Not everyone is aware of the complex technical lineage involved in the parts nestled inside their smartphone, even if they know of names like Qualcomm or MediaTek and the chips these companies make. Basically all consumer Android devices use a CPU design and architecture created by a company called ARM, tweaking and customizing bits as they see fit. Today ARM announced its new CPUs and GPUs, which, over the next year or two, will trickle down into actual hardware released by chipset manufacturers and eventually end up in a phone you see online, on a store shelf, or maybe in your pocket. And by far the biggest announcement is ARM's new GPU, which supports hardware ray tracing — a feature mostly limited to recent game consoles and high-end desktop GPUs, though Samsung beat ARM to it.

Immortalis-G715 and the new GPUs

The new GPU, Immortalis-G715, is based on the same Mali architecture many ARM GPUs use. This is the part where some folks groan because Mali designs are sometimes seen as a few steps behind in performance compared to some other companies in the space, but ARM maintains that there are big performance gains to be had here.

chrome_058lRjqRTx

In a pure feature sense, the Immortalis picks up hardware-based ray tracing support (compatible with the Vulkan API), variable-rate shading, matrix multiply instruction support, and increased developer support and tools. These numbers aren't particularly useful, but ARM claims a 15% performance improvement over prior solutions and a 2x increases in machine learning performance. The company claims it's also 15% more energy-efficient, which could equate to reduced power consumption or operating temperatures.

Ray tracing means fancier and better reflections and lighting in rendered scenes, and it's worth stressing here that ARM isn't actually first to announce a mobile chipset with the feature. The Samsung Exynos 2200 revealed in January has an AMD-based GPU that also claims to offer hardware-accelerated ray tracing.

Many of the details presented by ARM were later caveated by explaining that particulars and benefits will vary based on the implementation chosen by chipset partners — i.e., ARM has these designs, but they can scale up or down accoridng to manufacturer requirements. Ultimately, they can use and tweak it within those limits as they see fit, adjusting things like the number of cores or other metrics to fit their design. The G715 can be configured with 10 cores or more (we weren't provided with an upper limit).

A raytracing demo provided by ARM

ARM hasn't directly mentioned any partners that will use the tech, but MediaTek did congratulate the company for its announcement, which could be leading. MediaTek has also used Mali GPUs in its products before.

ARM had fewer details to share regarding the smaller GPUs, but the Immortalis has a pair of younger brothers that it shares its architecture with. The G715 replaces last year's G710 and inherits some of the features that the G715-Immortalis has, like variable-rate shading. Based on the shared nomenclature, we have to assume the two designs share a lot of guts, but the G715 won't have hardware-based ray tracing. It will be available in 7-9 core designs.

GPUs blog image1 (NEW).jpg.png-1039x498
Source: ARM

The slightly lower-end G615 replaces last year's G610 and offers the same features but comes in designs of 6 cores or less.

Cortex-X3 and the new (and old) CPUs

The new Cortex-X3 is the easiest core to get excited about as it's the biggest, newest high-performance ARMv9 CPU design. ARM again claims double-digit IPC improvements. The company doesn't even mince words, outright stating that it's targeting things like benchmarks, but it can also deliver a high level of performance for other workloads. Cumulatively, this means we have a big and fast new core to look forward to in the highest-end flagship chipsets.

Technical improvements include a 50% larger L1/L2 branch target buffer capacity and 10x larger L0 capacity.

CPUs blog image 8
Image source: ARM.

Even outside smartphones, the X3 can be used in up to 12-core configurations for things like laptops or desktops, and ARM says that the new design delivers 34% improved single-threaded performance compared to the latest mainstream laptops.

While many of us were waiting for Qualcomm's Nuvia plans for ARM desktop/laptop computing outside Apple to finally take off, it sounds like ARM's generic designs are getting to a similar point. We'll have to wait and see how they end up being used.

The Cortex-A715 offers improvements over the A710 when it comes to things like branch prediction accuracy and data prefetching, while also offering "consistent IPC gains." In fact, ARM says that the A715 matches the performance of the Cortex-X1, the first of ARM's still relatively new "super-big" cores that landed in 2020 as part of its custom CPU program. Compared to the older A710, it's 20% more power-efficient when operating at the same performance level and 5% faster at peak.

There's not actually a new ARMv9 little core this year. Instead, ARM has increased the efficiency of last year's Cortex-A510 by marginally improving its efficiency at the same performance level, saving around 5% in power, according to ARM. The entire second-gen ARMv9 lineup also includes a handful of security architectures benefits — acronyms for specific Linux kernel features and other tweaks that you won't understand, but which mean the cores have slightly more hardened security.

TC22 blog image1.png-1040x0
Image source: ARM

This is all a part of ARM's "Total Compute Solution 2022," which is what it calls its relatively recent end-to-end designs for chipsets — think of it like something closer to a drop-in solution a chipset manufacturer can build upon for GPU, CPU, machine learning, and basically everything. It has designs ranging from premium phones and laptops, mid-range phones, Chromebooks, and XR, down to budget devices, AR, and even set-top boxes. Here ARM is delivering other improvements to its machine learning solutions (double-digit uplift claimed in certain tasks like object detection and classification) and further hardening security.

ARM also stressed in the Q&A portion of a recent briefing that its second-gen ARMv9 solutions are really meant to drive home the push to 64-bit and encourage an end to 32-bit support, poking fun at the Chinese market's apparent insistence that 32-bit legacy support be maintained, which has held back hardware advancement to a certain degree. We should also point out that A510 still has some optional 32-bit support. If ARM puts its foot down, we might see 32-bit support really and truly end soon.

Zoom_o9CiU5Sn0d

The company's future hardware roadmap was also shown off, with names like Titan and Krake corresponding to future GPUs, "CXC23" and "CXC24" designs planned as Cortex X-series successors (associated with a new Hayden DSU cluster), and a new Hayes small/efficiency core design planned for 2023 and 2024. Other seemingly more fundamental architectural changes also look to be planned for 2024, with what look to be interconnect changes coming.

Ultimately, all this means that ARM has some new designs — for CPU cores, GPUs, and even full chipsets — that you might start to see landing in devices soon. ARM hinted that at least one vendor using the Cortex X3 would be announced later this year, with real-world hardware expected to land early next year.