This article has been reproduced by the Supernet. Other media reprints are subject to the approval of Supernet.
NVIDIA's performance has risen in the past two years, and stocks have also made rockets generally rise from more than 30 US dollars two years ago to 250 US dollars. One important reason is that NVIDIA is leading in emerging fields such as deep learning and AI artificial intelligence. CEO Huang Renxun has been Some media are called 'AI Godfather'.
In contrast, AMD has struggled in the game market for the past two years, but AMD CTO said in an interview with the media a few days ago. AMD has a comprehensive machine learning strategy that will provide an AI engine for the data center and edge computing markets.
EEtimes recently interviewed Mark Papermarster, AMD's chief technology officer, about AMD's AI strategy, 7nm Vega, Zen 2 processor and future process choices.
However, everyone can rest assured that Papermarster did not mention any new products, new technologies and new processes. The company's secrets are strictly adhered to. The more important content is to confirm that AMD will make great achievements in the AI market.
Mark Papermarster claims that AMD has a comprehensive machine learning strategy that will provide AI engines for data centers, edge computing, and more.
But the specific details are gone. What kind of products will AMD provide for the AI market? What optimizations will be made for AI, deep learning, etc? These questions are not to be said.
EEtimes mentioned that AMD launched GPU acceleration for deep learning as early as 2016. This is the Radeon Instinct accelerator card. The Radeon Instinct MI25 uses a new generation Vega architecture with 64 NCU units, which translates to 4096 SP streams. With 16GB of HBM2 memory, the memory bandwidth is up to 484 GB/s, and the semi-precision floating-point performance has been greatly improved, reaching 24.6TFLOP, and the single precision is also 12.3TFLOPS.
After that, Google introduced the TensorFlow processor for deep learning, adding MAC (multiply accumulating unit) to the hardware to accelerate deep learning. In May 2017, NVIDIA introduced Volta graphics card, which is the first to join the MAC unit. The graphics card, called the Tensor unit by NVIDIA.
Intel earlier this year plans to transfer its acquired Movidius accelerator unit to a PC platform running Windows ML. Analysts believe that Intel will eventually integrate the Movidius unit into the CPU.
Mark Papermarster did not confirm whether AMD will release the MAC unit on the Vega to be released at the end of this year or the 7nm Zen2 processor released early next year, but he said that the upcoming Vega GPU supports 16-bit floating-point accidental operations.
This is actually not news. Before the release, AMD mentioned that 7nm Vega has added new deep learning instructions and supported 8-bit operations.
Now major companies are researching algorithms that accelerate deep learning. ARM has added 8-bit computing to ML Core, NVIDIA is already working on 2-bit operations, and Intel is planning unit operations.
