The mobile GPU space has been heating up for a while now. Apple held the crown of the fastest GPU for a long time but was bested by Qualcomm this generation, and Qualcomm's dominance relative to the rest of the Android world was also incredible. With Arm's Immortalis GPUs, though, things are picking up pace and making the space ever the more competitive.
We recently sat down with Steve Raphael from Arm at this year's MWC and followed it up with another conversation with Andy Craigen to talk us through Arm, its current trajectory, and the work the company has completed around its new line of Immortalis GPUs, starting with the G715. As Raphael puts it, “we are so proud of what the team has done. If you look at some of the comparisons, it’s holding its own out there, as a flagship GPU."
Our testing definitely confirms that, and it can go reasonably toe-to-toe against the Adreno 740 of the Snapdragon 8 Gen 2. However, what actually goes into making the Immortalis G715 into a formidable competitor?
Dissecting the Immortalis name
First and foremost, "Immortalis" is a new name for Arm's highest-tier GPUs, and the reason for its existence is out of necessity and to not confuse consumers. “In order to not create confusion in the market, we need to create some separate categories” Craigen tells me. Immortalis has features that Mali doesn't, though it encompasses all of the features that Mali has.
As a result, you can actually get a Mali G715 GPU too, and it's not just the Immortalis G715. Both GPUs are more or less the same, except the Mali G715 doesn't have a Ray-Tracing Unit, or RTU, and also has fewer core options; even if those cores themselves are all the same on a technical level. There's no telling whether there will be further deviations in the future between the two, but it's setting the stage for there to be significantly more differences in the future if Arm deems it relevant.
As for what the name "Immortalis" means, it's Latin for "Immortal," meaning that something is immune from death. It suggests power and strength, which is something Arm appears to be leaning into a lot more as it produces flagship-class GPUs.
Deep-diving into the Immortalis G715
Craigen spent time with me walking through what changes were in the G715 that led to it being so performant, and some of those changes are incredibly interesting. The biggest change was the overall improvement of performance efficiency of the entire chipset, with the shrinkage of the Fused-Multiply Add module. This module is used for floating-point calculations, which are incredibly important in GPUs.
As a result of this, Arm was able to double the number of FMAs per execution engine, one in each shader core, while also adding an MMUL, which meant significant gains in the form of a two-times improvement in performance at only 1.3 times the silicon area. However, as Craigen notes, "if we double the amount of FMA in the execution unit, you have to make sure there are no bottlenecks elsewhere in the machine. Otherwise, it’s pointless."
Because of that, Arm improved multiple other modules across the GPU, too. These are dubbed "PPA improvements," or power, performance, and area improvements. The Command Stream Frontend (CSF) is made faster, there's a three-times improvement in triangle throughput via the Tiler, two-times FP16 blender throughput, and a few other improvements too. All of these improvements are what result in a much faster GPU.
If there's a 100% improvement in compute, then why is there not a 100% improvement in GPU, as per Arm's own figures that they shared? The company boasts a modest 15% improvement, but even in our own testing, a lot of the results we collected are a lot more than just a 15% improvement in FPS. Craigen tells me the reason is that the company tests across a lot of different use cases, and compute is merely one of them. Other improvements aren't as extreme, and it's a totality of the GPU that they test for when they share their improvement metrics. It makes sense and would be misleading to advertise a 100% compute improvement without clarifying what that actually means for the GPU.
The limitation of being in a smartphone is primarily the size. “Mobile devices are very different to consoles or PCs where you have, certainly on PCs, an almost limitless thermal budget depending on your fans [and other criteria]," Craigen tells me. It's a perfect balancing act between the thermal output of a chip and the power that it's capable of, and a lot of OEMs put their own spin on it, too.
The Immortalis G715 is the most flagship-level GPU from Arm yet
For years, Mali GPUs have faltered behind the rest of the competition in the mobile gaming space. They're adequate for basic tasks and for hobbyist mobile gaming, but anyone who is serious about getting the best performance would typically like to veer away from a Mali GPU. With the Immortalis G715, Arm went for a new branding that Craigen says is "where we’re saying these are the best performance features and functions for the premium smartphone market."
Where things get a little bit confusing is that there is still is also a Mali G715 GPU, but its only difference is that it doesn't have ray-tracing support. Everything else remains the same between the Immortalis and Mali G715 GPUs, though that doesn't mean that will be the only difference in future models. As more flagship features come to Arm's GPUs, the need to discern between Immortalis and Mali further may become apparent.
As for how these technologies are implemented, Arm pushes to use Vulkan wherever possible for complete interoperability. Craigen says that “proprietary APIs cause fragmentation, which nobody particularly likes," whereas using the Vulkan API to implement ray-tracing, for example, means that apps that make use of it can just call the Vulkan API, and it doesn't matter what hardware is in place so long as it has ray-tracing support. He told me that “we are very much focused, and have been for a long time, on Vulkan."
What comes next from Arm?
In speaking to both Raphael and Craigen, it's clear that a lot goes into GPU designing. As Craigen says, “there’s a fairly long period between how we develop an IP and then getting it into an SoC to then getting it out into a phone. That’s quite a significant period of time, roughly in the order of five years." I asked Raphael what to expect from the future, and he took a coy approach. He told me to expect great things and that from this generation, there will be a "wide range of implementations, you’ll see all sorts of stuff.”
With such massive jumps being made from one generation to the next, I'll be curious to see if the next generation of Arm GPU will have similar improvements. Arm nearly went toe-to-toe against Qualcomm in this generation, and for typical day-to-day usage, both GPUs are essentially the same in power. Testing even reveals that Arm's ray tracing seems to perform better than Qualcomm's.
The future appears to be bright for Immortalis GPUs, and we're excited to see another player truly enter the space and take on Qualcomm.