Industry Expert Blogs
|
Scaling up vs scaling down. The real scoop on power-efficient GPUs for laptopsKristof Beets - Imagination Technologies Group Ltd.May 25, 2022 |
If you've bought a laptop recently, the chances are that your system is architected differently now than it was even a couple of years ago and designs are continuing to evolve. Laptops, and even PCs, are moving away from the traditional model of discrete components, toward higher levels of integration; much like the SoCs that have been used for many years in mobile. This is becoming more important as usage models change (think more hybrid work scenarios), making the need for energy-efficient laptops that still deliver a high level of performance ever more important.
The traditional desktop component companies no longer have the PC space to themselves and companies with mobile expertise are moving into this area. These companies are also leveraging their expertise at keeping power consumption to a minimum whilst maximising performance. They are doing this by incorporating functionality such as graphics, neural network acceleration, security, and I/O, into a single SoC with a unified memory architecture. As such, the modern computer is moving away from just being a box containing a CPU, memory and plug-in graphics cards inside. Instead, many who might have opted for a traditional desktop will now consider a laptop featuring a single, highly integrated chip, with far fewer system-slowing PCI Express connections. In addition to being lower power and less bandwidth-hungry, such an architecture is easier and less expensive to make.
With this high level of integration, today's high-end computers, be they mini desktops or portable laptops, could soon look very much like smartphone processors, and this is setting the stage for a shake-up in the PC processor game. At Imagination, we've already been helping our customers design products like this for decades.
Laptop GPUs reconsidered
With this in mind, companies that have traditionally supplied GPUs for desktops PCs will have to reconsider their graphics architectures to keep pace with evolving requirements. They can no longer focus on designing the highest performance chips with no concern for power or bandwidth since today's laptops are likely to spend a lot of time unplugged from power.
While some companies recently released GPUs have been designed to deliver better power consumption, a chip with a power consumption figure of say, 25 Watts is still a beast of a processor! A laptop with a GPU consuming this much power requires a large battery and fan, and it will need to be plugged in much of the time.
Compare this to a GPU designed with a mobile-first philosophy, where the focus from the start is on delivering high performance within extremely efficient hardware. Mobile GPUs are designed to excel at only 1 W of power as that is all the thermal budget of a very thin phone allows. It's easier to scale up from a design that has been highly optimised for mobile than it is to try to shoehorn a 'brute force' solution into an efficient power envelope. Instead of taking a GPU designed for a high-end PC and scaling it down, the smarter choice is to start with a GPU that was designed first for mobile and to scale it up.
The importance of scalability
At Imagination we know how to deliver high performance in highly efficient hardware. We have delivered the 3D graphics technology for more than 10 billion mobile devices to-date through our PowerVR graphics. Over the decades, we've developed numerous techniques and technologies to deliver the industry's most highly optimised mobile GPUs. This includes tile-based deferred rendering (TBDR) technology.
The Apple M1 chip is an excellent example for how a mobile-first approach pays dividends and how a fully-integrated SoC can boost performance, using TBDR to improve efficiency. In fact, the M-series chips development started with the A4 chip in the 2010 iPhone 4, as a mobile solution. After more than 10 generations, the A14 chip represented the final step in Apple's project. The A14 has been boosted with an additional two CPU cores (on top of its existing two high-performance and four energy-efficiency cores) as well as four additional GPU cores, resulting in a powerful new M1 architecture. This scalability is only achieved through a mobile-first approach with full integration of all advanced functionality into a single SoC with a unified view of memory.
The latest example of Imagination's expertise in combining performance and efficiency is our highly scalable CXT GPU architecture – the industry's first mobile-optimised ray tracing architecture. In a mobile power envelope, CXT delivers the realistic lighting, reflections and shadow effects that were previously reserved for high-end desktop graphics.
The efficiency and scalability of CXT are apparent when you think about the 25W of power consumed by those new GPUs from traditional PC silicon providers. If instead, you use a GPU based on CXT IP, you can get more performance – 2.5x in fact – within the same power budget. This enables you to offer not only more performance it can also deliver features such as ray tracing, and all within that same power envelope.
This kind of thinking can enable silicon vendors and OEMs to differentiate their products.
Another possibility to create differentiation is by integrating efficient AI processing on-chip. While the M1 provides this capability, it isn't yet the norm for PCs. Such AI capability usually requires massive compute power, but Imagination's neural network accelerator (NNA) IP cores enable integration of highly efficient, highly reliable and silicon-proven AI inferencing capability into an SoC.
The future is heterogeneous
With increasingly diverse workloads, chip vendors need to look at optimisations across the chip that provide the best performance whilst meeting power and efficiency requirements. This means not only choosing highly-optimised GPUs and NNAs, but also integrating these processors with highly-efficient CPUs and other compute elements in flexible heterogeneous arrays.
Imagination's Catapult RISC-V CPU is designed to work in a world of highly-scalable compute. This makes it ideal not only for PCs, but for a wide range of applications, such as 5G modems, storage, ADAS/autonomous vehicles, data centre and high-performance computing. In addition, our Ethernet Packet Processor (EPP) IP cores enable dataflow between processors and the network to avoid CPU bottlenecks.
Winning in the new PC processing era
The biggest tech companies are increasingly moving away from using off-the-shelf silicon to developing bespoke solutions to gain greater control of their designs, and thus gain an edge in the market. OEMs that aren't designing their own silicon face a growing challenge with few highly competitive off-the-shelf silicon solutions available.
This presents an opportunity for SoC manufacturers to develop highly-optimised solutions that can give OEMs an edge. It's all about providing more performance and features within the same power envelope – delivering a better user experience, whilst extending battery life.
At Imagination, we've designed all our IP cores on mobile design principles because we know that is what our customers need to create forward-looking designs. With efficient and scalable processors designed for heterogeneous architectures, SoC manufacturers can choose the right combination of compute elements for their applications and provide winning solutions for their customers.