![GPUs
• If a company knows its AI strategy includes advanced neural networks
and AI algorithms, choosing a CPU chip processor would require
numerous accelerators to match the multicore, fast processing speed
of a GPU.
• The current industry standard is to build your AI system using GPUs.
GPUs are optimized to render graphics and images but have the
speed and computational power to support AI, machine learning and
neural network development. Nvidia, Intel and Arm are some of the
primary GPU vendors.
• GPUs are very effective for training. If you have enough deep learning
[models that you're training], use an architecture that [is] designed for
that: GPUs.](https://image.slidesharecdn.com/aihardware-191207183817/85/AI-Hardware-7-320.jpg)
AI Hardware
- 1. AI Hardware Artificial Intelligence Presentation
- 2. Introduction • The field of artificial intelligence (AI) has witnessed tremendous growth in recent years with the advent of Deep Neural Networks (DNNs) that surpass humans in a variety of cognitive tasks. • The algorithmic superiority of DNNs comes at extremely high computation and memory costs that pose significant challenges to the hardware platforms executing them. • Currently, GPUs and specialized digital CMOS accelerators are the state-of-the-art in DNN hardware. • However, the ever-increasing complexity of DNNs and the data they process have led to a quest for the next quantum improvement in processing efficiency.
- 3. Introduction • The AI hardware team is exploring new devices, architectures and algorithms to improve processing efficiency as well as enable the transition from Narrow AI to Broad AI. • Approximate computing, in-memory computing, machine intelligence and quantum computing are all part of the computing approaches being explored for AI workloads. “IBM Research is developing new devices and hardware architectures to support that support the tremendous processing power and unprecedented speed that AI requires to realize its full potential.”
- 4. Why AI Hardware? • Every enterprise is trying to implement AI and machine learning. But, before AI, before clean data and before platform comparison, enterprises need to find the best hardware to support AI. • Companies need to evaluate hardware before considering how to utilize AI software and products. • This hardware evaluation needs to include memory and processing requirements and whether conventional CPUs or more specialized GPUs and AI chips are necessary. • Your initial choice in hardware -- and, most importantly, your chip selection -- will branch out and affect your long-term AI strategy.
- 5. Choosing the Right Chip • If there are roles for both CPUs and GPUs, developers may wonder where each fits as they build their AI-optimized hardware infrastructure. • By AI experts it’s encouraged to evaluate your machine learning and AI goals and building your hardware infrastructure to match. If you want a basic AI strategy, choosing CPUs with accelerators and software as needed could be enough to power your machine's general-purpose computing and a light AI workload. • If you want to develop deep learning AI, choosing GPUs is likely going to deliver the best results. GPUs, as they are fairly new to enterprise AI use cases, can get pricey. But, if you want to train your own deep learning and AI models, the processing efficiency and speed of GPUs might be worth it.
- 6. CPUs • CPUs power most a machine's basic computing tasks. • Before specialized processing units, like GPUs and tensor processing units, gained traction in the machine learning field, CPUs did most of the heavy lifting. • However, one of the most important factors in choosing AI-optimized hardware is processing speed. A CPU-based machine can take longer than a GPU-based one to train AI models because it has fewer cores and doesn't take advantage of parallel processing the way GPUs do. • If you want to do light deep learning or you want to do a mix of deep learning and general purpose, CPU is the best machine to do that.
- 7. GPUs • If a company knows its AI strategy includes advanced neural networks and AI algorithms, choosing a CPU chip processor would require numerous accelerators to match the multicore, fast processing speed of a GPU. • The current industry standard is to build your AI system using GPUs. GPUs are optimized to render graphics and images but have the speed and computational power to support AI, machine learning and neural network development. Nvidia, Intel and Arm are some of the primary GPU vendors. • GPUs are very effective for training. If you have enough deep learning [models that you're training], use an architecture that [is] designed for that: GPUs.
- 9. Techniques for Better AI chips 1. Reduced Precision In order to increase accuracy, computers have gone from 16, to 32, to 64 bits, but it turns out that neural nets don’t need that level of accuracy. 2. In-memory computation using Phase Changing Memory. The awesome part about this technology though is its ability to perform simple arithmetic functions locally, rather than sending data to the processor then back. 3. Analog computation (yeah, that old tech’s coming back) It’s ability to lower energy usage and easily perform arithmetic makes it ideal for AI algorithms.
- 10. Companies Working on AI Specific Hardware • Google’s tensor processing units (TPU), which they offer over the cloud and costs just a quarter compared to training a similar model on AWS. • Microsoft is investing in field programable gate arrays (FGPA) from Intel for training and inference of AI models. • Intel has a bunch of hardware for specific AI algorithms like CNN’s. They’ve also acquired Nervana, a startup working on AI chips, with a decent software suite for developers as well. • IBM’s doing a lot of research into analog computation and phase changing memory for AI. • Nvidia’s dominated the machine learning hardware space because of their great GPU’s, and now they’re making them even better for AI applications, for example with their Tesla V100 GPU’s.
- 11. Moore’s law may be coming to an end, but hardware must keep improving to keep up with the increasing demands of software. Thanks