NVIDIA in deep learning
So far, we’ve seen how NVIDIA (NVDA) is moving ahead on its technology roadmap to tap the next big VR (virtual reality) opportunity.
Moving on to the next big trend of deep learning, the company has unveiled its most-awaited next-generation GPU (graphics processing unit) architecture, Pascal, which will house the company’s five technology hallmarks. Deep learning is a computing model that enables computers to find patterns in massive amounts of data and make forecasts.
The company unveiled its first GPU built on Pascal architecture, the Tesla P100, at the GTC 2016 conference. Tesla GPUs serve the data center market.
All you need to know about Pascal P100 GPU
The Pascal P100 GPU is the result of three years of research and $3 billion worth of capital. This is NVIDIA’s first GPU to use five new technologies: Pascal Architecture, HBM 2 (high-bandwidth memory), NVLink interconnect, AI (artificial intelligence) algorithms, and 16 nm (nanometer) FinFET (fin-shaped field effect transistor) technology.
The Pascal P100 GPU is the company’s largest chip, with a size of 600mm^2. The chip houses 150 billion transistors, 5.3 TeraFlops of FP64 compute power, and 4 MB (megabytes) of L2 cache. With NVLink, its speed is five times that of the PCIe 3.0, which is currently in use.
Moreover, the company’s transition to 16nm FinFET will significantly improve Pascal’s power efficiency over the current 28nm Maxwell architecture.
Pascal’s architecture is aimed at accelerating artificial intelligence and deep learning on HPC (high-performance computing). Web service companies Baidu and Google (GOOG) have adopted Pascal’s architecture for their deep-learning initiatives.
Pascal in production
The biggest announcement by NVIDIA CEO Jen-Hsun Huang at the GTC conference was that the Pascal P100 GPU is currently in mass production at Taiwan’s (EWT) Taiwan Semiconductor Manufacturing Company’s (TSM) foundry.
However, Huang didn’t state the launch date of the GPU. This could build pressure on Advanced Micro Devices (AMD), which is due to launch its Polaris GPU. The Polaris GPU will compete head-to-head with Pascal.
NVIDIA plans to launch its P100 servers in 1Q17. We’ll explore this move in the next part of the series.