Initiative AMD Radeon Instinct, aimed at accelerating deep learning includes the hardware and software
The event Tech Summit, which last in SONOMA, California, AMD has released another batch of information about the features of future models, including the Zen CPU’s and GPU architecture Vega.
Today, the manufacturer announced strategy to accelerate machine learning — Radeon Instinct. The new initiative includes the AMD hardware in the form of specialized accelerators and software open-source, designed to accelerate deep learning neural networks.
The event was presented three models of accelerators Radeon Instinct, designed for a wide range of machine learning tasks: Radeon Instinct MI6, MI8 and MI25. They are based on GPUs Polaris, Fiji and Vega respectively.
Accelerator Radeon MI6 Instinct built in GPU on the architecture of Polaris. It demonstrates the performance of 5.7 TFLOPS in computing, half precision (FP16), consuming 150 watts. Card is equipped with 16 GB of memory.
The basis of the accelerator GeForce Instinct MI8 is the GPU Fiji Nano. This compact model is characterized by the performance of 8.2 TFLOPS (FP16) with consumption of less than 175 watts. Feature Radeon Instinct MI8 — 4 GB high speed memory HBM.
The last of accelerators the most interesting because it is still not feature high end on the Vega architecture. It consumes up to 300 watts, has a new high-speed cache and better performance.
All of these hardware solutions have passive cooling and are exclusively manufactured by the company AMD.
At the same time they announced the AMD software components: a library MIOpen designed to speed up machine learning tasks and open-source and new software ROCm, optimized for the task deep learning.
Market accelerators Radeon Instinct will appear in the first half of 2017.
Sources: AMD, own correspondent iXBT