Clouding the matter is the fact that the lines between the CPU and GPU have blurred significantly in recent years. On the desktop, it is becoming increasingly more common for applications to be accelerated on the GPU. And GPUs are revolutionizing the way many different types of data are processed on the enterprise level. These ongoing changes in the industry have spurred much discussion about how PCs will be designed in the future and have sparked heated debates between users. Even the companies involved have traded shots. Nvidia CEO Jen-Sun Huang, when responding to questions regarding Intel’s future discrete graphics product Larrabee, went so far as to say that his company would be opening a can of “whoop ass” on Intel.

To gain some inside perspective and a better understanding of where the industry is headed, we decided to speak with representatives from a handful of the most influential companies in PC technology. We spoke with Intel spokesperson Nick Knupffer; AMD’s Vice President of Advanced Marketing Pat Moorhead; the general manager of platform business at Nvidia, Drew Henry; and VIA’s Centaur design center president, Glenn Henry. In addition, we got an outsider’s perspective by talking with Jon Peddie, president of Jon Peddie Research. Not surprisingly, we received a number of opinions on the future of CPUs and GPUs.

Past, Present & Future

When asked about the current relationship between CPUs and GPUs, Peddie explains things succinctly: “[It’s] totally symbiotic. One cannot live without the other.

“The GPU fulfills two opportunities for computer users: It’s a graphics accelerator (which is an ASAP [application-specific application processor]) and also a high-performance coprocessor for complex SIMD [single instruction, multiple data] or vector-based datasets used in scientific, financial, and multimedia applications in a heterogeneous architecture. The CPU cannot do the parallel operations that a GPU does, and a GPU cannot do the scalar operations that a CPU does. They need each other, and neither can replace the other.”

In other words, there’s no single processor in a PC that is good at both highly parallel and scalar operations. As software evolves and computing applications change over time, finding the right balance of CPU and GPU performance becomes increasingly more important for a good user experience.

In response to a similar question about the current and future relationship between CPUs and GPUs, AMD’s Moorhead states, “Before talking about what’s ahead, let’s first take a look at where we are now. Today’s model focuses almost exclusively on adding more of the same CPU cores to chip designs. For certain market segments, such as commercial servers, this model will continue to be the right choice for many years. This is because commercial segments already have extensive access to applications capable of benefiting from greater numbers of homogenous cores.”

Moorhead continues, “For the consumer segments, however, the number of applications that benefit from homogenous multicore CPUs drops considerably, especially as the number of CPU cores increases above four.” According to Moorhead, an increasing number of software vendors and users are becoming more interested in how they can benefit from accelerating applications beyond CPU optimization. For example, GPUs’ highly parallel processing architectures are best suited for 3D applications, working with HD media, and certain other nonvisual applications.

Moorhead’s point plays out in modern benchmarks. If an application can’t take advantage of multiple threads, adding more cores has little to no impact on a system’s performance. Moorhead also hints at the increasing importance of using the GPU for complex parallel computations.

The Impending CPU/GPU Fusion

One of the hottest topics right now is CPU/GPU fusion. Both AMD and Intel have announced plans to incorporate graphics processors on the CPU die. However, what this means to companies such as Nvidia and AMD (and their partners), which sell many motherboard IGPs, or to the discrete graphics card business, remains to be seen. And although a combined CPU/GPU is likely to offer better performance than current IGP technology because of the closer coupling of the CPU, GPU, and system memory, we won’t know just how good performance will be for some time.

When asked about how CPU/GPU fusion will impact the industry, however, Jon Peddie confidently states that, “Embedded GPUs will not be able to deliver the performance needed for interesting multimedia and professional applications like video and picture editing, HD movie viewing, CAD, DCC, HPC applications, and games.” Jon’s position is one shared by many in the industry, but you wouldn’t think that was the case with all of the marketing material surrounding CPU/GPU fusion.

Although he doesn’t make any outward performance claims, Moorhead revealed some interesting information regarding AMD’s future plans in this area.

“AMD views the combining of CPU cores and accelerator cores on-chip as representing a new category of mainstream microprocessor, which it has titled the Accelerated Processing Unit [APU],” he says. “While the integration of a GPU core within APU designs is a one of our near-term priorities, the APU category represents much more than merely ‘CPU+GPU’ [in the long term].

“In the future, AMD envisions APUs that include a varying mix of CPU cores (scalar processing cores), GPU cores (parallel processing cores), and fixed-function accelerator cores—all on-chip.” Moorhead adds that AMD won’t leave out off-chip accelerators, preserving the right hardware combinations for every level of the market. To take advantage of these new architectures, AMD is working toward simplifying the programming environment.

This is an interesting development, even if it is one we have speculated on in the past. Down the road, GPU cores won’t be the only specialized cores integrated onto the CPU die. “With the ‘Swift’ APU, our goal isn’t to make the integrated graphics chipsets of today obsolete,” Moorhead says, “rather, it’s to improve the overall computing experience while improving the cost and energy consumption associated with that type of experience.” With the Swift APU, AMD should have a platform that is both cost- and energy-efficient.

A War Is Looming

Aside from CPU/GPU fusion, other recent developments in the CPU and GPU industries have left some companies, predominantly Nvidia, in precarious positions. We don’t have to rehash AMD’s ATI acquisition here: The results of the merger are well-known. But Intel plans to re-enter the discrete GPU space with the Larrabee product, as well as bolster performance on its integrated GPU. This leaves many to wonder what Nvidia will be able to do without a total platform package. After all, AMD and Intel will be able to win over consumers and OEMs with top-to-bottom offerings consisting of CPUs, chipsets, and GPUs. Lacking its own CPU, Nvidia has gone on the offensive with a campaign designed to educate consumers on the importance of the GPU and a balanced, optimized PC.

“For mainstream PCs, processing pictures, video, and 3D graphics have become the top reasons for purchasing a PC,” Nvidia’s Henry says. “GPUs are perfectly designed for these applications. Great new applications are starting to come out, such as RapiHD from Elemental Technologies. This application speeds up the transcoding of SD and HD videos to post on YouTube or watch on your iPod by 10 times or more. GPUs are inherently parallel, so any application that benefits from this parallelism would run great on a GPU.”

With all of the fuss regarding GPUs encroaching on territory historically reserved for CPUs, we tend to forget that graphics aren’t important to all users. Glenn Henry makes an interesting point: “Compute and data servers typically have little need for graphics. For example, we have over a thousand high-end PCs in our compute center (and we’re a very small company), none of which have any display output.”

He adds, “The current ‘mainstream’ PC definitely has need for lots of graphics and image processing power—not just for gaming, but also for video, image processing, fancy 3D interfaces, etc. But these types of systems also have requirements for lots of CPU performance. I’m not sure which need is greater, but a balance, as Nvidia suggests, is certainly required.”

What’s On The Horizon?

Opinions on where the industry is headed were decidedly mixed, but some common themes did arise in our conversations, namely platformization, parallelization, and integration. “In most cases, mainstream software applications today are already heavily challenged to utilize the full compute capabilities of four CPU cores, let alone 16 or more,” says Moorhead. “By combining the optimum mix of CPU cores with differing types of on-chip and off-chip accelerators, platforms can be better tailored to accelerate the types of software a particular end-user segment cares about most. The concept of focus on the platform is such that the right hardware can be used to achieve the hardware balance that’s desired, while providing software abstraction layers that make it possible for programmers not to know about these hardware changes.”

“Intel drives and defines our products on the platform level, including CPU, chipset, and graphics. It is difficult to predict beyond the announcements Intel has already made on plans to integrate CPU and graphics in a single device,” says Intel’s Knupffer. Knupffer also says that Intel is currently researching the following areas: integrated 2D mesh networks to directly connect cores and other resources (80-core Terascale chip); speculative multithreading to extract thread-level parallelism at the hardware level; various circuits to accelerate specific tasks such as HD encode and network I/O; new kinds of denser on-chip memory circuits, as well as 3D stacking; resilient circuits and architectures to push performance higher and reduce voltage thresholds; and fundamental changes to the cache hierarchy to improve memory sharing.

Although not quite as specific, VIA’s Glenn Henry makes similar statements. “I see two major thrusts, more powerful, specialized instructions and integration of more function onto [the] same chip as [the] processor: memory controller, graphics controller, most southbridge functions,” he says.

Drew Henry, on the other hand, offers more pointed comments. “GPU performance is critical for most users,” he says. “We expect to be well ahead in performance and features than the products which integrate a GPU onto a CPU core. Visual computing applications are advancing fast and demanding more and more GPU processing power. The GPU is the processing architecture for the future of computing.”

Of course, what AMD, Intel, and Nvidia have to say doesn’t always jibe completely with outside observers. Peddie says, “I’ve said for several years now that the GPU is the processor and CPU is the co-processor. The user looks at and responds to the screen the entire time he or she is in front of a computer.”

The End Game

Exactly how all of the impending changes to CPU and GPU architectures will affect the industry remains to be seen. One thing is for certain, though: There are going to be some fireworks.

by Marco Chiappetta