ARM debuts Cortex-A75 and Cortex-A55 with AI in mind
May 29, 2017 — by Eric Brown 3,645 viewsARM unveiled three chip designs that support DynamIQ load sharing: an up to 50 percent faster Cortex-A75, a Cortex-A55 companion, and a Mali-G72 GPU.
Two months ago when ARM announced its DynamIQ multi-core load balancing heir to its Big.Little scheme, it claimed that DynamIQ-enabled SoCs would be optimized to deliver up to a 50 percent boost in AI performance over the next 3-5 years relative to today’s high-end Cortex-A73 based systems. Now, the Softbank-owned chip IP designer has unveiled the first ARMv8 chip designs to support DynamIQ.
The Cortex-A75 follow-on to the year-old Cortex-A73 “delivers a massive 50 percent uplift in performance and greater multicore capabilities,” claims ARM. It’s designed to work in DynamIQ Big.Little configurations with the lower powered Cortex-A55, which replaces the older Cortex-A53, which has appeared in 1.5 billion products to date. The Cortex-A55, which will likely appear on its own in dual- and quad-core clusters on the hacker boards of late 2018 and 2019, offers “dedicated AI instructions and up to 2.5x the performance-per-milliwatt efficiency” compared to the Cortex-A53, claims ARM.
— ADVERTISEMENT —
The new chip designs may first appear in 1Q 2018 in a hexa-core Qualcomm Snapdragon 845 SoC, according to a recent rumor from MyDrivers, reported on HotHardware. The Snapdragon 845 is said to combine a pair of Cortex-A75 cores with four Cortex-A53 cores and an Adreno 630 GPU. The 10nm-fabricated Snapdragon 845 is rumored to include a Snapdragon X20 LTE modem, and support faster LPDDR4X RAM and UFS 2.1 storage.

Typical SoC combining Cortex-A75 and A55 and Mali-G72
(click image to enlarge)
ARM’s own preference for a GPU pairing with the Cortex-A75 and -A55 is its own newly announced Mali-G72, which is designed to work closely with both architectures using DynamIQ. This second-gen Bitfrost architecture GPU, which follows last year’s Cortex-A73-paired Mali-G71, is said to improve performance while “reducing bandwidth for a 17 percent machine learning efficiency gain.”
ARM repeatedly promoted the new designs as AI chips, which may be overstating things a bit. True, there are some AI extensions here, and the designs offer greater performance, which is demanded by artificial intelligence technologies like machine learning, visual computing, and voice-activated agents like Alexa, Google Assistant, Watson, Cortana, and Siri. In addition, DynamIQ, which enables more flexible core configurations that were not possible with Big.Little, makes it easier for CPU cores to better collaborate with the GPU and other co-processors that are increasingly tapped for AI.
Yet, this is not a radical new AI redesign such as Google’s cloud-server oriented Tensor Processing Unit (TPU) chips or Apple’s upcoming Apple Neural Engine. (No doubt, ARM has its own true “AI chip” in the pipeline.) Rather, these are smartphone and high-end embedded and automotive SoCs that are designed to offload some of the AI algorithms currently processed in the cloud to edge devices ranging from IoT gateways to laptops. This new generation also raises the bar for RISC-V newcomers like SiFive, which could pose significant new competition for ARM to add to its ongoing battle with Intel and AMD.
Inside the Cortex-A75
The Cortex-A75 supports up to quad -A75 clusters at up to 3GHz with 10nm fabrication, and offers “significantly improved integer performance, and substantial enhancements in floating point and memory workloads” compared to the Cortex-A72 and -A73, says ARM. Yet, it also retains the same power efficiency levels, claims the chip designer.

Cortex-A75 architecture
(click image to enlarge)
Thanks to an advance to ARMv8.2 and the new DynamIQ, the memory subsystem has been improved with benefits like access to an optional shared cluster L3 cache from 512KB to 4MB. Each core also has a private, 256KB to 512KB L2 cache with half the latency of “traditional high performance processors,” as well as 64KB L1 I- and D-cache allotments, says ARM. Other memory subsystem features include support for asynchronous frequencies, and “potentially independent voltage and power rails for individual CPUs or groups of cores,” says ARM.
The Cortex-A75 is part of a Sophia family of core designs that started with the 32-bit Cortex-A17 and was followed by the Cortex-A73. “ARM’s focus has shifted from improving power efficiency and thermal headroom for A73 to improving performance and adding new features for A75,” says a May 29 AnandTech report on ARM’s new chips. The enhancements to integer processing “allows the A75 to be more speculative, improving its ability to execute instructions out of order,” says the story.
Cortex-A55
While the Cortex-A75 is more about performance than efficiency, the Cortex-A55 delivers up to 18 percent more performance at 15 percent better power efficiency compared to the Cortex-A53. In addition to supporting smartphones, the up to octa-core ready design can be used in IoT gateways, high-end wearables, and other automotive, consumer, and embedded applications.

Cortex-A55 architecture
(click image to enlarge)
The Cortex-A55 provides a configurable L2 cache from 64KB to 256KB that provides more than 50 percent reduced latency for memory access compared to the -A53. It also provides an optional L3 cache (512KB to 4MB) that can be shared across up to eight cores. L1 caches are configured at 16KB to 64KB. DynamIQ technology will enable designs such as 7x or 3x Cortex-A55 with a single Cortex-A75.
The Cortex-A55 also adds support for Virtual Host Extensions (VHE), which AnandTech says will prove important for automotive and industrial safety and reliability. A new Int8 dot product instructions offering is said to be useful for accelerating neural networks.
Mali-G72
The up to 850MHz Mali-G72 builds upon the major breakthroughs of ARM’s first Bitfrost-architecture GPU, the Mali-G71, including support for up to 32 shader cores and support for 16nm FinFET fabrication. Improvements include increased tile buffer memory for up to 16x multi-sample anti-aliasing processing “at minimal performance cost,” says ARM.

Mali-G72 architecture
(click image to enlarge)
Complex machine learning and mobile gaming applications can run algorithms at up to 25 percent higher energy efficient and 20 percent better performance density, claims ARM. This is said to lead to a 40 percent greater overall performance compared to the Mali-G71.
The new Cortex-A designs are supported with ARM’s fully suite of computer, media, display, security, and system IP. These features are supported by the ARM Compute Library and a range of new System Guidance for Mobile package offered free to ARM partners. As usual, ARM TrustZone is provided for security.
Further information
The new Cortex-A75, Cortex-A55, and Mali-G72 designs appear to be available now to ARM partners. No availability information was available for final silicon, but based on the Snapdragon 845 rumors, it appears we’re looking at Q1 2018. More information may be found in ARM’s announcement, as well as its Cortex-A75, Cortex-A55, and Mali-G72 product pages.
very interesting, is there a possible build of Rasperry or odroid with that in the near future? I really enjoy the idea about such processors in projects!