What Is NVMe? The Complete Guide to Non-Volatile Memory Express

Storage protocols designed for mechanical drives have been constraining flash performance for over a decade. While SSDs deliver microsecond-capable hardware, legacy protocols like SATA and SAS add hundreds of microseconds of unnecessary latency through their single-queue architectures and protocol translation layers—based on industry benchmarks and real-world deployments.

NVMe (Non-Volatile Memory Express) is a storage protocol designed specifically for solid-state drives that connects directly through the PCIe interface, eliminating the bottlenecks of disk-era protocols. Instead of funneling commands through a single queue like SATA, NVMe enables up to 64,000 queues with 64,000 commands each—fundamentally changing how storage communicates with modern multi-core processors.

But what most discussions miss is that simply adding NVMe drives isn't enough if your system still translates between protocols, converting NVMe to SCSI and back again at various points in the data path.

This guide examines NVMe's architecture, quantifies its real-world performance advantages, and explains why end-to-end NVMe implementation matters.

How NVMe Revolutionized Storage Architecture

For two decades, storage protocols were designed around mechanical limitations. SATA and SAS assumed storage devices needed time to physically seek data, building in command overhead that made sense when disk platters had to rotate into position. These protocols funnel all commands through a single queue—adequate for mechanical seeks but catastrophic for flash memory capable of microsecond responses.

The protocol mismatch becomes clear in the numbers. SAS supports up to 256 commands (per the SAS-3 specification)in its single queue, while enterprise SSDs handle thousands of simultaneous operations. These legacy protocols require multiple translation layers: applications send NVMe commands that get translated to SCSI, then to SATA or SAS, then potentially back to NVMe at the drive level. Each translation adds 50-200 microseconds of latency.

Why Flash Needed Its Own Protocol

NVMe emerged in 2011 to eliminate translation penalties. Rather than retrofitting disk protocols, the NVM Express consortium designed a protocol assuming no mechanical components. NVMe streamlines the command set, eliminating parsing overhead while maintaining full functionality.

The protocol connects storage directly to CPUs via PCIe lanes, the same high-speed interface used for graphics cards. This positions storage as a peer to other high-performance components rather than relegating it behind translation layers. With PCIe Gen 4 delivering 64GB/s of bandwidth, NVMe allows flash to operate without constraints.

How NVMe Works: Architecture and Components

NVMe's architecture fundamentally rethinks storage communication. Instead of traditional host bus adapters, NVMe storage appears to the CPU as memory-mapped I/O, allowing direct access without kernel overhead for critical operations.

Queue Architecture and CPU Optimization

Modern processors contain dozens of cores, yet legacy storage protocols funnel them all through a single I/O queue. NVMe assigns dedicated queue pairs to each CPU core, eliminating lock contention and enabling true parallel processing.

When an application needs data, it places commands in submission queues via simple memory writes—no system calls required. The NVMe controller processes commands independently and places results in completion queues. This asynchronous model means CPUs spend virtually no cycles waiting for storage.

PCIe Lanes and Bandwidth

NVMe devices connect via PCIe lanes, with each lane providing bidirectional bandwidth. A typical NVMe SSD uses four PCIe lanes, delivering up to 8GB/s with PCIe Gen 4. Enterprise arrays aggregate multiple devices for even higher throughput.

But bandwidth alone doesn't determine performance. Latency—the time between request and response—often matters more for transactional workloads. NVMe's direct PCIe connection eliminates multiple bus transitions and protocol conversions that plague SATA implementations.

NVMe Performance Benefits: Real Numbers, Not Marketing

Storage industry marketing often makes vague claims like "blazing fast" or "ultra-responsive." However, NVMe delivers real advantages.

According to industry testing and vendor specifications, raw NAND flash reads take approximately 100 microseconds. However, SATA SSDs typically deliver total latencies of 100-200 microseconds, while NVMe SSDs achieve 20-100 microseconds—demonstrating how protocol overhead can equal or exceed the actual media access time.

IOPS and Real-world Impact

A single NVMe device can deliver over 1 million IOPS for 4KB random reads—performance requiring dozens of SATA SSDs. Oracle databases on end-to-end NVMe show:

• More transactions per second
• Reduction in query response time
• Fewer storage-related wait events

Power Efficiency

NVMe's efficiency compounds its performance benefits. By eliminating protocol overhead:

• SATA SSD: ~10,000 IOPS per watt
• NVMe SSD: ~50,000 IOPS per watt

NVMe-oF: Extending NVMe beyond Direct Attachment

NVMe over Fabrics extends NVMe's benefits across data centers, enabling shared storage without sacrificing latency advantages. But implementation choices dramatically impact performance.

NVMe over Fibre Channel (FC-NVMe)

FC-NVMe leverages existing SAN infrastructure, making it attractive for enterprises with Fibre Channel investments. It requires Gen 5 (16Gb) or Gen 6 (32Gb) switches that support NVMe forwarding—older switches claiming "NVMe support" often perform protocol translation, reintroducing overhead.

NVMe over RoCE

RoCE promises the lowest network latency through kernel bypass—RDMA operations complete in around a microsecond. But RoCE requires lossless Ethernet with Priority Flow Control across every switch and adapter. One misconfigured port can cause a performance collapse. The reality is that many "RoCE" deployments actually run iWARP because true RoCE proves too fragile. When properly implemented, RoCE can deliver 160-180 microsecond storage latency.

NVMe over TCP

NVMe/TCP runs over standard Ethernet without special hardware. Critics dismiss it as "slow," but modern implementations can achieve 200-250 microsecond latency—faster than SATA SSDs despite crossing the network.

The key advantage: simplicity. NVMe/TCP works with existing switches, standard NICs, and cloud provider networks.

Implementing NVMe in Production

Simply installing NVMe drives rarely delivers expected benefits. The entire storage stack must support end-to-end NVMe operations.

The Protocol Translation Trap

Many organizations buy NVMe SSDs for existing arrays and expect transformation. The drives communicate via NVMe, but the controller translates everything to SCSI for compatibility. This translation adds microseconds, negating NVMe's advantages.

OS and Migration Requirements

NVMe requires a modern operating system to support. Each requires specific configurations—interrupt affinity, multipath modules, and queue depth adjustments.

For a successful migration:

1. Start with non-critical workloads for validation
2. Implement latency monitoring at every layer
3. Prioritize latency-sensitive databases first
4. Verify end-to-end NVMe with tools like nvme-cli

NVMe for AI and Modern Workloads

Expensive GPUs often sit idle, waiting for data. NVMe changes that through GPU Direct Storage—enabling drives to transfer data directly to GPU memory.

For AI training, this means:

• Faster epoch training
• Faster checkpoint writing
• Increased GPU utilization 
• Freed up CPU for preprocessing

Databases benefit beyond raw speed. NVMe's predictable sub-200 microsecond latency eliminates query planning uncertainty. Optimizers make better decisions knowing data arrives quickly. Applications designed for slow storage behave differently when storage becomes predictable.

The Pure Storage End-to-end NVMe Advantage

While the industry debates adoption strategies, Pure Storage has deployed end-to-end NVMe across thousands of customer deployments, generating telemetry that reveals what actually works. The differentiator is eliminating every protocol translation between the application and NAND flash.

DirectFlash: Eliminating Hidden Overhead

Traditional NVMe SSDs contain redundant controllers and overprovisioning. Pure Storage® DirectFlash® modules expose raw NAND directly to the array's NVMe interface, delivering:

• More usable capacity
• Lower power consumption
• Predictable latency without garbage collection
• Global wear leveling across all flash

End-to-End NVMe Architecture

Purity software maintains NVMe from host to NAND while supporting legacy systems. For NVMe hosts, it provides direct namespace access. For legacy hosts, it translates once at the array edge—not internally.

Pure Storage FlashArray//X™ delivers consistent sub-200 microsecond latency by eliminating internal protocol conversions:

• Pure Storage arrays: 150μs average latency
• Traditional "NVMe" arrays with internal translation: 400-600μs
• The difference: elimination of protocol translation overhead

Non-disruptive Evolution

Pure Storage Evergreen architecture enables NVMe adoption without forklift upgrades. Controllers upgrade to NVMe-capable versions without data migration.

The Future of NVMe

NVMe's evolution extends beyond speed. The NVMe 2.0 specification introduces computational storage—processing within the storage device itself. Database filtering, compression, and AI inference happen where data lives, eliminating movement overhead.

Conclusion

NVMe represents the elimination of artificial bottlenecks constraining applications for decades. When implemented end-to-end without protocol translation, NVMe delivers 150-microsecond latency that transforms everything from database transactions to AI training.

The critical insights: Protocol translation destroys NVMe's advantages. NVMe-oF extends benefits across data centers, but implementation matters. Modern workloads require the predictable, low latency only end-to-end NVMe provides.

A Pure Storage end-to-end implementation, validated across thousands customers, proves 150-microsecond latency is an operational reality. Through DirectFlash modules, organizations achieve the performance NVMe promises. As storage evolves toward computational capabilities and memory speeds, Pure Storage Evergreen architecture ensures today's investments deliver tomorrow's innovations without disruption.