Introduction

As one of the world leaders in digital technology, Samsung pretty much makes any type of electronic device you can think of. Their products are used by millions of people around the world.

Being a leader in DRAM and flash memory production, it comes as no surprise that they are also a huge player in the SSD business. Their EVO and PRO Series SSDs are highly popular among upgraders, system builders, and enthusiasts.



The Samsung 990 Pro was announced late last year. While everybody expected it to be a PCI-Express 5.0 solid-state-drive it's actually built using the PCI-Express 4.0 interface. While the 980 Pro used a controller called "Elpis," the 990 Pro is built using "Pascal"—not much else is known, except that Pascal is optimized for energy efficiency and uses the same 8 nanometer production process as Elpis. The NAND flash has been upgraded, too, to the company's seventh generation 176-layer 3D TLC NAND (the 980 Pro uses 128-layer flash). As expected for a high-end drive, a DRAM cache chip is included, too.

The Samsung 990 Pro is available in capacities of 1 TB ($80) and 2 TB ($150), a 4 TB version was promised last year, but never materialized. Endurance for these models is set to 600 TBW and 1200 TBW, respectively. Samsung includes a five-year warranty with the 990 Pro SSD.

Packaging

The Drive

The drive is designed for the M.2 2280 form factor, which makes it 22 mm wide and 80 mm long.


PCI-Express 4.0 x4 is used as the host interface to the rest of the system.


On the PCB you'll find the controller and two flash chips, a single DRAM cache chip is included, too.


On the back of the drive a sticker with copper is installed, acting as heat spreader. The front sticker is just thin plastic.

Chip Component Analysis

This is Samsung's new PCIe Gen 4 controller called "Pascal." It is produced on a 8 nm production process in Samsung's foundries.


The two flash chips are Samsung 176-layer V-NAND v7 TLC. Each chip has a capacity of 1 TB.


One Samsung LPDDR4-4266 chip provides a total of 2 GB of fast DRAM storage for the controller to store the mapping tables.

Test Setup

Synthetic Testing

• Tests are run with a 20-second-long warm-up time (result recording starts at second 21).
• Between each test, the drive is left idle for 60 seconds, to allow it to flush and reorganize its internal data.
• All write requests contain random, incompressible data.
• Disk cache is flushed between all tests.
• M.2 drives are tested with a fan blowing on them; that is, except for the results investigating uncooled behavior on the thermal testing page.

Real-life Testing

• After initial configuration and installation, a disk image is created; it is used to test every drive.
• Automated updates are disabled for the OS and all programs. This ensures that—for every review—each drive uses the same settings, without interference from previous testing.
• Our disk image consumes around 600 GB—partitions are resized to fill all available space on the drive.
• All drives are filled with random data to 80% of their capacity
• Partitions are properly aligned.
• Disk cache is flushed between all tests.
• In order to minimize random variation, each real-life performance test is run several times, with reboots between tests to minimize the impact of disk cache.
• All application benchmarks run the actual application and do not replay any disk traces.
• Our real-life testing data includes performance numbers for a typical high-performance HDD, using results from a Western Digital WD Black 1 TB 7200 RPM 3.5" SATA. HDDs are significantly slower than SSDs, which is why we're not putting the result in the chart, as that would break the scaling, making the SSDs indistinguishable in comparison. Instead, we've added the HDD performance numbers in the title of each test entry.