Micron 9100 Max NVMe 2.4TB SSD Review

Teardown, Architecture And CPU Utilization

Teardown

Architecture

Image 1 of 4

The Micron 9100 MAX comes in the standard HHHL (Half-Height Half-Length) AIC form factor, which is an immediate advantage over large FHHL (Full-Height Half-Length) competitors, such as the Mangstor MX6300, because the small form factor will slot into a wide range of servers. The 9100 connects via PCIe 3.0 x4. When we view the SSD from an angle, we note the internal air channel that leverages the linear airflow patterns within the server, thus boosting heat dissipation.

Like many other enterprise SSDs, the 9100 will begin throttling at 85°C. SSDs have progressed in this respect; now they initiate a dynamic throttling curve to reduce temperature to avoid a steep performance drop. The 9100 AIC model requires 300 LFM of airflow, whereas the U.2 (2.5") requires 450 LFM. This is the only differentiation between the two models, other than the different form factor. Micron also allows users to adjust the thermal profile of the 9100 series by selecting either the MAX, 25W or 20W thermal modes. The 25W and 20W modes restrict performance by limiting power consumption, which in turn reduces the thermal envelope. This allow users to customize the SSD for dense and coiling-challenged environments. This feature is not currently active (Micron will enable it soon), so we are not able to provide various test results at each thermal threshold.

The I/O bracket features a constellation of small holes that allow air to flow somewhat unimpeded out the rear of the server, which encourages thermal dissipation. We also spot two rectangular slots that house the two LEDs that provide a visual status indicator. The two liquid-filled electrolytic capacitors jut up near the bracket. These capacitors provide enough capacitance to flush data in flight from all caches and buffers to the non-volatile NAND in the event of an unsafe host power-loss event. The 9100 requires two minutes upon reboot after a power-loss event to resume normal operation. The 9100 features the typical 2 million hour MTBF and 1 per 10^17 UBER ratings.

Image 1 of 6

The Microsemi Flashtec NVMe1032 89HF16P04CG3 16-channel SSD controller comes into view when we remove the heatsink. This mature controller debuted in 2014 and continues to power a broad range of SSDs. Microsemi allows its customers wide latitude with the firmware, which fuels an incredible amount of differentiation between competing products.

The previous-generation Micron products utilized a 32-channel controller that supported four targets (LUN) per channel, but the Flashtec controller supports up to eight targets for each of the 16 channels. This provides a similar amount of parallelism through the benefits of interleaving. Each target can consist of several NAND packages to increase density. Each of the 9100's 32 x 16nm NAND packages features eight 128Gbit NAND die per package, and each channel serves two individual packages.

The SSD also houses nine 512MB DRAM packages, but only 4GB of DRAM is available for use due to the ECC arrangement, which adds an extra layer of protection. The 9100 uses thermal pads on all of the critical components, such as the ASIC, NAND and DRAM, to facilitate cooling. Micron uses a 31:1 RAIN ratio for SSDs larger than 1TB, and 15:1 for SSDs 1TB and under.

QoS Specifications

The SSD vendors are beginning to focus on more refined aspects of flash performance as SSDs, and our understanding of how applications extract data from them, evolves. In the past, performance specifications were relegated to basic random and sequential speeds and feeds, but now the major vendors also include QoS (Quality of Service) measurements to quantify performance consistency.

Micron joins the ranks of vendors focusing on the all-important metric, and according to our measurements (which we cover on the following pages), the company has conservatively stated its QoS specifications, which is somewhat of a Micron trait. We routinely measured values that were much lower, and thus better, than the rated specifications.

CPU Utilization

Image 1 of 4

All of our test samples feature the NVMe interface, and the developers focused on reducing CPU usage during the protocol design. As such, there is little variation between our standard NVMe samples. The Intel DC P3608 is the lone exception to the rule, because it requires an RSTe-based RAID 0 implementation of its two separate internal volumes, which increases its host CPU consumption.