The final main replace to our client SSD benchmark suite was launched in 2017. So much has modified within the SSD market over the previous a number of years, and our SSD testing strategies want to vary in response to these tendencies. SSDs typically have been getting extra complicated, so there are extra methods for benchmarks to finish up being deceptive or unrealistic. This could be on accident, or as a part of a deliberate PR technique to obfuscate the actual world efficiency behind a ‘holy’ unified metric. [pun intended]
The nature of testing {hardware} has additionally modified, particularly in mild of the necessity to assist the most recent applied sciences like PCIe 4.0, We additionally discover it related to extra carefully resemble the sorts of {hardware} our readers are actually more likely to be utilizing. This article introduces the brand new 2021 version of our client SSD check suite. It covers the objectives and rationale behind the design decisions, the {hardware} and software program configurations used, and the main points of every benchmark and its goal. It is our hope that this new check suite can be appropriate for the following a number of years of SSD evaluations.
SSD Benchmarking: The AnandTech Heritage
To begin, we have to evaluate a bit about how SSDs work. Anand’s basic articles from 2009, reminiscent of The SSD Anthology and The SSD Relapse, doc lots of the early challenges SSDs needed to overcome on the way in which to taking on because the mainstream client storage expertise. More than a decade later, these articles are nonetheless beneficial background info on how at present’s SSDs work. All of the elemental challenges of coping with NAND flash reminiscence nonetheless apply: the shortcoming to make in-place modification to knowledge saved in NAND flash, the disparity between NAND web page sizes and erase block sizes, the necessity for put on leveling and the ensuing efficiency drops when a drive is full are all nonetheless related. Today’s SSDs are a lot greater, quicker and cheaper, however their controllers and firmware nonetheless must cope with these points. Today’s SSDs additionally layer on some additional challenges that make their inside workings much more sophisticated, and require extra subtle testing to correctly assess.
Anyone keep in mind Intel’s X25-M, utilizing 50nm NAND?
When I began reviewing SSDs for AnandTech in 2015, TLC NAND and NVMe have been solely starting to creep into the patron market: TLC on the low finish, and NVMe on the excessive finish. Both have now taken over: MLC has all however disappeared from the market, and NVMe is the default selection for brand spanking new techniques. (In Micron’s most up-to-date earnings name, they shared that their consumer SSD shipments are actually over 90% NVMe by variety of bits shipped.) The larger efficiency allowed by PCIe and NVMe reasonably than SATA make beautiful variations in benchmark scores, however the migration to TLC was really extra vital from the angle of benchmark design. That’s as a result of client TLC SSDs rely closely on SLC caching.
In a drive, the extra bits you write to a cell, the extra complicated (and slower it’s). Modern drives run three bits-per-cell, generally known as TLC, and Four bits per cell, generally known as QLC. Both of those are slower for writes than 1 bit per cell, or SLC. By treating a portion of the drive cells as one bit per cell, generally known as ‘SLC mode’. this permits the drive to soak up a burst of writes extra rapidly. This is as a result of programming the voltage of an SLC cell could be achieved in a single step and would not require as a lot precision as when storing a number of bits per cell.
The draw back is that knowledge written to SLC NAND should later be re-written into blocks which can be working as MLC/TLC/QLC, storing a number of bits per cell. This course of (typically referred to as folding) is often routinely carried out by the drive throughout idle time, the place latency will not be essential. This freee up the area within the SLC cache for future use. SLC caching is usually not utilized by…
