The evolution of XPU and the critical role of software
Contributors: Jason Lawley, Director Technology Leadership Marketing and Chandan Damannagari, Software Technology Expert at Intel
The long-predicted data explosion is here, and one thing is clear — the days of a single architecture to process data are behind us. The scale of the explosion can be seen in the ~1GB/day of data generated by an individual internet user, 50GB/day by a smart car, 3TB/day by a smart hospital, and 50PB/day by a smart city¹. Not only is the amount and diversity of data today staggering, but also where, when, and how it’s processed is rapidly evolving. Whether a workload is run on the edge or in the cloud, or if it is an AI workload or a storage workload, having the right architecture and software to utilize that architecture can make or break a project. At Intel, we are on a journey to provide every developer with the right combination of hardware architectures and software to build the optimal solution for their particular use case. We are doing this both through our XPU strategy and by enabling an open, standards-based cross-architecture programming model called oneAPI. But what is an XPU? It’s not a new processor or product, but a portfolio of architectures (CPU, GPU, FPGA and other accelerators). This blog explains the concept of XPU and how software is critical to the success of Intel’s XPU future.
THE NEED FOR DIVERSE COMPUTING ARCHITECTURES
If raw data is the proverbial ore, then data processing is how gold is extracted from that raw data. Processing the most amount of data with the lowest power has always been the goal of most compute architectures. In recent years, an important aspect to this gold analogy is time. The value of data is fleeting and often without processing that data at the right time, it becomes useless. Think of things like real-time ad services, automotive sensing, or live video transcoding. They are all sensitive to how long it takes to process the data. With this in mind, there are a few key characteristics that developers look for in their hardware and that computer architects strive to deliver. They are high performance, low latency, and low power.
- Performance — The amount of data that can be processed per unit of time. This might be video analytics or encoding, searching a database, or a variety of other data processing activities. Performance might be measured in GByte/s or images/s.
- Latency — The length of time required to process the data and to get a result. In some applications, latency is not an issue while in others like real-time digital assistants or automotive video processing, it is a critical factor. Latency is often measured in milli-seconds or even micro-seconds.
- Power — The amount of power needed to process the data. Power is important both at the edge and in the cloud. As devices move to the edge, where only batteries or other limited power sources are available, power consumption rates are increasingly critical. Even in the data center, cooling and electricity costs constitute a significant percentage of total operating costs. Power is often measured in watts or joules.
Unfortunately, there is no longer a one size fits all architecture for the diversity of today’s data. Specialized architectures have shown that they can optimize performance, power, or latency as needed, but getting the right balance for a gaming GPU is vastly different than what is needed for a battery-operated security camera. Architectures also continue to evolve as the types of data and workloads diversify. The ability to create a broad set of architectures targeting an even broader set of workloads is the essence of Intel’s XPU strategy.
XPU AND oneAPI — TWO PIECES OF THE SAME HETEROGENEOUS COMPUTING PUZZLE
Having the right architecture for the right application is critical to our customers success. CPUs and the vast applicability of their general-purpose computing capabilities are central to our XPU strategy. We also offer an already successful lineup of GPUs, FPGAs, VPUs, and other architecture types to meet the processing needs of the huge variety of different workloads that exist today and into the future. We classified the range of architectures into Scalar, Vector, Matrix and Spatial as we explained in Harnessing the Power of a Heterogenous Computing Future. Our mission is to offer the broadest suite of XPU silicon in the industry, to enable our customers and communities in this next era of computing.
Having multiple architectures is just one piece of the puzzle though. The bigger challenge is to quickly determine which architecture makes the most sense for a desired workload. Often, the best way to evaluate a workload is to try that workload on the different architectures and compare. The migration process between architectures can be time consuming and error prone. Enter oneAPI, a cross-industry, open, standards-based unified programming model. Using oneAPI, developers are given the freedom to choose the best architectures to meet their workload/application needs, while using a single code base across multiple architectures. This allows developers to maximize cross-architecture performance and minimize development costs, while also giving them the freedom to expose and exploit cutting-edge features across diverse XPU architectures. By abstracting hardware heterogeneity with oneAPI, Intel is driving a new era of accelerated computing and helping to best position our customers to harness the power of XPUs now and into the future.
This is the first blog in a series of blogs where we will explore the scalar, vector, matrix and spatial architectures in greater detail. You can also check out this blog from James Reinders to learn more about how DPC++ and oneAPI can help unleash the power of XPUs. We have made significant strides on our journey to become an XPU company and we continue to strive to provide class-leading solutions to harness the limitless potential of the world’s data.
Notes and Disclaimers
© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subs. Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex
