Flash-based removable media, such as SD cards, have a host of use-cases in products ranging from content capture devices to portable game consoles. Many computing systems (including PCs and smartphones) also employ them to augment the available storage capacity. There are two main standards bodies in this segment - the SD Association (SDA) and the CompactFlash Association (CFA), with the SDA is responsible for the Secure Digital cards (SD / microSD), and the CFA has the CompactFlash and CFast card markets.

The products currently servicing this market segment are based on technology standards drafted when SATA was the pinnacle of internal storage performance, but in fact both standards bodies jumped on to the NVMe bandwagon in the late 2010s. Removable memory cards based on PCIe / NVMe are slowly starting to appear in the market - CFexpress from CFA, and SD Express from the SDA. Both ADATA and Lexar have announced plans to launch their SD Express cards within the next few quarters. The cards from both vendors are based on the Silicon Motion SM2708 controller.

Silicon Motion sampled us its SM2708 reference design (with a 250GB capacity) to put through our rigorous memory cards evaluation suite. This review takes a detailed look at the performance of the card in conjunction with Realtek's RTL9211DS card reader reference design. It also serves as a preview of what consumers can expect from SD Express cards appearing in the market over the next few quarters.

Lets Go: The Rise of PCIe and NVMe

SD cards and CompactFlash cards have emerged as the storage media of choice for content capture devices such as digital cameras and camcorders. SD cards have also enjoyed support in the portable game console space and as boot drives for single-board computers. Currently popular SD and CF cards are based on standards developed when SATA ruled the roost as the transfer protocol of choice for internal storage devices.

As PCIe-based NVMe took the SSD storage market by storm, both the SDA and CFA introduced removable card standards based on PCIe. The CFA had actually jumped on to PCIe quite early, with a 1Gbps standard in 2011 - the XQD card format. However, lack of backwards compatibility with existing CompactFlash readers meant that the format never really took off despite making its way into retail. While the original CF cards were based on PATA (the precursor to SATA), a faster version based on SATA was introduced in 2009 as CFast. In 2016, the CFA announced plans for the CFexpress standard and it was published in mid-2017. CFexpress cards retain the XQD form-factor, and a few CFexpress cards are already in the market.

SD cards are also popular in multiple market segments. In order to serve the needs of all serviced markets, the SD Association introduced the NVMe-based SD Express standard (SD 7.0) in 2018, with a SD 8.0 follow-up in 2020. SD cards as well as card readers based on these new standards have been making the rounds at various trade shows since 2019. However, none went on to appear in the retail market. That is about to change in the coming months, with both ADATA and Lexar announcing plans to launch their SD Express cards based on the Silicon Motion SM2708 card controller within the next few quarters.

What We're Testing Today: SD 7.1 Reference Designs

Consumers need both SD Express cards and card readers in order to take advantage of all the benefits of PCIe / NVMe in the SD form-factor. While SD Express cards are being enabled by controllers such as the SM2708, card readers platforms based on JMicron and Realtek have also appeared in various trade-shows (such as the RTS5261 demonstrated at Computex 2019). Silicon Motion coordinated with Realtek to sample their 250GB SM2708 reference design along with a card reader based on the the Realtek RTL9211DS and RTS5261 for this review.

The Silicon Motion SM2708 card controller supports both SD UHS-I and PCIe Gen 3.0 x2 on the upstream side. On the flash side, both Toggle 3.0 and ONFI 4.1 NAND interfaces at 800 MT/s are supported. The controller is restricted to two-channel operation with 8 enables per channel, and both 3D TLC and QLC can be used.

Silicon Motion has qualified the SM2708 for operation with SanDisk / Kioxia 96L TLC (2-plane flash) and Micron B27B (96L) and B47R (176L) (both 4-plane flash). With 4 plane flash, Silicon Motion claims write performance of more than 700 MB/sec with four NAND flash pieces. The write performance is highly influenced by the number of flash pieces, NAND tPROG, and plane numbers.

The 250GB reference design sampled for review comes with the SanDisk / Kioxia BiCS 4 96L TLC 2-plane flash.

The Realtek card reader is a dual-chip solution that uses the RTL9211DS and RTS5261. The RTL9211DS is more commonly seen in storage bridges that enable M.2 NVMe SSDs to be used with a USB host. It has a USB 3.2 Gen 2 (10 Gbps) upstream interface, and a PCIe 3.0 x2 downstream interface.

In the card reader, the functionality is the same - in SD Express mode, the PCIe lanes directly connect to the SD card pins (with the RTS5261 PCIe pins acting in bypass mode). However, in legacy mode, the UHS-I interface operation is implemented with the RTS5261, and the RTL9211DS only acts as a host interface component. The communication between the RTL9211DS and the RTS5261 is handled via firmware in this case.

The SD Express Standard

The SDA's SD Express standard aims to facilitate the manufacturing of removable cards in the legacy SD and microSD form-factors while retaining basic backwards compatibility. Towards this, the SD Express cards support both the PCIe / NVMe interface, as well as the UHS-I interface.

The pin layout of the SD Express cards are similar to that of UHS-II cards. Depending on the host capabilities, the card controller can switch between legacy SD UHS-I or NVMe modes. The card controller recognizes the host's capabilities based on the supply voltage - PCIe / NVMe requires a 1.8V supply and uses both rows of pins on the card, while the legacy mode is the default and uses only the top row of pins for data transfer. It must be noted that SD 8.0 introduced in 2020 allows for a third row of pins for supporting a second PCIe lane. Due to form-factor limitations, microSD Express doesn't support the two-lane feature.

While initialization may be in either mode, accessing through PCIe enables the card to present itself as a standard NVMe device to the host. This is especially relevant to our evaluation scheme, as S.M.A.R.T access via NVMe allows for temperature tracking and card health monitoring among other features.

Testbed Setup and Evaluation Methodology

Direct-attached storage devices (including SD Express cards) are evaluated using the Quartz Canyon NUC (essentially, the Xeon / ECC version of the Ghost Canyon NUC) configured with 2x 16GB DDR4-2667 ECC SODIMMs and a PCIe 3.0 x4 NVMe SSD - the IM2P33E8 1TB from ADATA.

The most attractive aspect of the Quartz Canyon NUC is the presence of two PCIe slots (electrically, x16 and x4) for add-in cards. In the absence of a discrete GPU - for which there is no need in a DAS testbed - both slots are available. In fact, we also added a spare SanDisk Extreme PRO M.2 NVMe SSD to the CPU direct-attached M.2 22110 slot in the baseboard in order to avoid DMI bottlenecks when evaluating Thunderbolt 3 devices. This still allows for two add-in cards operating at x8 (x16 electrical) and x4 (x4 electrical). Since the Quartz Canyon NUC doesn't have a native USB 3.2 Gen 2x2 port, Silverstone's SST-ECU06 add-in card was installed in the x4 slot. All non-Thunderbolt devices are tested using the Type-C port enabled by the SST-ECU06.

The specifications of the testbed are summarized in the table below:

The testbed hardware is only one segment of the evaluation. Over the last few years, the typical direct-attached storage workloads for memory cards have also evolved. High bit-rate 4K videos at 60fps have become quite common, and 8K videos are starting to make an appearance. Game install sizes have also grown steadily even in portable game consoles, thanks to high resolution textures and artwork. Keeping these in mind, our evaluation scheme for memory cards involves multiple workloads which are described in detail in the corresponding sections.

• Out-of-the-box performance using CrystalDiskMark and fio Sequential Access workloads
• Extended usage simulation using custom robocopy workloads and real-world access traces using PCMark 10's storage benchmark
• Long-term usage effects evaluation using CrystalDiskMark and fio Sequential Access workloads
• Performance restoration testing

The SM2708 reference design was evaluated in two modes - one using the Realtek card reader in SD Express mode, and another using the Lexar Professional Workflow SR2 SDHC / SDXC UHS-II USB 3.0 Reader (used in our standard evaluation of SD / uSD cards) in UHS-I mode. The remaining sections in this review will cover the performance of the card in both these modes along with some observations on power consumptions and thermals.

Out-of-the-Box Performance Evaluation

Silicon Motion claims read speeds of up to 895 MBps and write speeds in excess of 700 MBps in their marketing material for the SM2708. However, these speeds are heavily dependent on the NAND flash in the card. For the sampled reference design, Silicon Motion mentioned that they were seeing 895 MBps reads and around 420 MBps writes in terms of peak performance. Real-world speeds are bound to be much lower, depending on the particulars of the access trace.

CrystalDiskMark [Fresh]

CrystalDiskMark serves as a quick check to ensure that the card can meet the performance claims of the manufacturer. The workloads were processed for the SM2708 card in both SD Express and UHS-I modes.

fio Sequential Workload [Fresh]
TOP: SM2708 SD 7.1 250GBSM2708 UHS-I 250GB	BOTTOM: SM2708 UHS-I 250GBSM2708 SD 7.1 250GB

The reference design has a SLC cache of around 5.5 GB up to which write speeds of up to 375 MBps can be sustained. Beyond that, we have a 75 MBps region, and further on, 30 MBps. Once the controller has been subject to this traffic, the reads start off around 200 MBps before moving higher and higher and ending up at around 700 MBps. On the other hand, the SLC caching effect is a bit more difficult to track in the UHS-I mode. Write speeds vary from as low as 10 MBps to peaks around 62 MBps. Reads are consistent at 72 MBps.

Simulating Extended Usage

The performance of memory cards tends to go down over time as wear and tear on the NAND takes its toll. In order to simulate long-term usage, we subject the card to heavy traffic - similar to what one might do with direct-attached storage devices such as external drives. This traffic is also monitored to estimate performance consistency and relative performance numbers. Thanks to the exposure of the SD Express card as a standard NVMe device, the internal temperature of the SD Express card is also monitored.

AnandTech DAS Suite

Usage scenarios for memory cards may involve transfer of large amounts of photos and videos. Other usage scenarios include the use of the unit as a download or install location for games in portable game consoles, and importing files directly from it into a multimedia editing program such as Adobe Photoshop (for quick edits). Some users may even opt to boot an OS off a memory card in single-board computers.

The AnandTech DAS Suite tackles the first use-case. The evaluation involves processing four different workloads:

• AV: Multimedia content with audio and video files totalling 24.03 GB over 1263 files in 109 sub-folders
• Home: Photos and document files totalling 18.86 GB over 7627 files in 382 sub-folders
• BR: Blu-ray folder structure totalling 23.09 GB over 111 files in 10 sub-folders
• ISOs: OS installation files (ISOs) totalling 28.61 GB over 4 files in one folder

Each data set is first placed in a 29GB RAM drive, and a robocopy command is issue to transfer it to the memory card (formatted in exFAT).

robocopy /NP /MIR /NFL /J /NDL /MT:32 $SRC_PATH $DEST_PATH

Upon completion of the transfer (write test), the contents from the unit are read back into the RAM drive (read test) after a 10 second idling interval. This process is repeated three times for each workload. Read and write speeds, as well as the time taken to complete each pass are recorded. Whenever possible, the temperature of the memory card is recorded during the idling intervals. Bandwidth for each data set is computed as the average of all three passes.

Full System Drive Benchmark Bandwidth (MBps)Full System Drive Benchmark Latency (us)Full System Drive Benchmark ScoreExpand All

Equivalent numbers for external flash drives can be found in this review featuring both NVMe and SATA SSDs behind a bridge chip. The numbers in the SD Express mode closely track the ADATA SC680 960GB sample - which happens to be a SATA SSD using the Silicon Motion SM2259XT DRAM-less controller but with more flash packages. Given that the reference design sampled to us only uses two packages and has 25% of the capacity (not much parallelism to exploit), it is only the NVMe interface / SD Express operation that allows the memory caard to reach the SC680's performance level.

Long-Term Performance Evaluation

A few of the benchmarks are repeated on the memory card after subjecting it to the extended usage simulation.

Sequential Access - fio Workload

Re-processing the fio workload in the used mode gives an idea of long-term performance consistency (whether there is appreciable degradation in performance as the amount of pre-existing data increases and / or the card is subject to wear and tear in terms of amount and type of NAND writes).

CrystalDiskMark [Refreshed] Benchmarks
TOP: SM2708 SD 7.1 250GBSM2708 UHS-I 250GB	BOTTOM: SM2708 UHS-I 250GBSM2708 SD 7.1 250GB

Upon formatting in SD Express mode, the performance gets restored to around 878 MBps / 415 MBps). However, formatting in UHS-I mode has no effect, as the refreshed CrystalDiskMark numbers are very similar to the numbers seen in the used case.

Miscellaneous Aspects and Concluding Remarks

The appearance of PCIe / NVMe in the memory cards space enables us to bring over some of our direct-attached storage evaluation metrics. Two important aspects are power consumption and thermal characteristics.

Power Consumption

Bus-powered devices can configure themselves to operate within the power delivery constraints of the host port. While Thunderbolt ports are guaranteed to supply up to 15W for client devices, USB 2.0 ports are guaranteed to deliver only 2.5W (500mA @ 5V). In this context, it is interesting to have a fine-grained look at the power consumption profile of the memory card / card reader combination. Using the Plugable USBC-TKEY, the bus power consumption of the drives was tracked while processing the complete test suite. The graphs below plot the instantaneous bus power consumption against time, while singling out the maximum and minimum power consumption numbers.

Power Consumption Profile [Card + Reader]
TOP: SM2708 SD 7.1 250GBSM2708 UHS-I 250GB	BOTTOM: SM2708 UHS-I 250GBSM2708 UHS-I 250GBSM2708 SD 7.1 250GB

The Realtek card reader and the SM2708 reference design combination has a peak power consumption of 4.58W, with the average being around 3.25W. On the other hand, in UHS-I mode, the Lexar card reader and the reference design operate at less than 1.67W, with the average being around 1.25W. The extra performance in SD Express mode comes at the cost of power consumption, but that is no surprise.

Thermal Profile

The performance consistency evaluation of the reference design in SD Express mode revealed the internal temperature of the SD card to be around 100C when subject to heavy stress. In order to rule out any errors in the S.M.A.R.T polling, we used a FLIR One Pro thermal imager to capture the thermal profile of the card while being subjected to the performance consistency test.

To our dismay, we found the surface temperature to be as high as 96C - something entirely unsafe for handling with bare hands. It does look like safety standards allow for such surface temperatures, provided the device is under 50mm x 50mm in size. Unfortunately, it is a sad reality that the SD card and microSD card form-factors do not allow for the thermal design necessary to sustain extended operation with NVMe-class speeds. In any case, Silicon Motion's customers can tweak the firmware's configurable parameters to achieve a different balance between performance and temperature. The SM2708 might also held back a bit by its 28nm fabrication node. Once SD Express achieves scale of deployment, it is likely that Silicon Motion can choose more advanced nodes to drive down the power consumption and stress temperatures.

Final Words

Reviewing the Silicon Motion SM2708 reference design has helped us greatly in setting the consumer expectations for SD Express using NVMe right. The advertised numbers can only be achieved in burst mode, and sustained writes may be affected by a combination of SLC cache sizes and thermals. Performance similar to that of a high-end SATA SSD or a PCIe 3.0 x2 NVMe SSD can be expected, though the nature and number of flash pieces inside the card can heavily influence the performance.

Silicon Motion has a couple of design wins ready to ship, with ADATA and Lexar in the final stages of bringing their SD Express products into the market.

Overall, consumers have been waiting for SD Express for quite some time. While CFexpress has taken the early lead with camera / camcorder makers, SD cards have a much wider range of markets to service and target. We are bullish on the prospects for SD Express in verticals such as portable game consoles and SBCs, but it might take a few years for mid-range cameras to start supporting it.

In the meanwhile, products such as the Silicon Motion SM2708 and Realtek RTL9211DS / RTS5261 can prime the market for the new standard on the PC side. A lot is dependent on the vendors not charging too much of a premium for SD Express over the standard UHS-II cards / readers.