NVIDIA tensor cores

Overview and evaluation of tensor cores

Tensor cores are firstly embedded in Volta architecture (CUDA 9), in Turing they are using the second generation, in Ampere the 3rd generation and 4th generation in Hopper

Key take aways
1

In Volta GV100, each Tensor Core performs 64 floating point FMA operations per clock, and eight Tensor Cores in an SM perform a total of 512 FMA operations (or 1024 individual floating point operations) per clock.

2

Tensor Cores provide up to 12x higher peak TFLOPS on Tesla V100 that can be applied to deep learning training compared to using standard FP32 operations on P100

3

Each Tensor Core operates on a 4x4x4 matrix and performs the following operation: D = A×B + C, where A/B can be FP16 and C/D can be FP16/FP32

4

Pasted image 20230630152326.png
Inputs are FP16; Full precision produce; FP32 accumulation.

Manipulate methods
  1. CUDA C++ API. The API exposes specialized matrix load, matrix multiply and accumulate, and matrix store operations to efficiently use Tensor Cores from a CUDA-C++ program.
  2. cuBLAS and cuDNN libraries.
Key take aways
  1. Add new INT8 and INT4 for inferencing.
  2. A new technique called Deep Learning Super Sampling (DLSS) is powered by Tensor Cores

Tensor Cores in Ampere

  1. Introduce hardware support for processing matrices with specific sparsity patterns at up to 2x throughput, by skipping the zero-valued elements.
  2. Add new precision mode TF32 and BF16.
  3. Note that GA10x GPUs do not include Tensor Core acceleration for double-precision (FP64) operations, as provided in A100.
1

For HPC, the A100 Tensor Core includes new IEEE-compliant FP64 processing that delivers 2.5x the FP64 performance of V100.

  • TF32 Tensor Core instructions which accelerate processing of FP32 data
  • IEEE-compliant FP64 Tensor Core instructions for HPC
  • BF16 Tensor Core instructions at the same throughput as FP16
2

Each of the A100 Tensor Cores can execute 256 FP16 FMA operations per clock, allowing it to compute the results for an 8x4x8 mixed-precision matrix multiplication per clock

3

A100 accelerates tensor math with TF32 while supporting FP32 input and output data (right), enabling easy integration into DL and HPC programs and automatic acceleration of DL frameworks.

FP64 facility

Comparison of these three generations

Important

$

Dissecting Tensor Cores via Microbenchmarks- Latency, Throughput and Numerical Behaviors

  1. Furthermore, Ampere Architecture redesigns the micro-architecture of Tensor Cores. Unlike Volta and Turing Architecture which have eight Tensor Cores per SM and each Tensor Core performs a 4×4×4 MM (i.e. m = n = k = 4), there are only four Tensor Cores per SM and each Tensor Core performs an 8×4×8 MM.

Pasted image 20230703173135.png
- From [[Dissecting the NVidia Turing T4 GPU via Microbenchmarking.pdf]]

Ways to manipulate tensor cores

  1. High-level libraries like cuBLAS and cuDNN
  2. CUDA C++ API (WMMA)
  3. PTX
  4. Device

Example of 1, 2 in Programming Tensor Cores in CUDA 9

Example of 3, 4 in [[NVIDIA cutlass, PTX to program tensor cores.pdf]]

Use CUDA C++ API (NVIDIA:wmma)

Programming tensor cores using nvcuda-wmma