For years, manufacturers of mobile processors have focused on optimizing designs for maximum performance within limited power budgets, storage space, and bandwidth. In the past, clearly these considerations have not been taken seriously in markets such as the data center or personal computer (PC). Today, the transformation of the traditional data center and PC market is quietly happening – changing the rules of processor design, forcing developers to rethink their chip architecture to achieve higher performance per watt.

Mobile Processor Design Principles Applied to PCs and Data Centers

Today, more and more cloud gaming, data mining, artificial intelligence/data analysis and high performance computing are implemented in the cloud. While the requirements of these applications vary, they are all the same in terms of the ever-increasing computational requirements.

Data centers cannot keep up with this demand by constantly expanding their physical footprint. To keep operating expenditures (OpEx) within acceptable limits and achieve Net Zero goals, enterprises need to increase computing density in a limited space for higher computing performance. Processing elements such as graphics processing units (GPUs), central processing units (CPUs), artificial intelligence (AI) accelerators, etc. must achieve the highest performance within the smallest power/cooling and area budgets. To this end, a processor design that follows mobile design principles is an ideal starting point.



Big tech firms differentiate with custom chip designs

With Moore's Law coming to an end, it's no longer possible for the industry to see performance improvements every two years. In this context, companies have joined this design "competition" one after another, vying to create the best user experience with the best chips.

The world's largest technology companies have long understood this and come prepared. They are focusing on designing their own custom chips for use in consumer products, PCs or data centers, etc. These companies have moved from off-the-shelf chips to custom chips, hoping to gain an edge by taking better control of their designs. So we see Amazon investing in Graviton CPU designs and Google launching TPU-centric Tensor CPUs. Apple's M1 processor will bring Mac computers a chip optimized for mobile design principles, providing higher integration and better performance per watt.

OEM alternative

The challenge for original equipment manufacturers (OEMs) without in-house hardware and software design teams and yet to develop custom chips is how to differentiate their designs against highly optimized architectures. Most of the chips these OEMs use are readily available, potentially putting them at a disadvantage. Many chips designed for PCs and data centers are "brute force" solutions that, while providing the required performance, are often too power hungry, too memory/bandwidth intensive, and uncompetitive. Additionally, these chips have limitations in the specific software and operating systems available.
SoC suppliers in some mobile markets have begun to enter the data center and PC markets, hoping to carve up part of the market share of existing players, but their number is too small to help OEMs achieve differentiation while innovating and controlling costs. As a result, some industry alternatives are emerging. We have seen many manufacturers are considering CPU solutions based on RISC-V architecture. However, a single CPU design cannot completely solve the fierce competition that OEMs are currently facing. OEMs need to look at the architecture of the entire data center to improve the innovation of the overall solution, thereby increasing their competitiveness.

Scalable heterogeneous architecture is key

Through heterogeneous computing, the CPU, GPU and other computing units are flexibly used to achieve the maximum utilization of hardware to achieve the optimization and improvement of computing performance, while satisfying the optimization of efficiency and power consumption ratio. Heterogeneous computing architecture provides flexible array work solutions for the ever-increasing computing requirements of data centers. At present, many semiconductor manufacturers are researching related products and applications to empower OEM market competitiveness. Imagination, a traditional GPU IP company, launched its CPU product line last year and strengthened the research and development of heterogeneous computing. It intends to provide customers with more complete heterogeneous computing solutions through the optimization and improvement of product portfolio, so as to provide better services. customers to meet the needs of future high-performance computing.

Mobile GPUs lay the foundation

Mobile GPUs are an ideal entry point for creating efficient heterogeneous designs. Rather than trying to force high-end GPUs into mobile power budgets, it would make more sense to apply mobile GPU upgrades to the data center and PC space. Because mobile GPUs are born to be "small and beautiful". Mobile GPU manufacturers have developed many patented technologies to maximize GPU performance and low power consumption. Under the background of extensive single pursuit of high-performance GPUs in data centers and PCs, these technical advantages can give mobile GPU manufacturers more competitive advantages and provide OEMs with more management added value.

When it comes to the patented technology of mobile GPU, I have to mention again Imagination, a veteran company that focuses on GPU design. Compared with its rivals, Imagination has focused its efforts on the GPU field for many years, especially in the more complex GPU rendering field. Imagination is a pioneer developer of many technologies, such as GPU hardware virtualization, block Deferred rendering (TBDR), real-time hardware ray tracing (Ray Tracing), etc. Tiled Deferred Rendering (TBDR) technology divides geometric data into small regions (tiles) and processes them uniformly. Since each tile is rasterized and processed individually, the rendered size is very small, so all data can be kept in fast-running on-chip memory. This technology laid the foundation for the M1's graphics processing.

For application scenarios such as Android cloud games, the data center needs to flexibly handle different game consumption scenarios of multiple users. The method of processing multiple small concurrent workloads on multiple small GPUs is more efficient than using traditional desktop GPUs. The cloud game industry chain is strengthening the development and application of GPU hardware virtualization technology to reduce costs. Mobile GPUs scale up the distributed multi-core mobile GPU architecture, enabling each GPU to support more users while providing greater energy efficiency for many users in the cloud.
Taking Innosilicon as an example, as the industry leader of the first domestic high-end GPU core, based on Imagination's mobile GPU IP, the company has extended the mobile GPU architecture to high-performance server-level hardware, aiming to break the desktop The current pattern of the computer graphics card market. In this high-end market, long dominated by duopoly, no one expected new competitors to emerge, but Innosilicon is offering an alternative using changing market forces and highly scalable and efficient technology.

Adding efficient on-chip AI processing (as shown by the M1) is another opportunity for OEMs. Since on-chip AI processing is not yet standard in PCs, OEMs can leverage this capability to support emerging applications such as super-resolution noise reduction, audio commands, security, and more. Such AI capabilities typically require enormous computing power, and efficient, highly reliable AI inference capabilities can be integrated on SoCs using Neural Network Accelerator (NNA) IP designed based on mobile design principles. In the field of end-to-end AI edge accelerators, compared with other competitors, Imagination's NNA edge accelerator hardware not only inherits the DNA of its GPU design with high performance and low power consumption, but also outperforms the competition in computing fields of different orders of magnitude Excellent performance by opponents.

Designing specialized chips – not just for big tech companies

SoC manufacturers need scalable IP cores based on mobile design principles to create energy-efficient, high-bandwidth, and high-performance designs. With this processor designed for heterogeneous architectures, they can create specialized, efficient new solutions. This can help OEMs to provide highly competitive and differentiated products with a firm grasp on the future direction of the business.

Leave a Reply

Your email address will not be published.