Over the past few days, AMD dedicated plans for the conclusion of new products based on advanced graphics architecture. The new architecture, in particular - VEGA, also promise to be AMD's ticket to the world of high performance computing, where for a long time NVIDIA feel good with Tesla products. So, yesterday, AMD unveiled a series of accelerators Radeon Instinct headed by the most productive solution MI25. Future MI25 adapter on the GPU generation VEGA provide the speed
calculations on operations with half-precision FP16 at 25 teraflops. For comparison, NVIDIA GP100 processor architecture Pascal performs FP16 performance of 21.2 teraflops. In other words, AMD solution will be slightly better competitor actual offer.
In closed Coduri Raja demonstration showed Module VEGA CUBE FP16 with a peak performance of 100 teraflops. If we assume that the performance of one unit is 25 teraflops in the face of MI25, the cube peak consumption can be up to 1.2 kW.
Unfortunately, the quality of the photos may not accurately estimate the size of the VEGA GPU and even bridge with the GPU and memory HBM. A rough comparison with the image of the bridge with the GPU NVIDIA GP100 may hint at approximately the same dimensions of the decisions of both companies.
Related Products :
|