Optimizing an NVMe Host IP for FPGA Storage

Optimizing an NVMe Host IP for FPGA Storage

NVMe host optimization in an FPGA starts by treating latency and throughput as separate acceptance targets, not interchangeable measures. The workload determines which one is the hard constraint, but the design decision should account for both.

For bandwidth, attach interface generation, transfer direction, unit, and device context to each target. Keep sequential and random I/O on separate assessment tracks.

The short answer

For FPGA NVMe host optimization, target both latency and throughput, then let the workload set the primary constraint. Write down the requirement and evaluation method for each metric, and keep them fixed when comparing designs. Match datapath goals to the intended PCIe generation, transfer direction, unit, and device context; do not treat an SSD result as proof of host performance. Assess sequential and random I/O separately, and place write caching where the architecture assigns policy and state ownership. Consider SATA IP only for a SATA requirement, or custom development when a defined interface or behavior gap remains.

What should an FPGA NVMe host optimize: latency, throughput, or both?

For NVMe host optimization in an FPGA, optimize both latency and throughput, then let the workload determine which target leads. We recommend measuring them separately and judging the design against both stated requirements.

Electronic Design describes NVMe specification version 1.0b as giving the industry a standardized, low-latency, open-source host driver and command set that enables interface and solid-state memories to achieve a level of performance. Read Electronic Design’s explanation of SATA and NVMe. Treat that description as context, not as a substitute for targets defined for your FPGA storage design.

Start with the workload. Write down the latency and throughput requirements that matter to the design, and decide which one is the hard constraint. Use the other as a separate acceptance target rather than assuming that one metric represents both.

Optimization target Question to answer How to use it
Latency What latency requirement must the design meet? Set it as an explicit acceptance target.
Throughput What throughput requirement must the design meet? Track it as a separate acceptance target.
Both Does the selected design satisfy both requirements? Use both targets to make the final decision.

Our recommendation: optimize for both, with a clearly named primary constraint. Keep the workload definition, target values, and evaluation method fixed when comparing host-IP design choices, so the decision reflects the requirements you set rather than a change in the comparison procedure. If a design meets one target but misses the other, describe that result directly instead of calling it optimized overall.

How should PCIe bandwidth shape the host datapath target?

Set the host datapath target against the intended PCIe interface, not a generic headline number. EENews Europe reports that the CM7 Series enterprise NVMe SSDs nearly saturate PCIe 5.0 at 14 gigabytes/s read throughput.

Treat 14 gigabytes/s as a PCIe 5.0 read reference for the CM7 Series. The cited result does not establish a write-throughput target, a result for another SSD, or a minimum requirement for every FPGA host. Keep the scope attached whenever you reuse the number.

For a host specification, state the interface generation, transfer direction, throughput unit, and device context beside the target. Keep read and write targets separate: the cited CM7 figure is for read throughput, so it does not supply a write target. Do not treat an SSD result as proof of FPGA host throughput.

We recommend setting the target from the intended system configuration, then comparing it with the CM7 result only when the interface generation and read direction match; otherwise, label the comparison as a different scope. Keep target values distinct from reported results, too, so a design goal is not mistaken for the source’s reported figure.

When comparing datapath options, use the same named conditions for each entry: interface generation, direction, unit, and device context. If one condition differs, show that difference rather than presenting the values as directly comparable. Keep the scope explicit.

Which workload profiles should guide tuning priorities?

For NVMe host optimization, prioritize two profiles: latency-sensitive enterprise processing, and AI training or inference that reads and writes large data sets, as described in the Flashtec NVMe 5016 article.

Start with the workload.

Workload profile Source-described emphasis Tuning priority
Online transaction processing, financial data processing, and database mining The Flashtec NVMe 5016 description identifies these enterprise applications as sensitive to latency and performance. Keep latency and performance as explicit criteria for this profile.
AI model training and inference The controller description specifies higher throughput for reading and writing large data sets. Prioritize large-data read and write throughput, including storage-to-compute movement.

These profiles call for different evaluation criteria: enterprise examples are described as sensitive to latency and performance, while AI use cases emphasize higher throughput for large data-set reads and writes, according to the Flashtec NVMe 5016 article.

Our recommendation is to keep the profiles separate when setting tuning priorities. For transaction processing, financial data processing, or database mining, record latency and performance as the criteria under consideration. For AI workloads, describe the relevant large-data reads and writes and whether storage-to-compute movement is part of the target use case. That keeps each tuning decision tied to the workload it is meant to address.

Sources

How should sequential and random I/O be assessed separately?

Assess sequential and random I/O as separate tracks: use steady-state, large-block reads and writes for sequential throughput, then vary read/write mix and block size for random I/O, as described in SNIA’s throughput test and IOPS test.

Keep the tracks distinct. Use the table as a test plan, with each result tied to the workload that produced it.

Workload Test basis How to organize the assessment
Sequential I/O SNIA’s throughput test measures large-block sequential I/O at 100% read and 100% write under PTS Steady State conditions (SNIA throughput test). Keep the read and write cases separate, and label each with its block size and steady-state condition.
Random I/O SNIA’s IOPS test covers a broad range of read/write and block-size combinations (SNIA IOPS test). Organize selected combinations by read/write mix and block size; keep each result attached to its combination.

For sequential runs, name each case by transfer direction and block size rather than grouping both directions under one label. Separate read and write results so the direction being assessed remains clear when comparing throughput results. Keep the stated steady-state condition with those results, rather than treating it as an unrecorded detail.

For random runs, use read/write mix and block size as the two axes of the test plan. Choose the combinations that match the workload question, then record the selected values beside each IOPS result. A broad test range can be useful as a plan, but each result should remain tied to its specific mix and block size; do not combine unlike combinations under a single label.

Our recommendation is to review sequential throughput and random IOPS as separate decision inputs. Use the workload definition, not the metric alone, to interpret each result. That keeps the assessment aligned with the question being tested: large-block sequential reads and writes, or random I/O across selected read/write mixes and block sizes.

Where should write caching sit in the FPGA storage path?

Our recommendation: place write caching at the boundary that owns its policy and state; put it in the FPGA NVMe host IP only when your architecture assigns that responsibility there.

Use the workload profile established earlier in this article as an input, then identify which block sets cache policy, owns cache state, and defines when a write counts as complete. Keep ownership explicit.

Candidate placement Choose it when Design question
Inside the FPGA NVMe host IP Your architecture assigns cache policy and state ownership to the host IP. Can the host-IP contract define the cache’s completion point?
At another boundary in the FPGA storage path Your architecture assigns those responsibilities to another block. Can that block state its responsibility and handoff clearly?

Before settling on a location, distinguish the design’s meaning of “write accepted” from “write complete,” if the interface treats them as separate terms. Otherwise, specify the completion point used by the host-IP contract. Then check whether cache policy, state ownership, and completion language point to the same boundary; if they do not, resolve that mismatch before choosing a placement.

When should SATA IP or custom FPGA development be considered instead?

Our recommendation: use SATA IP only for a SATA requirement, and consider custom FPGA development when the IP options under review do not match the required interface or behavior.

Electronic Design describes SATA as using a serial 6Gbps physical layer (PHY) and an instruction set it calls very useful for hard-disk drives (HDDs) and SSDs; use that description as context, not as proof that a particular SATA IP meets a design requirement.

Start with the interface.

Option Consider it when Decision boundary
NVMe host IP The design requirement specifies an NVMe host. Keep it in scope when NVMe is required; do not treat SATA IP as an equivalent substitute.
SATA IP The design requirement specifies SATA. Choose it for a SATA requirement, not simply because the design connects to an SSD.
Custom FPGA development The required interface or behavior is not covered by the IP options under review. Define the specific gap before deciding whether custom development belongs in scope.

Keep the decision tied to the requirement, rather than to a broad label such as “storage.” Write down the required interface and the behavior the host must support, then compare each candidate against those points. If SATA is specified, the cited SATA description gives relevant context, but it does not establish that a particular IP implementation satisfies the rest of the design.

Use the interface requirement to make the first cut. Then state what remains unmet, if anything. If a candidate IP meets the stated scope, keep the comparison focused on that option; if neither candidate fits, use the defined gap to frame a custom-development decision. That keeps the choice between SATA IP and custom FPGA development connected to the actual design need, without assuming that one is a drop-in replacement for another.

In summary

Write down the workload, the latency and throughput requirements, and which one is the hard constraint. Define the PCIe context for each datapath target, then assess sequential and random I/O separately.

Before choosing where write caching belongs, identify which block owns its policy and state and where a write counts as complete. Use those requirements to compare the host-IP options and decide whether SATA IP or custom FPGA development is in scope.

How we wrote this article

The sources we consulted are listed at the end of the article. Where we cite none, it is our opinion, and we write it as such.

This article gives buying criteria and our point of view. It does not replace a professional’s advice on a specific situation.

It is published by Logic Design Solutions, which therefore has an interest in the subject.

Sources
  1. What’s The Difference Between SATA And NVMe? | Electronic Design, electronicdesign.com
  2. Enterprise NVMe SSDs drive performance with PCIe 5.0 …, eenewseurope.com
  3. High-performance PCIe® Gen 5 SSD controller family …, eenewseurope.com
  4. TP (Throughput) Test | SNIA | Experts on Data, snia.org
  5. IOPS (I/Os per Second) Test | SNIA | Experts on Data, snia.org