Industry update — On 4 August 2026, NVM Express published its 2026 specification set, extending features such as controller-based rate limiting and subsystem migration — all defined above the physical layer, a reminder that the Gen4-versus-Gen5 decision on an FPGA is set by the PCIe link and tile, not by the NVMe command set (NVM Express, 4 August 2026).
Updated 16 July 2026 · Written by the Logic Design Solutions engineering team
Bottom line up front: Gen4 vs Gen5 for an NVMe Host IP core on FPGA is decided by the PCIe PHY tile your device carries, not by the IP. Gen5 doubles per-lane bandwidth over Gen4, but it only pays off when your sustained write rate exceeds a Gen4 x4 link.
Two NVMe Host designs can share almost identical protocol logic and still perform an order of magnitude apart. The difference is rarely the IP core itself — it is the PCIe layer underneath it. Choosing between PCIe Gen4 and Gen5 for an NVMe Host IP core on FPGA is, in practice, a decision about the PHY: which hardened PCIe tile your device carries, and whether your board can hold the link together at the higher signaling rate. This guide walks through what actually changes when you step up a generation, and how to read your own FPGA to pick the right one.
- The short answer
- What “at the PHY” actually means
- How much faster is Gen5, lane by lane?
- Which FPGAs support Gen4 and Gen5?
- What stays the same in the NVMe Host IP
- How to choose Gen4 or Gen5
- Can Gen4 x8 replace Gen5 x4?
- Why signal integrity is the real cost of Gen5
- Feeding the link is still on you
- When Gen5 is the wrong call
- Frequently asked questions
The short answer
- The PHY tile sets your ceiling. You can only run the generation your device’s hardened PCIe block supports. Among common FPGA families, Gen5 today effectively means the Altera Agilex 7 R-Tile.
- Gen5 roughly doubles per-lane bandwidth over Gen4 (32 GT/s vs 16 GT/s per lane, per the PCI-SIG specification). That only matters if your sustained write rate genuinely exceeds a Gen4 x4 link.
- The NVMe Host protocol logic barely changes between generations. What changes is the tile beneath it and the signal-integrity budget on your board.
- Gen4 is the lower-risk default for most recorders. Reach for Gen5 only when the rate demands it and the board is engineered for the tighter channel.
What “at the PHY” actually means
In an FPGA-based NVMe Host, the path to the SSD is layered. At the top sits your application, streaming capture data into the host core. The NVMe Host IP implements the protocol — the command set, the submission and completion queues, the doorbells, the data movement. Underneath that is PCIe, split into the transaction, data-link and physical (PHY) layers. On modern FPGAs the PCIe layers are provided by a hardened block or tile in silicon, so the fabric is free for protocol and application logic.
The PHY is the part that actually drives the serial lanes at multi-gigahertz signaling rates. It is also the part that is fixed in silicon: a given tile is built for a given maximum PCIe generation. That single fact is why the Gen4-versus-Gen5 decision is a PHY decision. The protocol above can be written to support either; the tile below decides which one you can physically run, and how hard the board has to work to run it.
How much faster is Gen5 than Gen4, lane by lane?
Each PCIe generation roughly doubles the per-lane signaling rate of the one before it. The figures below are the theoretical rates defined by the PCI-SIG specification, using 128b/130b encoding; real designs land somewhat below them once protocol and system overheads are included.
| Generation | Per-lane signaling | Approx. usable / lane | Approx. usable x4 |
|---|---|---|---|
| PCIe Gen3 | 8 GT/s | ~1 GB/s | ~3.9 GB/s |
| PCIe Gen4 | 16 GT/s | ~2 GB/s | ~7.9 GB/s |
| PCIe Gen5 | 32 GT/s | ~4 GB/s | ~15.8 GB/s |
The practical reading is simple. A Gen5 x4 link carries roughly the bandwidth of a Gen4 x8 link, and about double a Gen4 x4. For an NVMe Host core, that headroom shows up as higher sustained write throughput — the figure that decides whether a high-speed recorder can keep pace with its capture stream without dropping data. If your rate already fits comfortably inside a Gen4 x4 link, the extra Gen5 bandwidth is capacity you are paying for but not using.
Which FPGAs support Gen4 and Gen5?
Because the ceiling is set by the hardened PCIe block, the honest first question is not “Gen4 or Gen5?” but “what does my device actually carry?” The table below sketches, at a general level, where the common families sit. Exact capabilities vary by specific device and speed grade, so always confirm against the vendor documentation for your part.
| Family | Hardened PCIe (typical max) | Gen5-capable? |
|---|---|---|
| AMD (Xilinx) Artix-7, Kintex-7 | Integrated block, up to Gen2 | No |
| AMD Kintex / Zynq UltraScale+ | Integrated block, up to Gen3 | No |
| Altera Cyclone, Arria | Hard IP, up to Gen3 | No |
| Altera Stratix 10 | H-Tile (Gen3), P-Tile (Gen4) | No |
| Altera Agilex 7 | P-Tile (Gen4), R-Tile (Gen5) | Yes (R-Tile) |
| Microchip PolarFire SoC | Hard block, Gen2 | No |
On the AMD side
The 7-series workhorses — Artix-7 and Kintex-7 — carry an integrated PCIe block topping out around Gen2. The UltraScale+ generation (Kintex and Zynq UltraScale+) raises that to Gen3. Gen4 and Gen5 hardened PCIe appear on newer AMD devices outside these particular families. So within the AMD parts most commonly paired with storage IP, Gen4 x4 is not something the integrated block provides — the practical high-bandwidth ceiling is Gen3, and you scale bandwidth with lane width rather than generation.
On the Altera side
This is where Gen4 and Gen5 are cleanly separated by tile. Stratix 10 offers Gen3 through its H-Tile and Gen4 through the P-Tile. Agilex 7 carries the P-Tile for Gen4 and the R-Tile for Gen5. If your requirement genuinely needs Gen5, the Agilex 7 R-Tile is the mainstream route to it among these families — a path we cover in depth in our guide to NVMe Host IP on the Agilex 7 R-Tile.
On the Microchip side
PolarFire SoC provides a hardened PCIe block at Gen2. It is a strong low-power, cost-sensitive platform, but it is not the device to reach for when the whole point is maximum sustained NVMe write throughput.
The takeaway across all three vendors: among the families most designers reach for, Gen5 today effectively means the Agilex 7 R-Tile. That narrows the “Gen4 vs Gen5” debate considerably before you have written a line of RTL.
Not sure which tile your target device carries, or whether your rate really needs Gen5? A short design review answers it before you commit to silicon.
What stays the same in the NVMe Host IP
It is worth being explicit about what does not change when you move a generation. The NVMe protocol is agnostic to the PCIe generation beneath it. The command set, the queue model, the way the host rings doorbells and reaps completions — all of that is identical whether the link is Gen4 or Gen5. A well-designed NVMe Host IP core presents the same interface to your application in both cases.
What differs sits below the protocol boundary: the hardened tile it binds to, the number of lanes, and the physical constraints of running those lanes reliably. In other words, choosing a generation is mostly an integration and board decision, not a rewrite of your storage logic. That is good news for portability — a design targeting Gen4 today can, in principle, move to a Gen5 tile later without re-architecting the storage stack, provided the board is built for it.
How to choose Gen4 or Gen5
If you are still choosing between NVMe and SATA in the first place, that is a separate decision we cover in SATA vs NVMe for FPGA storage; here we assume NVMe is the target. With the PHY realities in view, the decision comes down to a handful of factors. Work through them in order — the first two usually settle it.
| Factor | Points to Gen4 | Points to Gen5 |
|---|---|---|
| Sustained write rate | Fits inside a Gen4 x4 link | Exceeds a Gen4 x4 link |
| Available PHY tile | P-Tile, UltraScale+, Stratix 10 | Agilex 7 R-Tile |
| Board signal-integrity budget | Standard, comfortable margin | Tight; needs careful layout and validation |
| Design effort and risk | Lower | Higher |
| Typical fit | Most single-source recorders | Highest-rate or multi-stream recorders |
The single most useful test is the first row: measure or estimate your worst-case sustained write rate, compare it against Gen4 x4, and only look at Gen5 if you genuinely exceed it. Newest is not automatically best; the right generation is the lowest one that clears your rate with margin.
Can Gen4 x8 replace Gen5 x4?
Doubling the generation is not the only way to double bandwidth. A Gen4 x8 link carries roughly the same aggregate as a Gen5 x4 link, because bandwidth scales with both generation and lane count. If your device has a Gen4-capable tile with lanes to spare, widening the link can reach your target without stepping up to the tighter Gen5 channel.
Which route wins depends on specifics: how many high-speed lanes the tile and package pinout expose, how the SSD and connector are wired, and where your board’s signal-integrity margin is most comfortable. Sometimes a Gen4 x8 path on a mature tile is the lower-risk way to hit a rate that a Gen5 x4 path would also reach. It is a question worth asking explicitly rather than defaulting to the newest generation.
Why is signal integrity the real cost of Gen5?
The higher a PCIe generation signals, the less margin the channel has. At Gen5’s 32 GT/s, the PCB stack-up, connectors, vias and overall routing have a tighter budget than at Gen4. The silicon supports the rate, but realizing it depends on a board engineered for it — controlled impedance, careful length matching, attention to insertion loss, and validation that proves the link holds at rate under real conditions, not merely that it trains up at power-on.
This is the practical price of the extra bandwidth, and it is why a Gen5 decision should never be made on the datasheet alone. Budget the verification effort. A marginal Gen5 board that intermittently drops to a lower rate, or fails to sustain throughput under thermal load, is worse than a solid Gen4 design that meets the requirement comfortably.
Feeding the link is still on you
Finally, a faster PHY does nothing if the rest of the path cannot keep up. NVMe throughput comes from keeping many command queues full in parallel; a link that is starved of commands or fed by a data source that cannot sustain the rate will deliver no more at Gen5 than at Gen4 — for considerably more design effort. Before committing to a generation, confirm that your data producer, your buffering, and the host core’s queue handling can actually saturate the link you are choosing. The PHY sets the ceiling; the rest of the design decides how close you get to it. Getting capture data into the core efficiently is its own design task — see our guide to connecting NVMe Host IP to the FPGA fabric over AXI.
When Gen5 is the wrong call
Stay on Gen4 (or lower) when:
- Gen4 already covers your rate. If a Gen4 x4 — or x8 — link comfortably clears your sustained write need, Gen5 adds cost and signal-integrity risk for no benefit.
- Your device has no Gen5 tile. Outside the Agilex 7 R-Tile, the common families here do not offer hardened Gen5. Scale with lane width on the tile you have rather than forcing a generation the silicon does not provide.
- The board cannot hold the channel. If the PCB and connectors cannot meet the Gen5 budget, the achievable rate is set by the board, not the tile — and you inherit intermittent-failure risk on top.
- The rest of the path is the bottleneck. If your data source or buffering cannot sustain more than Gen4 delivers, a Gen5 link buys nothing but complexity.
Frequently asked questions
What actually determines whether I can run Gen5 on my FPGA?
The hardened PCIe PHY tile or block on the device, not the NVMe Host IP. The IP implements the protocol above PCIe; the tile below fixes the maximum PCIe generation you can physically run. Confirm your specific device’s PCIe capability before committing to a generation.
Which FPGA families support PCIe Gen5 today?
Among the families commonly paired with storage IP, Gen5 effectively means the Altera Agilex 7 R-Tile. Stratix 10 and the Agilex 7 P-Tile reach Gen4; AMD UltraScale+ tops out around Gen3 and the 7-series around Gen2; PolarFire SoC provides Gen2. Newer AMD devices outside these families add Gen4/Gen5. Always confirm per device.
Is Gen5 always twice as fast as Gen4 in practice?
No. Gen5 doubles the per-lane signaling rate, but you only see the benefit if the rest of the path keeps up: full command queues, a data source that sustains the rate, and a board whose signal integrity holds at 32 GT/s. Under-feed the link and Gen5 delivers no more than a well-fed Gen4 one.
Does the NVMe Host IP itself change between Gen4 and Gen5?
The NVMe protocol logic is essentially the same — the command set, queues and completions do not depend on the PCIe generation. What changes is the hardened tile the core binds to, the lane configuration, and the signal-integrity budget on the board.
Can I use Gen4 x8 instead of Gen5 x4?
Often, yes. A Gen4 x8 link carries roughly the same aggregate bandwidth as a Gen5 x4 link, since bandwidth scales with both generation and lane count. Which is better depends on the lanes your tile and pinout expose and where your board has the most signal-integrity margin.
What is the main design risk when moving to Gen5?
Signal integrity. At 32 GT/s the channel budget is tighter than at Gen4, so the PCB stack-up, connectors and routing have less margin. The verification effort should prove the link sustains its rate under real conditions, not just that it trains up at power-on.
Weighing Gen4 against Gen5 for a storage path? Tell us your target device and sustained write rate, and we will help you confirm the generation, the tile and the lane width before you commit.
