
Vulkan Version
?
Vulkan is a cross-platform graphics and compute API by Khronos Group, offering high performance and low CPU overhead. It lets developers control the GPU directly, reduces rendering overhead, and supports multi-threading and multi-core processors.
1.3
3.0
OpenCL Version
3.0
4.6
OpenGL
4.6
6.1
CUDA
6.0
12 (12_1)
DirectX
12 (12_1)
8-pin EPS
Power Connectors
1x 8-pin
96
ROPs
?
The Raster Operations Pipeline (ROPs) is primarily responsible for handling lighting and reflection calculations in games, as well as managing effects like anti-aliasing (AA), high resolution, smoke, and fire. The more demanding the anti-aliasing and lighting effects in a game, the higher the performance requirements for the ROPs; otherwise, it may result in a sharp drop in frame rate.
96
6.7
Shader Model
6.4
600W
Suggested PSU
600W

Benchmarks

FP32 (float) / TFLOPS
Tesla P40
11.995 +28%
Tesla P100 PCIe 16 GB
9.335
Blender
Tesla P40
802
Tesla P100 PCIe 16 GB
1200 +50%
OctaneBench
Tesla P40
163
Tesla P100 PCIe 16 GB
217 +33%