AMD launched Polaris based graphics processing units earlier this year and its a high time for the company to focus on its next gpu flagship, the Vega that will be released next year but the company’s first ever Vega based gpu accelerator has been emerged online, thanks to Videocardz.com.

The Vega powered GPU is known as the Radeon Instinct MI25 Accelerator, targeted for deep learning applications. The gpu accelerator reportedly outperforms NVIDIA’s Pascal Based Titan X in the benchmarks for deep learning application captured from the slides. The slide also mentions the rest of two accelerators, Polaris based MI6 and Fiji based MI8. The MI8 is designed as a small form factor accelerator.

On technical terms, the MI6 features 5.7 TFLOPs while having a memory bandwidth of 244 GB/s and has a TDP of less than 150W. While the MI8 features 8.2 TFLOPs, having a memory bandwidth of 512 GB/s and has a TDP of less than 175W. Whereas the Vega powered MI25 is not provided with that much information, except the TDP which is less than 300W. Some reports claim that the MI25 features around 25 TFLOPs of FP16 compute performance (the highest for any GPU till date).

It is also unknown whether the gpu accelerators will have FP16 performance or FP32 performance but some experts believe that it will feature the FP16 compute, following the benchmark slides below;

According to these benchmarks, the MI25 is almost two times faster than the Pascal based Titan X. Even the MI8 outperforms the Pascal based Titan X, but these are the first party benchmarks. The AMD Vega powered graphics processing units are expected to arrive in the first half of 2017 and according to the reports, the Vega 10 gpu features two times more GPU cores than the Polaris 10.

See also: First Details on next-Gen AMD VEGA 10, VEGA 20 and VEGA 11

AMD seems to be interested in entering the high-performance computing (HPC) market as the market for artificial intelligence grow. How these gpus will perform is the question time will let us know. Stay tuned for more updates!