Interface & Protocols: SATA, PCIe, UFS, eMMC, and New Standards

Efficient movement of data between storage media and processors hinges on underlying interfaces and protocols. As storage technologies—from spinning disks to NAND flash and beyond—have advanced, the interconnects that link them to CPUs and SoCs have had to evolve in tandem. In this article, we’ll explore the major interfaces—SATA, PCIe, UFS, eMMC—and examine emerging standards that promise even higher bandwidth, lower latency, and smarter data handling.

1. Why Interfaces Matter

Before delving into specifics, it’s useful to understand why the choice of interface/protocol matters at all:

• Bandwidth Ceiling: An interface’s raw throughput limit directly caps sequential read/write speeds. For example, a SATA III link tops out at 6 Gb/s (≈600 MB/s after overhead), whereas a PCIe Gen4 ×4 lane can deliver nearly 8 GB/s.
• Latency & Overhead: Legacy protocols (e.g., ATA/AHCI over SATA) carry extra command‐processing overhead, increasing per‐I/O latency. Newer protocols like NVMe are streamlined for low‐latency, high‐parallelism flash media.
• Power Efficiency: Mobile‐focused protocols (e.g., UFS, eMMC) often include deep‐sleep states and optimized power‐management features. In smartphones and IoT devices, quiescent current can be as important as raw throughput.
• Software Compatibility & Ecosystem Support: Some embedded OSes and bootloaders have native support for eMMC or UFS. Desktop/server platforms universally support SATA and PCIe/NVMe, while specialized “storage‐class memory” or computational storage may require updated drivers or firmware.

By matching storage media to the right interface—taking into account workload (random vs. sequential), latency sensitivity, power budgets, and cost constraints—systems can achieve maximum practical performance.

2. SATA (Serial ATA)

2.1 Overview & Evolution

Origins: Introduced in 2003 as a successor to parallel ATA (PATA), SATA employed a serial, point‐to‐point link rather than PATA’s ribbon cables.

Versions:

• SATA I (1.5 Gb/s) – debut generation, ≈150 MB/s effective.
• SATA II (3 Gb/s) – doubled link speed to ≈300 MB/s.
• SATA III (6 Gb/s) – current mainstream, ≈600 MB/s after 8b/10b encoding overhead.

Each revision was backward‐compatible, so a SATA III SSD could still function in a SATA II port (albeit limited to ≈300 MB/s).

2.2 Command Layer: AHCI vs. SATA

AHCI (Advanced Host Controller Interface)

• Defines how the OS communicates with a SATA controller.
• Supports a single submission queue with a depth of 32 commands.
• Designed originally for mechanical drives, AHCI adds overhead when interacting with NAND media—especially as SSD performance began to approach SATA’s raw link speed.

SATA’s Native Command Queuing (NCQ)

• Permits reordering of up to 32 commands to optimize head movement in HDDs.
• When SSDs entered the market, NCQ still provided a way to batch multiple I/O commands and reduce latency—but the single‐queue architecture eventually became a bottleneck as SSD media got faster.

2.3 Performance Characteristics

Feature	Typical Modern 2.5″ SATA SSD	Typical 3.5″ 7,200 RPM HDD
Sequential Read	≈550 MB/s	≈150 MB/s
Sequential Write	≈520 MB/s	≈140 MB/s
Random 4 KB Read Latency	≈0.08 ms (80 µs)	≈12 ms
Random 4 KB Write Latency	≈0.3 ms	≈12 ms
Queue Depth	NCQ up to 32	NCQ up to 32
Power (Active)	2 W–3 W	6 W–9 W

Saturation Point: By around 2013, high‐end SATA SSDs like the Samsung 840 Pro were saturating the 6 Gb/s link, delivering ≈550 MB/s reads and ≈520 MB/s writes.

Latency: Even with AHCI overhead, SSDs on a SATA link achieved ≈80 µs random read latency—over 100× faster than an HDD’s ≈12 ms.

2.4 Use Cases & Limitations

• Use Cases
  • General‐purpose laptops/desktops (boot drives, midrange SSDs).
  • Upgrading legacy systems: Many older motherboards expose only SATA ports.
  • Cost‐sensitive storage expansions: SATA HDD arrays remain common for bulk data, backups, and archival.
• Limitations
  • Bandwidth ceiling at ≈600 MB/s means new NAND generations (3D TLC/QLC) easily exceed the link.
  • AHCI’s single queue and legacy layers add some latency overhead.
  • Power efficiency is reasonable for desktops but suboptimal for ultra-thin laptops or mobile devices—that’s where UFS and eMMC excel.

3. PCIe & NVMe

3.1 PCIe Fundamentals

• Lane Structure: PCI Express lanes pair differential signal pairs for transmit/receive. Each lane’s raw bit rate depends on the generation:
  • Gen3: 8 GT/s → ≈1 GB/s per lane after 128b/130b encoding overhead.
  • Gen4: 16 GT/s → ≈2 GB/s per lane.
  • Gen5: 32 GT/s → ≈4 GB/s per lane.
  • Gen6: 64 GT/s (PAM4) → ≈8 GB/s per lane.
• ×4 vs. ×8 Configurations: Most M.2 NVMe drives use ×4 lanes. A Gen4 ×4 link thus offers ≈8 GB/s raw, while a Gen5 ×4 link hits ≈15.8 GB/s. Enterprise U.2 or EDSFF cards may use ×8 or ×16 lanes to push beyond 16 GB/s in Gen5.

3.2 NVMe (Non-Volatile Memory Express)

Protocol Layer

• Developed specifically for non-volatile media, NVMe lives above the PCIe fabric.
• Supports up to 64 K submission queues, each up to 64 K commands deep—enabling massive parallelism across multi-core CPUs.
• Reduces CPU overhead by streamlining command submission compared to AHCI/Ata commands.

Latency Advantages

• Under light load (queue depth = 1), high-end NVMe SSDs can hit ≈20 µs–30 µs random read latency.
• At queue depths of 32–64, NVMe can maintain tens to hundreds of thousands of IOPS, far beyond what SATA SSDs could manage.

3.3 PCIe Generations Impact

Gen	Lane Speed	×4 Bandwidth	Common Years	Target SSD Class
Gen3	8 GT/s	≈3.94 GB/s	2012–2019	Early NVMe SSDs (3 GB/s peak)
Gen4	16 GT/s	≈7.88 GB/s	2019–2022	Gaming/Workstation SSDs (7 GB/s peak)
Gen5	32 GT/s	≈15.75 GB/s	2022–Present	High-end Desktop/Enterprise (12–14 GB/s)
Gen6	64 GT/s (PAM4)	≈31.5 GB/s	Expected 2025+	Next-Gen Enterprise & AI

Impact on SSDs

• Gen3 ×4: Typically delivered 3,000–3,500 MB/s reads, 2,500–3,000 MB/s writes.
• Gen4 ×4: Pushed sequential reads to 7,000 MB/s and writes to 5,000–6,000 MB/s.
• Gen5 ×4: Further improvements, with flagship drives now topping ≈12,000 MB/s sequential. Random IOPS scale from ≈600 K in Gen3 to ≈1 M–1.2 M in Gen5 under QD32.

3.4 Use Cases & Ecosystem

• Desktop & Workstation
  • Gamers and content creators benefit from ultra-fast texture paging and scratch disk performance.
  • Niche overclockers and enthusiasts push memory timings, but storage bandwidth becomes critical for high-resolution asset streaming.
• Enterprise & Data Center
  • Servers use U.2 (2.5″) or EDSFF E3/E1.S form factors to take advantage of more lanes (×8 or ×16), redundant power, and hot-swap capability.
  • NVMe over Fabrics (RoCE, iWARP, Fibre Channel) extends NVMe performance across the network—ideal for hyperconverged infrastructure.
• Emerging Trends
  • CXL (Compute Express Link): Although technically a separate protocol, CXL 2.0/3.0 uses the same physical layer as PCIe Gen5/Gen6 to enable coherent memory sharing between CPUs, GPUs, and accelerators. Future “CXL SSDs” may combine NVMe functions with CXL.mem for pooled, byte-addressable storage.
  • Computational Storage: Offloads data processing (compression, encryption, database scans) onto the SSD itself, reducing CPU usage and data movement.

4. UFS (Universal Flash Storage)

4.1 Origins & Architecture

Designed for Mobile: UFS was developed by the JEDEC Solid State Technology Association to replace both eMMC (embedded MultiMediaCard) and various slower eMMC alternatives in smartphones, tablets, and embedded systems.

Versions & Performance:

• UFS 2.0 (2013): Introduced a full-duplex, point-to-point architecture. Each lane runs at up to 5 Gb/s (≈500 MB/s after 8b/10b encoding). Since UFS 2.0 supports two lanes (fast + slow), peak sequential read/write was ≈1,000 MB/s.
• UFS 2.1 (2016): Added features like Write Booster (pseud-SLC caching), Deep Sleep mode for ultra-low power, and improved background garbage collection.
• UFS 3.0 (2018): Doubled lane speed to 11.6 Gb/s (≈1.15 GB/s per lane after 128b/130b), giving ≈2.3 GB/s peak. Lower power per bit than UFS 2.1.
• UFS 3.1 (2020): Introduced Host Performance Booster (HMB) to allocate host DRAM for caching, Deep Sleep standby (hardware control), and Write Booster enhancements.
• UFS 4.0 (2022): Lane speed jumped to 23.2 Gb/s (≈2.9 GB/s after encoding) per lane—so ≈5.8 GB/s sequential throughput. Compared to UFS 3.1, UFS 4.0 cut latency by ≈30% and reduced active power by up to 50% per bit.

4.2 Protocol Advantages

• Full-Duplex, Low Overhead:
  • Unlike eMMC’s half-duplex, shared bus, UFS implements a serial, point-to-point link with dedicated uplink/downlink lanes—enabling simultaneous read and write operations.
  • Logical layers (UFS Transport Layer, UFS Link Layer) streamline command sets, borrowing from SCSI’s UFSHCI (UFS Host Controller Interface).
• Power Management:
  • Deep Sleep modes (UFS 2.1 and later) allow the device to be virtually off, waking in ≈100 µs—critical for mobile battery life.
  • Advanced throttling and dynamic lane shutdown minimize energy usage during idle.

4.3 Performance Metrics

UFS Version	Lanes	Per-Lane Speed	Total Bandwidth	Random 4 KB Read Latency	Random 4 KB Write Latency
2.0	2 (1 fast + 1 slow)	5 Gb/s (≈500 MB/s)	≈1,000 MB/s	≈200 µs	≈800 µs
2.1	2 (1 fast + 1 slow)	5 Gb/s (≈500 MB/s)	≈1,000 MB/s	≈180 µs	≈700 µs
3.0	2 (High-Speed Lanes)	11.6 Gb/s (≈1.15 GB/s)	≈2.3 GB/s	≈120 µs	≈400 µs
3.1	2 (High-Speed Lanes)	11.6 Gb/s (≈1.15 GB/s)	≈2.3 GB/s	≈100 µs	≈350 µs
4.0	2 (High-Speed Lanes)	23.2 Gb/s (≈2.9 GB/s)	≈5.8 GB/s	≈80 µs	≈250 µs

Sustained Throughput: High-end UFS 4.0 eUFS modules (embedded UFS) can exceed 4,000 MB/s sequential reads and 2,000 MB/s sequential writes in flagship smartphones (e.g., 2023–2024 Android flagship SoCs).

Random IOPS: UFS 3.1 and 4.0 achieve ≈150 K–200 K random read IOPS (4 KB) at queue depths as low as 1–4—nearly on par with SATA SSDs in light workload scenarios.

4.4 Use Cases & Ecosystem

• Smartphones & Tablets
  • Almost every flagship Android smartphone from 2020 onward uses UFS 3.1 or UFS 4.0 for onboard storage.
  • High sustained read speeds accelerate application launches and camera buffer writes for burst photography.
• Embedded Systems & Automotive
  • UFS’s power management and small footprint make it a fit for infotainment systems, industrial controllers, and in-vehicle infotainment (IVI) where durability and low standby power are crucial.
• Comparison vs. eMMC
  • Throughput: eMMC 5.1 peaks at ≈400 MB/s, whereas UFS 2.1 already hit ≈1 GB/s.
  • Architectural Overhead: eMMC’s shared bus is half-duplex, so read or write can’t occur simultaneously, increasing I/O latency. UFS’s full-duplex lanes enable concurrent operations.
  • Security & Boot: Both support secure boot and boot partitions, but UFS often integrates more robust encryption offload and faster boot times.

5. eMMC (Embedded MultiMediaCard)

5.1 Overview & Versions

Origins: Derived from removable MMC cards popular in digital cameras and early embedded platforms, eMMC emerged as a “chip-on-board” soldered NAND solution.

Common Versions:

• eMMC 4.5 (2012): The first ubiquitous standard for low-end smartphones and simple embedded systems. Offered ≈200 MB/s reads, ≈100 MB/s writes.
• eMMC 5.0 (2013): Boosted to ≈250 MB/s reads and ≈125 MB/s writes with HS200 mode (200 MHz DDR).
• eMMC 5.1 (2015): Added HS400 (400 MHz DDR) mode—peak ≈400 MB/s reads, ≈200 MB/s writes, improved command queue (CQ) up to depth 32.
• eMMC 5.1x (Beyond 2016): Minor tweaks in power states and hardware reliability, but effectively saturated at ≈400 MB/s.

5.2 Protocol & Architecture

Parallel Bus & Half-Duplex

• eMMC uses an 8-bit parallel data bus running at up to 400 MHz DDR in HS400 mode.
• Unlike UFS’s serial, full-duplex lanes, eMMC can only read or write at any given moment—cannot do both concurrently.

Controller Functions

• On-die ECC: Each eMMC package contains a controller that handles wear leveling, bad-block management, and ECC—abstracting it from the host SoC.
• Boot Partitions: eMMC reserves up to two boot partitions (each up to ≈32 MB) that can be mapped as “hidden” boot areas for faster OS loading.

5.3 Performance & Use Cases

Performance Characteristics

• Sequential read/write ceilings at ≈350 MB/s–400 MB/s and ≈150 MB/s–200 MB/s, respectively, in HS400.
• Random 4 KB reads often hover around 8 K–12 K IOPS—sufficient for lightweight UI and OS tasks but far below UFS or SATA SSD levels.

Use Cases

• Entry-Level Smartphones & Tablets: Cost-sensitive devices still rely on eMMC 5.1 for moderate storage performance.
• IoT & Embedded Controllers: Single-board computers (e.g., early Raspberry Pi models), industrial controllers, and POS terminals often use eMMC due to its low cost and simple integration.
• Automotive & Appliances: Simple infotainment screens, telematics control units (TCUs), and digital instrument clusters that don’t require multi-GB/s throughput.

5.4 Decline & Transition

As UFS became affordable and more energy-efficient, many OEMs transitioned from eMMC to UFS in midrange devices circa 2018–2020. However, eMMC remains entrenched in ultra-low-cost and deeply embedded sectors where ≤400 MB/s is sufficient and power envelopes are tight.

6. New & Emerging Standards

6.1 PCIe 6.0 & Beyond

Bandwidth Leap

• PCIe 6.0 doubles per-lane bandwidth versus Gen5, moving from 32 GT/s to 64 GT/s with PAM4 signaling.
• A ×4 link in Gen6 could deliver ≈31.5 GB/s—blurring lines between PCIe SSDs and memory fabrics.

Encoding Overhead: PAM4 requires more sophisticated equalization and error detection/correction to maintain signal integrity at high frequencies.

Implications for NVMe SSDs:

• Consumer M.2 form factors will likely adopt Gen6 ×4 drives in late 2025 or early 2026, pushing sequential reads/writes beyond 16,000 MB/s.
• Enterprise EDSFF E3/E1.S cards may leverage Gen6 ×8 or ×16 for multi-GB/s sustained throughput—ideal for AI/ML training datasets.

6.2 CXL (Compute Express Link)

Protocol Overview

• CXL builds on the PCIe physical layer (Gen5 and soon Gen6) but adds three sub-protocols:
  • CXL.io – PCIe 5.0-compatible, handles configuration and standard I/O.
  • CXL.mem – Enables cache-coherent memory accesses to attached memory devices (DRAM, persistent memory).
  • CXL.cache – Allows a device (e.g., accelerator) to cache host memory coherently.
• CXL 2.0 (2021) introduced memory pooling, allowing multiple hosts to share a set of CXL memory modules.
• CXL 3.0 (2023) adds switching and fabric topologies, so data centers can build disaggregated memory/storage pools across racks.

Use Cases:

• Memory Disaggregation: Instead of fixed DRAM slots in each server, a pool of DRAM/PMem modules can be allocated dynamically—maximizing utilization.
• Persistent Memory Expansion: CXL.mem devices can offer large-capacity, byte-addressable persistent memory (e.g., DDR5 PMem DIMMs)—benefiting in-memory databases, virtualization, and caching layers.

6.3 NVMe Over Fabrics (NVMe-oF)

Beyond Local PCIe:

• NVMe-oF extends NVMe’s performance over network fabrics like RDMA (RoCE, iWARP) and Fibre Channel (FC).
• Reduces CPU overhead by offloading NVMe command processing to smart NICs or HBAs.
• Achieves consistent sub-100 µs latency across the network—critical for scale-out storage and hyperconverged clusters.

Key Technologies:

• RoCE (RDMA over Converged Ethernet) – runs NVMe commands over Ethernet with RDMA semantics.
• NVMe/FC – NVMe commands tunneled over Fibre Channel, compatible with existing FC SANs.

6.4 Zoned Namespaces (ZNS)

Principle:

• ZNS partitions an SSD into multiple sequential write “zones.” The host is responsible for writing data sequentially within each zone, eliminating internal FTL (Flash Translation Layer) overheads associated with random writes and garbage collection.

Benefits:

• Reduced Write Amplification: By avoiding random writes, write amplification drops significantly—improving endurance, especially for QLC/PLC NAND.
• Predictable Latency: With host-managed zone pointers, latency spikes due to background GC are minimized.

Adoption:

• Hyperscale data centers (e.g., key-value workloads like RocksDB) and large-scale logging (Kafka) benefit from ZNS’s write-optimized design.
• Linux’s lightnvm subsystem and Windows’s ZNS API support host-managed SSDs as of 2022–2023.

6.5 Open-Channel SSDs

Host-Managed Flash:

• A step beyond ZNS: the host controls every aspect of LBA-to-physical-block mapping. The SSD effectively becomes raw flash (with no internal FTL), placing wear leveling, garbage collection, and error management in the host’s domain.

Trade-Offs:

• Pros: Maximum performance tuning—ideal for specialized workloads with predictable write patterns.
• Cons: Complex software stack required (specialized drivers, patch kernels). Limited to large hyperscalers or niche applications.

6.6 UFS 4.0 & Beyond

UFS 4.0 Recap:

• Doubles lane speed to 23.2 Gb/s (≈2.9 GB/s after encoding) per lane, total ≈5.8 GB/s.
• Reduces latency by ≈30% compared to UFS 3.1, and active power per bit is roughly halved.

UFS 4.1 (2024–2025):

• Introduced new security features (Secure Write Protection, improved encryption offload), faster startup times, and further power-state optimizations.
• Peak reads ≈6 GB/s, writes ≈3 GB/s.

Anticipated UFS 5.0:

• Expected around 2026, likely leveraging PCIe Gen5's physical layers to push beyond 10 GB/s on mobile devices. Will further close the gap between embedded and desktop/performance storage.

7. Comparative Overview

Interface	Max Link BW	Typical Media	Primary Use	Half/Full Duplex	Latency (4 KB Read)
SATA III	6 Gb/s (600 MB/s)	HDD, SSD	Desktops, Laptops, Legacy Servers	Half-Duplex	≈80 µs (SSD)
PCIe Gen4 ×4	16 Gb/s per lane → ≈7.8 GB/s total	NVMe SSD, GPUs, Add-in Cards	High-Performance Workstations, Servers	Full-Duplex	≈18 µs (NVMe SSD)
UFS 3.1	11.6 Gb/s per lane → ≈2.3 GB/s total	Smartphones, Tablets, Embedded	Mobile Boot Drives, Embedded Storage	Full-Duplex	≈100 µs
eMMC 5.1	400 MB/s (HS400)	Entry-Level Mobile, IoT, SBCs	Low-Cost Embedded Systems	Half-Duplex	≈400 µs
PCIe Gen5 ×4	≈31.5 GB/s	Next-Gen NVMe SSDs, CXL Devices	Ultra-High-Performance Systems	Full-Duplex	≈15 µs

---

8. Best Practices & Future Outlook

1. Match Workload to Interface
   • HTPC/Everyday Desktop: A SATA III SSD still delivers sub-0.1 ms boot and application loads at a low cost.
   • Enthusiast Rig/Content Creation: PCIe Gen4/Gen5 NVMe SSDs yield the fastest scratch-disk performance for 8K video editing and high-resolution rendering.
   • Flagship Smartphone: UFS 3.1 or UFS 4.0 modules provide instantaneous app launches, snappy camera buffer writes, and minimal power usage.
   • Embedded IoT & SBCs: eMMC 5.1 remains widespread where ≤400 MB/s is sufficient, power budgets are tight, and cost matters most.
2. Plan for PCIe Gen6 & CXL
   • Enterprises evaluating new server platforms in late 2025+ should look for native PCIe Gen6 and CXL 3.0 support. That will enable future-proof memory disaggregation, persistent memory expansions, and extremely low-latency NVMe oF fabrics.
   • Hyperscalers adopting ZNS or Open-Channel SSDs in Q4 2024–Q1 2025 gain endurance and predictable performance in write-intensive workloads.
3. Firmware & Driver Updates
   • For NVMe, keep firmware current: new controller microcodes often implement better thermal management and drive life optimizations.
   • UFS OEMs occasionally release updated drivers for better power management or bug fixes—ensure your smartphone or embedded board uses the latest vendor-supplied UFS driver stacks.
4. Security Considerations
   • Modern interfaces like UFS 4.0 and eMMC 5.1 support built-in AES encryption and secure boot regions. Use device-level encryption wherever possible to protect data at rest.
   • Enterprise NVMe drives often include TCG Opal or FIPS-certified SSD controllers; ensure compliance if deploying in regulated environments.

Interfaces and protocols are the unsung heroes beneath every storage solution. While NAND flash, DRAM, and emerging memory media attract headlines, it’s the link between storage device and processor that ultimately determines real-world performance, latency, and power consumption.

SATA III gave us the first mainstream flash-accelerated desktops and laptops, replacing mechanical HDD bottlenecks with ≈0.1 ms latency for under $100.

PCIe + NVMe unlocked multi-GB/s throughput and microsecond-level random I/O, powering everything from high-end gaming rigs to hyperscale data centers.

UFS enabled flagship smartphones to approach SSD-class performance in a tiny, power-sipping package—shrinking OS boot times, speeding app launches, and enabling 8K video recording.

eMMC still maintains a foothold in cost-sensitive embedded and IoT markets, offering reliable 200–400 MB/s storage at minimal BOM cost.

Looking ahead, PCIe Gen6, CXL, ZNS, and computational storage promise to further collapse the divide between “memory” and “storage,” enabling coherent, byte-addressable data pools and in-situ data processing. Whether you’re selecting a boot drive for your next desktop build, architecting a flash tier in a cloud cluster, or designing the next generation of flagship smartphones, understanding these interfaces and protocols will help you choose—and optimize—the right storage solution for every use case.