Already last year, AMD had first shown their upcoming Radeon Instinct compute accelerators as well as the Radeon Open Compute platform (ROCm). These accelerators are designed to challenge NVIDIA in the deep learning and machine intelligence field of application. At the time, AMD didn’t share much information about the specifications of these accelerators. Nevertheless AMD has finally announced the MI25, which is their new flagship model.


Like the MI6 and the MI8 compute accelerators, the MI25 is a passively-cooled card for deep learning and HPC applications. It packs 64 compute units and 4096 stream processors. It’s capable of hitting 12.3 TFLOPS of FP32 and 24.6 TFLOPS of FP16 performance. At the moment AMD hasn't announced the official GPU's clock rates but we believe that the clocks will be somewhere around 1500 MHz. The MI25 will come with 16GB of HBM2 ECC memory. In combination with the 512-bit memory interface, the MI25 should achieve 484 GB/s of memory bandwidth. AMD is once again keeping secrets about the High Bandwidth Cache Controller (HBCC) and apparently it looks like they don't want to share all the details yet.

Speaking about power consumption, the MI25 will require at maximum 300W. Furthermore AMD is claiming that regarding FP16 energy efficiency AMD is capable of outpacing NVIDIA’s Tesla P100-16 as well as the Tesla P100-SMX2. AMD plans to finally release the Radeon Instinct compute accelerators in the third quarter of this year.







Source: AMD