Architectures for the Future: PowerVR, Furian, ARM, DynamIQ

Imagination Technologies recently unveiled its next-generation GPU architecture, Furian, suited for 7 nm silicon architectures and targeted at high performance computing in a wide variety of new applications, and for higher density screen resolutions. Compared to the previous-generation Imagination Technologies GPU architecture, Rogue, Furian is claimed to have higher density in the same process technology and at the same clock rate:

• 35% GFLOPS more per silicon area in the same process rate
• 80% fill rate improvement, driven largely from architectural and routing changes
• 70% to 90% density improvement for gaming applications

Goals stated by Furian include:

• Moving from OpenGL embedded systems to Vulkan programming.
• Shift from rendering scenes for 1080p @ 60 fps to 4K @ 120 fps, with 10-bit high dynamic range (HDR)
• Support for artificial intelligence (AI) computation in the form of neural nets/convolutional neural networking (CNN)
• Automotive advanced driver assistance (ADAS)
• New form factor devices, such as convertible tablets
• Improved GPU pipeline with ability to handle deferred rendering
• ARM unveiled its DynamIQ architecture. This is a more CPU-centric approach, and to a lesser extent, is a system-centric approach compared to Imagination Technologies’ GPU-centric approach. Key features include:
• Up to 50X AI performance boost in 3-5 years.
• Up to 10X faster response to accelerators.
• Specific conformance targeting for automotive sector (ASIL D) around resilient systems, adaptive computing (ability to use smaller or larger cores based on workload), and faster responsiveness (lower latency on interrupt response).
• Improved heterogenous processing, extending the big.LITTLE approach to cover any types of general-purpose computing cores, special-purpose computing cores, or fixed function blocks.
• Additional speed control and sleep modes.

Both companies have made these architectural announcements in line with the early architectural design cycles of leading silicon manufacturers. In other words, the ink is not dry yet. These architectures have yet to be realized in specifically disclosed cores, where real power-area-performance tradeoffs are met and can be compared against real workloads. Overall, Imagination Technologies’ announcement was much more specific in terms of specific instruction set and architectural changes helping to achieve their benefits. ARM’s announcement was as much about reinforcing the role of ARM CPU’s in the mobile ecosystem—and an expanding portion of the mobile ecosystem—as it was about specific architectural directions and capabilities, not features and functions.  

Interestingly, however, we start to see that the vision of computing is very common across both companies—the battle between these IP providers (those that enable core elements of the designs behind little known names like Apple’s A7 chip, Samsung’s Exynos, Qualcomm’s Snapdragon, Broadcom, MediaTek, and other silicon providers) and other silicon providers is also very present:

• Intel has both the Intel Architecture and Altera FPGA platform under the hood. While ARM again can claims some traction in the supercomputing domain, Intel owns the data center and has a lead on 5G virtualized network architectures. Intel’s significant investment in the automotive space (Mobileye, Movidius) shows it is taking a clear look at visual computing, even as separate investments in virtual reality (VR) are being fed into its sports directions (Voke, Replay Technology).
• nVidia switched gears from the PC- and gaming-oriented GPU markets (which it still dominates) to the automotive sector.  nVidia’s strongest barrier is that it still consumes significant power per operation compared to mobile architectures, such as those by Qualcomm.

Common threads in the future computing market include:

• Increased role for computer vision (see, for example, ABI Research’s report entitled, Machine Vision in Augmented and Virtual Reality Markets: [AN-2406])
• More heterogeneous processing elements, with specialized CPUs, GPUs, vision engines, engines for convolutional neural networks, etc.
• Continuing providing increased performance driving next-generation interfaces forward and making our machines anticipate our needs more and more of the time.

All in all, IP suppliers are chasing after more vertically-specific market sectors, with critical areas including:

• Embedded AI and Machine Vision
• Cloud-based Big Data and AI—sifting through large amounts of data with the ability to find out what matters, and predict the future outcomes.
• Vision within the automotive context
• 3D workloads for VR and AR
• IoT—small, power-efficient nodes providing data to larger systems.

From an AR and VR perspective, specifically:

• A number of core functions, such as simultaneous location and mapping (SLAM) will occur between the camera and display processors. Moving functions such as location, registration, mapping, rotation, etc., to the display processor can improve overall efficiency and reduce latency.
• Higher levels of rendering complexity, and more efficient rendering, to achieve 4K display resolution and higher frame rates.
• Solutions, which decrease the latency between generation and display, can reduce the overall processing performance requirements significantly.