Choosing Memory Cells for SSDs

This blog post, published on April 18, 2022, was written by Sebastien Jean, CTO, Phison Electronics Corp.

As enterprises seek to enhance performance, efficiency, and reliability in today’s data centers, they are increasingly turning to SSDs over HDDs. Finding the right SSDs for their needs, however, isn’t as straightforward as some might think. SSDs come in a range of form factors and configurations, and can be highly customized.

Many factors go into customizing an SSD for specific applications. Each company has differing priorities on which SSD aspects are most important. Some enterprises require the highest level of performance with the least amount of latency-for applications such as real-time banking transactions, for instance, or crunching through massive ML datasets. Others prioritize low power consumption for energy savings, high endurance for a longer lifecycle or cost efficiency to align with strict budgets.

With so much variety in hardware and software design options for SSDs, enterprises should first determine their storage and server needs and then look for a vendor that can deliver SSDs with their exact requirements in mind.
 
Customizing SSDs for performance and endurance
Factors such as flash memory type, block allocation and connection interface can have a big impact on an SSD’s performance.

One of the most straightforward ways to improve performance without any negative side effects or impacts on endurance is to simply choose the best NAND available. Not all NAND is created equal, and flash memory vendors are continually improving and upgrading their NAND products to beat the competition. It pays to do some research before purchasing. Phison’s engineering teams work in partnership with the leading NAND wafer manufacturers. It evaluates new NAND technologies before they ship and then validates NAND for use in storage products. Firm’s expertise in NAND helps its customers select the optimal NAND for use in their storage applications.

 
NVMe vs. SATA
How an SSD connects to a server’s motherboard can increase throughput and therefore performance. When every enterprise was using HDDs in their data centers, the most common connection interface was SATA, or SATA. The latest version of the interface, 6Gb SATA, has a maximum bandwidth throughput of 600MB/s. SAS is popular, too, but it also relies on the SATA physical interface. Though the SAS interface is 2x as fast, the physical platter capability is roughly the same as SATA.

In 2011, as SSDs became more widely accepted, a new storage protocol was created, called NVMe. It quickly became the industry standard for connecting SSDs to a motherboard, because it uses the PCIe 3.0 bus, which is approximately 6x faster than SATA. That is because SSDs typically use x4 PCIe lanes, and they were achieving 3,500MB/s throughput compared to 600MB/s in 2011. Today, the same x4 lanes provide 14GB/s throughput using Gen5 signaling. Unlike spinning platters, digital NAND flash memory can support the full interface bandwidth. In addition to being fast, NVMe PCIe SSDs can achieve shorter data access and command queuing latency. NVMe SSDs also have excellent multitasking capabilities. All these improvements result in greater performance.
 
Memory cell type
SSDs are made up primarily of a flash controller and NAND flash memory cells that store the data. These memory cells come in several configurations that determine their speed and how many bits of data the cell can accept.
• SLC SSDs are the most basic configuration. Each cell accepts 1 bit of data. The advantage of SLC SSDs is their speed: they’re the fastest SSDs available. They’re also the most reliable and durable SSDs – and not surprisingly, the most expensive. Today all-SLC configurations are reserved for specialized applications such as the movie industry, medical imaging, high frequency trading and space-based sensor platforms.
• MLC SSDs accept 2 bits of data per cell. These are slower than SLC SSDs because writing 2 bits in a cell takes longer than writing 1 bit. They’re also less reliable and durable because more data is written on the cell and memory cells degrade with repeated use over time. This configuration is now largely deprecated in favor of TLC and QLC.
• TLC SSDs accept 3 bits of data per cell. As the most common type of enterprise SSD today, they offer greater capacity at a lower price than SLC. The tradeoffs, however, are reduced speed, reliability and durability. Though to give TLC its due, it is the baseline for all enterprise storage. SLC is considered a specialized NAND and QLC is used for cost-optimized or read-intensive applications.
• QLC SSDs accept 4 bits of data per cell. While the price positioning of QLC NAND is below that of TLC, current gens of QLC have comparably lower performance and durability vs TLC. Current gens of QLC NAND are a great fit for use in read-intensive applications that have minimal write I/O.
• Penta-Level Cell (PLC) SSDs encode 4 bits of data per cell. While not yet commercially available, they will offer larger storage capacities per NAND die, and lower cost-per-gigabyte in price at the retail level-at reduced performance and durability, however, vs. QLC or TLC NAND. The practical applications for PLC NAND are for infrequently used USB flash drives or flash memory cards that are roughly in line with HDD performance.

The challenge with all these memory cell types is figuring out how to balance the customer application’s need for performance, reliability and durability. Features such as read speed, write speed, program/erase (P/E) cycling, data retention and active power have direct impact on each one of those qualities.

For instance, mainstream enterprise storage today primarily uses TLC NAND. In comparison, an all SLC NAND SSD is 3x more expensive to make, but yields excellent low-latency performance. Target applications include high-frequency stock market trading or write-heavy applications like 8K video drones used for recording Hollywood action scenes. QLC NAND is typically priced under TLC NAND and provides 33% more storage than TLC in the same NAND die. QLC NAND works well for data-intensive applications like oil surveying, developing new jet turbines or performing genomic studies. Matching the NAND to the application is only one of the factors that must be considered, as we will see below.
 
Over-provisioning
NAND flash memory cannot be directly overwritten like magnetic media-the old data that is still valid must first be moved to a new block before the old block is erased and able to be re-written. The process of collecting data that is still valid is called “garbage collection”, though in this case it is referring to clearing out the old invalid obsolete data and collecting the use data that is valid. Erasing blocks is quite slow compared to read or write operations. If not handled correctly the erase operation slows down performance and pushes out the latency distribution. Over-provisioning (OP) the SSD by providing spare capacity that is not directly addressable by the user allows more time for data to naturally invalidate as new data is written. This reduces the garbage collection overhead and drive wear, while tightening up the latency distribution curve. Having extra OP allows the system to carry out program/erase cycles without disrupting or slowing down system performance.

Increasing OP also boosts steady state performance, diminishes Write Amplification Factor (WAF) and actually increases SSD endurance. OP decreases user capacity, but that can be offset by adding more NAND, which unfortunately increases both cost and active power.
 
Parallel processing
An HDD has only 1 or 2 sweeping arms managing the R/W head position relative to the media. Though each disk face has a dedicated head, all heads on the same arm must look at the same sector on each disk face, which greatly limits random performance. In contrast, a 4TB SSD has 64 independent NAND die, each with 2 to 6 independent planes. This means that an enterprise SSD can process up to 384I/O commands simultaneously every 60 to 80μs. This is how an enterprise SSD achieves 1 million to 2 million IO/s in comparison to HDD, which are limited to 150 to 400 IO/s. To put these numbers into perspective, an eye blink is approximately 1/3 of a second.

Additional improvements are possible when looking at the storage controller design. A typical enterprise controller will have 6 to 8 large processing cores, while Phison designs focus on 32 to 48 micro-cores. Though it takes additional effort to design a more complex processor, it brings several benefits including additional parallelism, reduced latency and improved power.
 
Customizing SSDs for power usage
As enterprises try to reduce their carbon footprint, the data center provides several opportunities to save energy and lower costs. One option is to replace power-hungry HDDs with more efficient SSDs, though some SSDs are more efficient than others. Beyond simply replacing rotating media for digital media, SSDs can be customized to reduce power consumption in several additional ways.

Client SSDs based on the Phison E18 include a power management feature that provides up to 20% of additional battery life while playing high-end games on laptops. The company’s Enterprise X1 controller enables SSDs with performance while reducing power by 30%.

One strategy to manage power consumption of the storage controller is to migrate to a smaller process node. Decreasing the process node size from 28nm to 12nm, for instance, allows the node to operate at higher frequencies with lower voltage used. By using less energy to move data over a bus or toggle the transistors, it reduces the energy used by the SSD. Using less power means the SSD generates less heat, which in turn reduces the transistor leakage current.

Another way to decrease power requirements is to reduce how many NAND channels an SSD uses. That capability is enabled by an improved ONFI bus speed, which is used to move data from the NAND to the SSD controller. Now you don’t need 8 or 16 channels to saturate the Gen4 and even Gen5 PCIe interface. You can potentially saturate the host interface with just 4 NAND channels. Reducing the number of back-end channels decreases the total SSD power by 20 to 30%.
 
Phison SSD customization partner
It takes expertise to develop customized flash storage solutions such as SSDs. Smart enterprises find partners that have that knowledge and expertise to create custom components.

Since producing the first single-chip USB flash controller 20 years ago, Phison has continued to grow and expand its knowledge base. It acts as an on-demand engineering service for its partners, and owns all the critical IP for NAND flash controllers in-house. With such an experience, the company offers greater flexibility when customizing products for OEM clients and enterprises.