ARM improves the performance of its storage chip with the Cortex-R82

Last week, ARM announced the launch of a Cortex-R82 chip, which will be able to both store and process data. Such a processor could lay the foundation for a whole new generation of storage devices capable of processing the data they store. Storage controllers, such as those made by Marvell, Psion, and other array vendors like EMC, handle I/O and disk and SSD management, but still delegate the data processing task to the CPU. This requires a back and forth of data between two separate devices for processing. But there is a type of component called “processing storage” that can process data exactly where it is stored, without having to move it in memory. Different methods – by indexing and by schema, for example – can be used to process this data in the storage space. In-place processing eliminates the latency of moving data and frees up the CPU. Obviously this can only be done on SSDs.

Previous ARM-designed R5 and R8 lines already offer similar capabilities, but since these processors are 32-bit, they are limited to 4 GB of memory. The R82, the first 64-bit processor, is able to access a much larger memory space, 1 TB in this case. ARM does not manufacture chips. The company designs designs that it sells under license to manufacturers who take care of their development. According to ARM, the Cortex-R82 chip can integrate up to eight processing cores. Even better, due to chip consistency, the cores all see the same memory. The designer claims that compared to the R8 generation, the performance of the Cortex-R82 chip has doubled.

On-chip AI processing

The Cortex-R82 chip is also capable of handling machine learning tasks using ARM’s Neon technology, an advanced extension of the SIMD (Single Instruction Multiple Data) architecture that accelerates signal processing algorithms and functions to increase the performance of certain applications, in particular machine learning tasks. ARM claims that Neon technology accelerates neural network performance by up to 14 times compared to the previous generation R8 chip.

Another relatively important fact to note: the optional Memory Management Unit (MMU) allows rich operating systems, such as Linux, to be run on the chip separately from the main operating system. Thus, a storage array can have its own multi-core processor, memory space, and operating system to perform both storage processing and data processing completely independently of the main system. We therefore have two operating environments on the chip and we can assign cores to both tasks. This is particularly important for machine learning because the size of datasets is growing at an astonishing rate, so the storage and ongoing processing of data could overwhelm core processors. The fewer calls to main memory and the CPU or GPU, the better.

Coming soon

The Cortex-R82 chip does not do speculative execution, which means that its processing capacity is not equivalent to that of a Xeon or Ampere chip. But it can perform real-time processing which could be very useful for cleaning up datasets, processing raw data, offloading the main chip from initial data processing, and freeing up the CPU for processing relevant tasks. For the moment, ARM has not indicated when its Cortex-R82 chip will be available.