AMD Unveils The VEGA-Based Radeon Instinct MI25 GPU

AMD has just announced the Instinct MI25 VEGA-based graphics solution which is designed for complex deep learning situations. The product has 12.5 TFLOPs of compute power in 32-bit operations and 25 FLOPs using FP16 calculations. This particular graphics solution is compared against the Titan X Pascal and offers 50% extra performance versus Fiji. Sadly, there are no details regarding the clock speed or memory configuration although it’s interesting to see the GPU drawing less than 300 watts. According to AMD’s internal testing, the MI25 comprehensively beats the Titan X Pascal during DeepBench GEMM.

Here we can see more details regarding AMD’s Naples platform with lower latency architecture and a high-density footprint:

VideoCardz has listed other intriguing information about the MI25 and its potential in server racks and instances where deep learning is a major focus:

Even though the information so far is relatively sparse, AMD looks like they’re making significant progress when it comes to deep learning. Of course, it provides an insight into VEGA’s performance which bodes pretty well for the consumer products. Perhaps, AMD will release more details about VEGA during today’s New Horizon event and it’s certainly worth tuning in to hear more. Overall, I’m really impressed with the deep learning performance data so far and the range’s varying form factors.