Your browser is not supported. Please update it.

14 November 2018

Imagination launches new auto GPU stack, market far from shaken out

Imagination Technologies has launched its latest automotive GPU design, the PowerVR Series8XT-A, a new IP design that it is pitching at next-gen automotive designs. A key part of the offering is the supporting software, tools, and support, which Imagination is saying will provide a ten-year lifespan.

The GPU design, which automotive OEMs can then license from Imagination and commission someone to build for them, makes the expected performance claims, but a big part of the launch marketing concerns its safety features. Automotive buyers are extremely sensitive to such issues, and the ISO 26262 and IEC 61508 compliance is apparently laid out in the extensive documentation.

Of course, the company is one of the automotive old-guard. Imagination claims that it has over a 50% market share by volume, thanks to customers shipping tens of millions of devices – with the most extensive use being in in-vehicle infotainment (IVI), digital cockpit dashboard instrument clusters, and ADAS (advanced driver assistance systems).

The new designs are being targeted at increasingly complex ADAS systems, which are evolving into semi-autonomous and then fully-autonomous features. Due to its conservatism, the automotive industry is looking to established vendors, and while newer entrants like Nvidia have carved out a slice, this is definitely not a market open to startups. Tesla is trying to go its own way, with regard to silicon, but the rest are sticking with what they know – largely.

Imagination Technologies’ CEO Leo Li said “the automotive industry is rapidly transitioning to assisted driving and fully autonomous vehicles. PowerVR has extensive experience in the automotive semiconductor market and as the requirements for power and performance continue to evolve, PowerVR is committed to remaining ahead of the curve. Imagination is excited to deliver this package that allows the growing group of automotive players to address the complex requirements of this market.”

The company cites Denso, Renesas, Socionext, and Texas Instruments, as flagship customers, but the company claims that its PowerVR package does lower the barrier to entry for new players – something that we don’t think is really the case, as automakers don’t want to buy from companies they don’t know and don’t yet trust.

As for technical details, the new Series8XT-A is based on Imagination’s Furian architecture, and is the automotive variant of the Series8XT family. New hardware and software fault detection features have been introduced, which Imagination says provides better recovery for the GPU should a problem arise. Error-Correcting Code (ECC) and Logic Built-in Self-Test (LBIST) are prominent too, as well as hardware virtualization to facilitate multiple separate concurrent applications, running in their own containers.

The more interesting addition is the Neural Network Accelerators (NNAs) that Imagination is trotting out to compliment the GPUs. The PowerVR Series2NX NNA is intended for use in SAE Level 3 to Level 5 functions, ranging from driver monitoring and gaze tracking, through to drivable path analysis and road user detection – apparently.

These AI-based functions are integral to the next-gen autonomous features, and Imagination is hoping to sell both the GPU and machine-learning silicon that will power the necessary chips. While the Series8XT-A is available to license now, the NNAs are going to be launching in 2019.  Given the long lead times of the automotive industry, ranging between 3-5 years, that provides a likely time-frame for when these designs begin appearing in vehicles.

Of course, Imagination has some serious competition in this market, and one of the largest automotive silicon providers, NXP, has got some reshuffling to do. After Qualcomm’s attempted takeover was abandoned, NXP now has to refocus on expanding up the stack, moving from basic microcontrollers and application processors, to rival the likes of Nvidia’s Drive platform – the extremely power central computer systems that Nvidia envisions powering the bulk of self-driving applications within a vehicle.

EETimes has a good interview piece with CTO Lars Reger, recounting an 11-hour flight on which Reger thought that the deal had finally achieved Chinese regulatory approval when he took off, and then found that the opposite was true when he landed. You can dive into it fully here.

Qualcomm’s $44bn purchase could have heavily shaken up the market. Qualcomm had essentially realized that the smartphone market was limiting, and wanted to expand its smaller forays into the automotive market. What better way than buying a leader in the sector, it thought, before geopolitical and antitrust tensions nixed the deal.

When that deal was announced, back in November 2016, Qualcomm was expecting to pay $47bn for the fifth-largest non-memory chip maker in the world. The price fluctuated somewhat, but Qualcomm was definitely keen to diversify its mobile-centric portfolio. It did have Snapdragon Automotive designs, but NXP brought a whole lot more to the table than just Snapdragon-powered IVI systems.

Shortly after the deal was announced, Qualcomm launched the Drive Data Platform, a sensor-fusion technology that would let Qualcomm’s Snapdragon 820Am SoC process all the sensor-inputs in a vehicle, and use its Neural Processing Engine (NPE) to make decisions.

So while Qualcomm will likely keep plugging away at the automotive sector, it is not going to be able to use NXP as a launching point. Similarly, NXP loses out on using Qualcomm’s SoCs as starting points for those sensor-fusion applications, which Reger said would have been combined with NXP’s Layerscape processors.

Currently, NXP’s BlueBox platform is the core of the advanced automotive portfolio. BlueBox is a modular architecture that lets customers build their own stacks according to their needs, rather than having to pick a chip that might not be quite right. Unveiled back in May 2016, before the Qualcomm announcement, NXP was keen to stress that a major advantage of modularity was that OEMs could combine chips from other vendors into their designs.

The BlueBox front has been very quiet in the passing years, likely because of the impact that Qualcomm’s acquisition would have had. It would have given NXP the large central processor needed for these autonomous functions, in Snapdragon. Now, it needs to either create its own, or partner and license a design. Its main options are Qualcomm, Nvidia, and also Intel’s Mobileye EyeQ range.

But Reger says that advanced autonomous vehicles are still years away, and that the evidence of this are the earnings warnings trickling out of the automakers due to the billions they have spent already in EV and autonomous research, as well as the partnerships that both they and their T1 suppliers are striking. Reger believes that no automaker or T1 is going to make money on SAE Level 5 vehicles in the next five to six years.

He added that in that time frame, any chip design is going to age poorly. “Six or eight years from now, the advanced process technology will move from 10nm to 7nm, or even down to 3nm. That simply means one thing – by the time your big SoC hits the volume market, you’re using a two-generation old technology.”

So while Imagination is trying to target the upper layers, those big SoCs, NXP is looking more at vision and radar applications, as well as a machine-vision chip – likely building on the Cognivue IP that it acquired when NXP bought Freescale for $11.8bn. The product names to keep an eye out for are the S32 family, with variants for radar (S32R), vision (S32V), and gateways (S32G). These have been years in development, and NXP is getting closer to finalizing them.