AMD introduced the machine-learning accelerators Radeon Instinct and revealed their characteristics
At the end of last year, AMD has shown the Instinct accelerators Radeon designed for machine learning tasks. Among them was model MI25 Instinct Radeon-based GPU Vega.
But then we uncovered only a part of the parameters of the two models, but now AMD has re-introduced these accelerators and shared details.
Let’s start with the most interesting: Radeon MI25 Instinct. This map is analogous to the Radeon Vega Frontier Edition. Here is the same — Vega 10 with 4096 stream processors and the same 16 GB of memory HBM2.
Performance is slightly lower: of 12.3 TFLOPS, which corresponds to a GPU frequency of approximately 1500 MHz. Memory bandwidth in this case is 484 GB/s Adapter is characterized by a TDP of 300 watts and is equipped with a passive cooler that is designed for use in ventilated racks. The length of the card is 267 mm.
In the diagrams above you can see the results of the comparison Instinct MI25 Radeon with Nvidia Tesla P100-16.
Model Radeon Instinct MI8 based on Fiji GPU with all active 4096 stream processors. There are only 4 GB of memory HBM, but the throughput is even higher than the previous model: 512 GB/s. Performance of the card reaches to 8.2 TFLOPS while consumption is about 175 watts.
Low card Radeon Instinct MI6 got Polaris 10 GPU with 2304 all active processors. Performance is 5.7 TFLOPS, memory bandwidth is 16 GB, is equal to 224 GB/s, and a TDP of 150 watts. Map differs odnoskatnoj cooling system, while the previous two used dual-slot.