Technology company media events can be broadly divided into two categories: Apple-style events, which showcase completed new products just before they hit stores, and Qualcomm-style events, which showcase new technology well before it is actually available The innovations will be included, sometimes without naming the first products. Instead of calling one type of event “better” than the other, I consider both to be valuable. Apple reveals “what is now” while Qualcomm reveals “what will be now”.
Last December, Qualcomm announced a breakthrough chipset for mixed reality devices called the Snapdragon XR2, which follows the “see what’s coming” strategy mentioned above. Designed specifically for standalone VR and AR headsets, the XR2 seemed ready to smoke the smartphone-grade Snapdragon 835 chips found in devices like the consumer Oculus Quest and the enterprise HTC Vive Focus, thanks to the performance improvements of the generation caliber Plus to be used. One diagram promised twice the CPU and GPU performance, four times the pixel throughput, six times the screen resolution and eleven times the dedicated AI TOPS. In short, XR2-based all-in-one VR headsets work far more like PCs than cell phones.
While Qualcomm hasn’t responded to a query on this point, there are indications that the R-1 won’t be the first Snapdragon XR2 headset to hit the market. This week, supposed Geekbench 5 results surfaced for an unannounced HTC Vive Focus model with the XR2, and the specs matched the familiar configuration of the new chipset: an 8-core 1.8 GHz CPU and Android 10. The results include single and multi-core scores of 924 and 3415, which are many times faster than Snapdragon 835 headsets like the Oculus Quest (267/746). Memory details added credibility to the results and was in line with expectations that the XR2 would generally be paired with 6GB of RAM, versus 4GB in previous generation solutions.
Business users may be interested in the new Focus – the Focus Plus is roughly half the cost that Lynx charges for the R-1 – but I don’t think these will be the only Snapdragon XR2 headsets available anytime soon. On the contrary, I believe the XR2 is on the verge of being adopted by consumer VR and AR headset makers, just as Qualcomm anticipated VentureBeat last year which means we are finally ready for the discrete mixed reality to take its long-awaited big leap forward.
What do XR2-based graphics really look like?
Note that while Qualcomm provides developers with both the chipset and reference platforms, finished XR2 headsets vary from company to company. For example, the fact that the chipset can theoretically power displays six times the resolution of previous VR screens, does not mean that a Focus sequel will actually deliver six times the previous resolution of a given headset. In the past, Qualcomm offered the chip-level capability. Each OEM then needs to select the parts that make sense for a particular audience, range of applications, and price.
I firmly believe that the general trend will be towards higher resolution VR displays, but with multiples in the 1.5-2X range instead of 6X. A few months ago I discussed the noticeable difference in visual clarity that Picos Neo 2 offers, which has 1,920 x 2,160 resolution per eye, almost twice as much detail as the dual 1,440 x 1,600 screens of Oculus Quest. Under the right circumstances, screens at this resolution can be largely devoid of the “screen door effect” that allows users to see individual pixels in black grids, requiring GPUs to provide enough polygon and texture detail to delight users.
Since the Neo 2 is powered by a Snapdragon 845 – a chip whose performance is roughly halfway between the 835 and the XR2 – I would expect the XR2 headsets to outperform Pico’s offerings. As I discovered in June, the Neo 2’s graphics look like higher-resolution versions of Quest graphics, with an imperceptibly faster refresh rate (75 Hz versus Quest’s 72 Hz). Based on Qualcomm’s statements about XR2 and Lynx’s statements for the Lynx-R1, it can be assumed that the new headsets will typically, if not generally, use a refresh rate of 90 Hz. This means PC-class display speeds with a lower chance of nausea.
Facebook surprised everyone with the Quest’s performance this generation, wrestling the Snapdragon 835 with more complex graphics than anyone could have imagined (except perhaps a John Carmack-caliber 3D graphics assistant). Qualcomm suggested that the XR2 has twice as much CPU and GPU power as the 835, but if the early benchmarks are to be believed, that’s a detriment to the capabilities of the new chipset. Since the Quest was able to approximate – not compare – the graphics in PlayStation VR titles, it is expected that XR2-based titles will compete with current generation console VR, if not entry-level PC VR . Beyond gaming, this means that social apps like Facebook Horizon and productivity / streaming apps like Virtual Desktop can look about as detailed, complex, and fluid on an XR2 headset as they do on the Oculus Rift.
One caveat: “Current Generation” and “Entry” are moving targets. Today we’re only two months away from a console generation transition, and even basic PCs keep improving their graphics performance every day. There is no way in the world that XR2 mobile-grade headsets completely eliminate the need for highly-specified dedicated machines. If the trend of the quest continues, the visual delta between tethered and tethered headsets will matter to fewer people.
What about XR2 AI?
While the raw data was announced last December, the AI processing capabilities of the Snapdragon XR2 are underestimated and could be a major factor in improving the performance of mixed reality headsets. Here, the gap between the Snapdragon 835 and XR2 is an 11-fold improvement – 15 trillion operations per second (TOPS) versus about 1.3 TOPS in the 835. That’s two-thirds more than Qualcomm’s Snapdragon 8cx Gen 2 and above laptop-class contains paper that matches the Snapdragon 865 flagship smartphone chipset.
I say “on paper” because the raw number of instructions doesn’t necessarily tell the full story of AI performance. Quantity is important, but so are quality and engineering and software considerations at the system level in order to be able to use the functions. However, if you think back to the early days of the Oculus Quest, the limited computer vision system magically delivered room-scale SLAM scans and 6DoF controller tracking, which Facebook then software updated to add impressively smooth hand tracking. Imagine what a headset with 10x more AI performance could do with tracking and cameras. Then assume that offloading some AI-related tasks from the CPU will also get better general-purpose performance.
There are other ways AI capabilities can affect the performance of mixed reality headsets, including empowering more intelligent computer opponents and generating solutions to semi-novel problems. They also enable more extensive voice control and help segment live graphics in augmented reality headsets so that they can be seamlessly combined with digital content.
Thanks to the COVID-19 pandemic and general economic uncertainty, the plans of many XR developers changed in 2020: Magic Leap imploded, the CEO of HTC resigned, Facebook renamed Oculus Connect and moved it completely online. Several VR developers turned around, sold themselves and / or shut down. To the extent that Lynx plausibly expected February to hit its promised summer date for R-1, it’s not exactly surprising that it didn’t.
Nothing is certain at the moment, but I’m pretty sure we’ll see Snapdragon XR2 headsets in stores – especially early ones with no 5G connectivity, rather than more deluxe models that have robust CPU / GPU / AI performance the freedom of a cellular connection. I wouldn’t expect this to happen until next year at the earliest, and even then only in areas with more robust 5G infrastructures than currently in the US. When that happens – and there is a when, not an if – we will see VR and AR everywhere, and that will be another huge step forward for this already exciting and rapidly evolving collection of technologies.