Contributed by Starblaze
Starblaze Taped Out its High-Performance SSD Controller SoC Based on Arm Architecture
One of the few bright spots of the COVID-19 pandemic is that we have witnessed an accelerated pace of digital transformation and automation to enable efficient remote work. Digital transformation has leaped forward by three to five years as organizations around the world ramp up their critical data-centric infrastructure to enable growth, innovation, and experiences in the new reality.
The rapid growth of data, coupled with the ever-increasing desire to make decisions based on analytics, has propelled innovators to push the envelope of digital infrastructure including compute, storage and networking. Arm, with our unique business model of licensing semiconductor technology building blocks, is enabling our silicon partners to innovate across a wide range of applications of the digital infrastructure. One of the key trends enabling the digital infrastructure is storage – specifically enterprise and cloud solid-state drives (SSDs), and one of our partners, Starblaze Technology, has built a futuristic product on Arm.
3 Key trends in cloud and enterprise storage and disk drives
1. Larger capacities and more specialized storage tiers
Hyperscale and cloud infrastructure need to carry a variety of workloads, which pose different requirements on storage, resulting in a slew of storage options in the cloud: object, file, block, data analytics, archive, database optimized storage, and more. The use-cases derived from these different storage tiers enable different demands on storage products.
Short video formats such as Tiktok, Youtube are one of the key driving factors for increased demand for SSDs due to their predictable performance, lower latency and high throughput. According to Gartner and Wells Fargo analysts, enterprise SSDs will see strong growth from 2021 to 2026 at a ~37% Compound Annual Growth Rate (CAGR), with total SSD capacity shipped growing to 32 percent of overall HDD/SSD capacity shipped by 2026 (vs its approximate ~20 percent share in 2021).
Hard Disk Drives (HDDs) are also growing at a healthy rate and continue to be an important storage medium of choice due to continued innovation on capacity and features focused on the cloud use case. Features such as Shingled Magnetic Recording (SMR) for improved density and capacity, dual actuator for performance, and SATA command duration limits (CDL) can help cloud operators to enhance performance, capacity, operational expenses, and Total Cost of Ownership (TCO).
At the disk level, SSDs with large capacities of 16TB and 32TB are becoming available now. To ensure best performance for these large-capacity SSDs, the Flash Translation Layer (FTL) tables in SSD controllers need 16GB to 32GB DRAMs if we assume 1GB DRAM per 1TB NAND. This means that the processor within the SSD controller needs to be able to support larger addressable memory. Arm Cortex-R82 supports 40 bits addresses up to 1TB of DRAM in a flat address map, while Cortex-A series are built with Memory Management Unit (MMU) to enable wider addressing.
2. Higher performance interfaces
One of the key design goals for SSD controllers is to saturate the host interface bus for maximum bandwidth. Majority of the new enterprise SSD controllers coming onto the market support PCIe Gen4 at 16GT/s per lane. Going forward, the advancement of Compute eXpress Link (CXL) is driving rapid transition to PCIe Gen5, which doubles the host interface throughput to 32GT/s per lane, and then PCIe Gen6 to 64GT/s per lane. CXL can empower SSDs, especially those with computational capabilities, to efficiently access host memory or cache with coherency. Higher SSD performance and throughput would need faster SSD controller processors both in the front end to handle transaction requests from the host and run more sophisticated wear-leveling algorithms and in the back end to manage the interface with media, the NAND channels.
3. More compute near storage
When the world shifts from compute centric to data centric as both the volume and importance of data grows, compute capabilities need to move closer to storage. As capacities increase, computational storage is gaining traction to improve network efficiency to process data at the source.
Even with the high-performance interfaces, the data movement can result in higher power consumption and higher latencies because the host CPUs need to spend compute cycles receiving the large volume of data and sending the results back to storage.
The thinking behind having more compute next to storage is in line with the Data Processing Unit (DPU) concept in SmartNICs, and we have already seen early products of Computational Storage Drives (CSDs) being used in hyperscale data center environments. The most common workloads to be offloaded to CSDs are video transcoding, AI inferencing, data compression/decompression, encryption/decryption and distributed database processing accelerations and many more.
The other advantage of adding compute into storage is that storage processing itself can become smarter. NAND flash media for SSDs have seen significant innovations from Single-Level Cell (SLC), Multi-Level Cell (MLC), Triple-Level Cell (TLC) to Quad-Level Cell (QLC) and even Penta-Level Cell (PLC) under development. This evolution increases storage density while decreasing performance, endurance, and cost, resulting in larger, cheaper but potentially lower performance and shorter life cycle SSDs. To compensate for differences, SSD controller hardware and firmware need to become smarter to take full advantage of these storage media innovations and use modern methodologies like machine learning to build better data placement methods as well as run more intensive wear-levelling algorithms. These advancements can have huge impact on key storage metrics such as random write IOPS, latency, quality of service and endurance.
Starblaze STAR2000: A next-gen SSD controller built on Arm
Recently, Starblaze Technology announced that its new generation of high-performance enterprise-class PCIe4.0 SSD controller chip STAR2000, designed with Arm Cortex-A55 multi-core processor and Arm China’s STAR-MC1 processor, has successfully taped out. The product will be launched to the market in the second half of 2022. Starblaze and Arm share common views over the direction of cloud, enterprise storage and SSD markets – making Arm IP the natural selection for the Starblaze STAR2000 SSD controller SoC.
STAR2000 is the first high-performance enterprise-level SSD controller in the China market that uses Arm’s Cortex-A55 CPU. Starblaze Technology chose Arm Cortex-A55 for its features and functions that enterprise SSD controllers focus on. Arm’s Cortex-A55 provided STAR2000 with a flexible and efficient multi-core consistency management mechanism, excellent instruction execution efficiency, extremely low cache access latency (shared L3), powerful input and output capabilities and computing performance. All these capabilities of Cortex-A55 helped the STAR2000 achieve higher performance efficiency and better compute performance to support the complex firmware of enterprise SSDs.
In the 8-core Cortex-A55 cluster, 4 cores are dedicated to FTL processing, while the other 4 can be used to run rich operating systems such as Linux. This enables easy offloading of host workloads onto the STAR2000 controller, as well as running more intelligent and sophisticated data placement and wear-leveling algorithms. Starblaze also included a Neural Network Processor (NPU) accelerator with 8 Tera Operations Per Second (TOPS) compute capability in its STAR2000 design to run AI/ML workloads more efficiently, which includes both SSD control operations themselves as well as compute offload. It can also help SSDs improve reliability and service quality, optimize power consumption, and realize intelligent self-checking and early defect detection.
The STAR-MC1 processor developed by Arm China is used for flash memory channel control, enabling STAR2000 to fully support various types of flash memory interface protocols and ensure the efficiency of flash memory access, and achieve excellent performance, power and area (PPA) balance.
STAR2000 is a high-performance SSD controller SoC for the data center and enterprise market. It adopts 12nm process and fully supports NVMe2.0 protocol. It meets the stringent requirements of enterprise-class SSDs in terms of reliability, fault tolerance and correction.
Starblaze is an early developer of high-performance enterprise SSD controller SoCs in China. It has demonstrated full-stack design capabilities in consumer-grade, enterprise-grade, and industrial-grade SSD controllers as well as intelligent storage solutions to solve the challenges around basic computing capabilities, I/O performance, and real-time responsiveness. On the cusp of the PCIe 5.0 SSD era with higher speed and higher bandwidth requirements, Starblaze has started scoping its next-generation PCIe Gen5 SSD controller SoC, to build on its profound domain expertise in the China local storage market and stay on the bleeding edge of innovation.
“In the future, Starblaze Technology will collaborate with Arm and Arm China to continue technological innovation to meet the needs of consumer and enterprise storage, data center, edge computing and other markets with higher performance and richer product portfolio. The parties will jointly strengthen the Arm ecosystem in the China storage market and continue to expand the ecosystem influence of China SSD space.” Sky Shen – CEO, Starblaze
Arm is committed to continuing its support for innovators such as Starblaze who choose to build their future on Arm architecture. For more information on Arm’s Storage solutions, please visit our webpage.