This week at Computex 2017 in Taipei, Taiwan, ARM Holdings announced their latest series of processor designs spanning mobility, datacenter and AI. While ARM Holdings doesn’t make chips themselves, they do design processor cores and other IP for most chip companies to utilize with their own designs. Just last month, ARM introduces their first ISP (image signal processor) for self-driving cars, which I wrote about here. ARM Holdings has been busy as ever with their latest processor and infrastructure designs, enabling them to be deployed from the datacenter to the smartwatch. Because ARM Holdings technology has shipped in over 100 billion chips in virtually all markets, the company is uniquely positioned to be at the edge of nearly everything that will happen in IoT and by extension, AI. Don’t be confused on the preferred Ai location, as on-device AI is as important as that in the cloud. By doing more AI and ML tasks on-device, ARM Holdings and their chip partners can decrease the privacy and security risks posed by moving such workloads rapidly and in huge volumes over to the cloud. On-device AI also helps with latency. So let me jump into what the company announced.
New Cortex -A75 CPU replaces A72 and A73
With the new ARM Holdings Cortex-A75 CPU core, the company is aiming to enable high-performance AI applications in smartphones and the datacenter as these processors can scale to the performance needs of these workloads. The new Cortex-A75 chips are ARM Holdings’ fastest chips for single-threaded applications and are designed to replace their A72 and A73 chips. The reason ARM gives as to how the A75 is replacing both chips is because the A75 is now capable of meeting the needs of the datacenter infrastructure chip vendors that would normally use an A72 and addressing the needs of the mobile-only A73. The A75 is fast enough that it will likely see lots of pickup among ARM Holdings SoC vendors who build chips for Chromebooks and Windows on ARM. ARM is claiming a performance uplift of over 50% compared to the A73 which has a lower per clock performance level as well as an overall lower top-end clock speed and is saying the A75 will clock up to 3.0 GHz, which is compared to the 2.4 GHz on the A73 which is currently in use in many smartphones. The A75 represents a tremendous amount of scale and if it works will be the only CPU I’m aware of that has successfully scaled that far.
New, smaller A55 CPU core
In addition to the Cortex-A75 CPU core, ARM Holdings also announced a new smaller, more efficient core, commonly known as the “LITTLE” cores as a part of ARM Holdings’ big.LITTLE multi-core processor design architecture. This new smaller core, the Cortex-A55, is what ARM Holdings calls its most efficient core ever and is designed to be used in conjunction with “big” cores like the A-75 to optimally balance performance and power consumption depending on the application.
big.LITTLE has matured a lot over the past 5 years and through those improvements, this approach has become a winner
In the past, many of ARM Holdings’ lower cost chip licensees used multiple Cortex-A53 clusters to create chips with 8-core CPUs. Likewise, they could theoretically do the same with eight A55 cores, however that may no longer make as much sense thanks to ARM Holdings’s new DynamIQ technology. That is because one of the benefits of DynamIQ is the ability to create single big.Little clusters of CPU cores that deviate from the traditional model of pairs or clusters of 2 and 4 cores, enabling new 3, 5 and 7-core clusters that utilize one A75 core paired with A55 cores to improve single-thread performance.
ARM Holdings announced DynamIQ earlier this year in March, but didn’t give many details about how it would work exactly or what processors would use it and why. That’s because new processors like the Cortex-A75 and A55 use ARM Holdings’s new DynamIQ technology to accelerate the performance of artificial intelligence, virtual reality, augmented reality and computer vision. DynamIQ technology enables ARM Holdings’s newest processors, when paired together to not only work together as a single cluster but also to accelerate artificial intelligence workloads. This is by creating DynamIQ clusters of up to 8 cores that utilize the new A75 and A55 cores along with a DSU (DynamIQ Shared Unit) which helps to optimized bandwidth and latency. The DSU reminds me a lot of what Apple talked about with their A10 Fusion “cache controller”. The traditional 4+4 core configurations are still possible with the new A75 and A55, but now they exist within a single DynamIQ cluster which gives them efficiency and latency advantages.
New Mali G72 GPU core
In addition to these new CPU cores and DynamIQ technology, ARM Holdings is also announcing their new Mali G72 GPU core technology. These new G72 cores are a step change over last year’s G71 currently in devices today, with optimizations for on-device machine learning. The G72, according to ARM, will deliver an efficiency improvement of 25% when compared to a G71 on the same process node. The same goes for their claim of 17% better machine learning efficiency, which is important for doing on-device machine learning. However, the improvements are not limited to machine learning only, ARM Holdings has also optimized the G72 for mobile VR with new features like Mobile Multiview, foveated rendering and MSAA (multi sampling anti-aliasing). These features will not only help to reduce the power consumed by running VR applications in a phone processor but also improve performance which will help to elevate the quality of mobile VR experiences.
For mobile VR to be successful it needs to attain a level of performance and efficiency that makes it palatable to both developers and consumers and the new G72 is yet another step in the right direction.
All of ARM Holdings’s biggest announcements at Computex 2017 are an indication of how seriously the company and its customers are taking machine learning and artificial intelligence. For that matter, it’s how seriously their customers and ARM take privacy, security and performance. Not just that, but that they intend to do a lot of the inference that machine learning and AI require to be done can be done increasingly on-device, helping to alleviate security and privacy issues. We can also see more workloads start to take advantage of the device’s local capabilities like Sensory to improve performance and ultimately user experience, which ultimately makes artificial intelligence more useful. Devices with these processors likely won’t be available in devices until next year, but we could see a handful of chip announcements this year. I would also expect that many of the chips that utilize these new cores will likely be using the leading-edge 10nm process nodes as well. This was a solid announcement for ARM Holdings, in step with industry as well as the needs of end users, and one step toward its goal of the next 100B devices in the next five years.