AMD has announced its brand new lineup of Radeon Instinct Deep Learning accelerators. These are the Radeon Instinct MI8, Radeon Instinct MI6, and Radeon Instinct MI25; all of which come with different graphics architectures.
First up is the Radeon Instinct MI8, which is based off of AMD’s Fiji GPU architecture. It features 64 compute units (4096 stream processors), 8.2 TLFOPS of FP16 (half-precision floating point) and FP32 (single-precision floating point) compute performance, 4GB of HBM1 memory, as well as a memory bandwidth of 512GBps.
Next is the Polaris-based Radeon Instinct MI6 Deep Learning accelerator. This card boasts 36 compute units (2304 stream processors) and 5.7 TFLOPs of FP32 and FP16 performance. When it comes to memory, the Radeon Instinct MI6 features 16GB of GDDR5 memory along with bandwidth speeds of about 224GBps.
The last and most powerful Deep Learning accelerator of the three would be AMD’s Radeon Instinct MI25. At the heart of this card is the company’s Vega GPU, which has 64 “next-gen” compute units (4096 stream processors) as well as 12.3 TFLOPS of FP16 and FP32 compute performance. Memory is handled via the HBM2 memory controller, which has a capacity of 16GB and bandwidth speeds of 484GBps.
According to AMD, all three brand new Radeon Instinct Deep Learning accelerators will be made available in the third quarter of 2017. Pricing for each card was not mentioned.