Industry news — On 6 July 2026 AMD introduced the Versal Premium Gen 2 MoP adaptive SoC, which places up to 32 GB of LPDDR5X on the package for as much as 288 GB/s of on-package memory bandwidth, exposes two PCIe Gen6 x8 interfaces with DMA and CXL 3.1, and cuts board area by up to 60 % — with sampling expected at the end of 2026 and production in the second half of 2027, a timeline that decides on its own whether the part belongs in the programme you are scoping now (CNX Software, 6 July 2026).
In short: Choosing an FPGA for storage IP is an elimination in a fixed order — interface, then transceiver rate and channel count, then the PCIe hard block generation and its Root Complex support — and only then the secondary axes of power, memory, lifecycle and tools, where AMD, Altera and Microchip each win a different one.
That announcement is a good part and a bad answer for most storage programmes reading it. Sampling at the end of 2026 and production in the second half of 2027 means a design that has to ship hardware in 2027 cannot build on it, however attractive the memory bandwidth looks on a slide. This is the recurring shape of the decision: the newest silicon is rarely the right silicon, because a storage design is selected by its availability window, its transceiver reach, its PCIe hard block and its thermal budget long before anyone compares peak specifications. This article is the method we use to get from a storage requirement to a shortlist of two or three parts across AMD, Altera and Microchip — in the order that eliminates the most candidates soonest.
- The short answer
- Start from the interface, not from the device
- The transceiver is the gate for storage line rates
- PCIe hard block generation, and the Root Complex trap
- Why a storage data path is BRAM-hungry, not LUT-hungry
- Power, configuration technology and SEU behaviour
- Control plane, lifecycle and toolchain cost
- A decision table for choosing an FPGA for storage IP
- What does not depend on the FPGA you choose
- Where this stops applying
- Frequently asked questions
The short answer
Choosing an FPGA for storage IP is a sequence of eliminations, not a scoring
exercise. Fix the interface first — SATA at 1.5, 3 or 6 Gbps, NVMe over PCIe, or storage
carried over Ethernet — because that alone sets the serial line rate you must reach and
removes most of the catalogue. Then check three hard gates in order: a serial transceiver
that reaches the rate at the speed grade you can actually buy, in the channel count your lane
budget demands; a PCIe hard block of the right generation that can operate as a
Root Complex, not only as an Endpoint; and enough on-chip memory for the buffering a
storage path lives on. Only then do the secondary axes matter — power, configuration
technology, control plane, lifecycle, tools — and each of those has a different winner among
AMD, Altera and Microchip. There is no overall winner, which is why this page routes to
device-specific ones rather than crowning a family.
Start from the interface, not from the device
The storage interface is the first constraint and it eliminates more candidates than every other
criterion put together. Engineers usually arrive at this decision holding a device family they
already like, and discover three weeks later that it cannot reach the line rate, carries the wrong
PCIe generation, or has the transceivers but not enough of them. Reverse the order and most of that
work disappears. Three storage interfaces are worth considering in an FPGA today, and they impose
different physical requirements:
- SATA runs at 1.5 Gbps (SATA I), 3 Gbps (SATA II) or
6 Gbps (SATA III) over a single serial lane per drive, defined by SATA-IO. One drive, one
transceiver channel. It is the least demanding of the three on the fabric and the most
forgiving on device choice, which is exactly why it survives in instrumentation and avionics long
after the consumer market moved on. - NVMe over PCIe runs at the PCI-SIG link rates — 8 GT/s per lane for
Gen3, 16 GT/s for Gen4, 32 GT/s for Gen5 — and a drive is normally x4, so the
channel count is lanes times drives. Two Gen4 x4 drives is eight transceiver channels running at
16 GT/s, and that requirement alone deletes the majority of every vendor’s catalogue. - Storage over Ethernet moves the physical requirement onto the network side: a
10 GbE link is one transceiver channel at roughly 10 Gbps, and the drives may sit behind
it rather than on the FPGA at all. The transceiver question is the same question, asked about a
different wire.
If you have not settled that choice yet, settle it before you open a device selector. Our
comparison of SATA versus NVMe for
FPGA storage covers the protocol and system-level trade, and if the answer is NVMe, the follow-on
question of Gen4 versus Gen5 and what the PHY
costs you is the one that will decide your device more than any other single input.
The transceiver is the gate for storage line rates
The practical rule is short and it is the one that survives contact with a real project:
you need a serial transceiver that reaches the required line rate at the speed grade you can
actually buy, and you need as many channels as the lane count times the number of drives.
Both halves matter, and the second half is the one that gets missed.
Line rate, and the speed grade you can actually buy
Every FPGA vendor publishes a maximum transceiver line rate per device family, and that number is
a property of the fastest speed grade in the family, not of the part you will end up ordering. The
speed grade is a binning of the same die by achievable performance, and the fastest bin
is typically the most expensive, the hardest to source and the first to go on allocation. LDS states
the same dependency across essentially its whole catalogue, in the plain wording that appears on its
product pages: an IP core’s operating frequency is
“FPGA speed grade operating frequency dependant”. Design against the grade your
procurement team can commit to for the life of the programme, not against the datasheet headline.
A worked example of why the headline is not the answer. AMD’s 7 series overview states that
“the GTP operates up to 6.6 Gb/s, the GTX operates up to 12.5 Gb/s, the GTH operates up to
13.1 Gb/s”. Read only that and an Artix 7 clears SATA III at 6 Gbps with room
to spare. Read the footnote in the same document and the GTP transceivers in the CP, CS, FT and FG
packages support data rates up to 6.25 Gb/s — still enough, but the margin has just gone
from 10 % to 4 %, and it moved because of the package, which is not a column
anybody checks first. In the UltraScale+ generation the same document family gives GTH at
16.3 Gb/s and GTY at 32.75 Gb/s, which is the step that puts Gen4 and then Gen5 signalling
in reach. On the Altera side the tiles are named rather than the transceivers, and the device
overviews state the PCIe result directly — the P-tile as PCIe 4.0 x16 at 16 Gbps and the
R-tile as PCIe 5.0 x16 at 32 Gbps.
There is also a floor, and it surprises people who have optimised for the ceiling. A transceiver
has a minimum line rate as well as a maximum, and the fastest ones do not go slowest. AMD’s
UltraScale documentation puts the minimum at 0.5 Gb/s for its GTH and GTY transceivers but at
9.8 Gb/s for the GTM — which means the very transceiver that reaches 58 Gb/s cannot
be used for a SATA link at 1.5, 3 or 6 Gbps at all. If your design mixes a fast NVMe path with
slow SATA ports, the transceiver class is not one decision but two, and picking the headline part
can leave you with channels that are physically unable to run your slower interface.
We quote no Microchip figure here, deliberately: we could not retrieve Microchip’s published
maxima from a vendor source while writing this, and a third-hand number is worse than none. Take
the figures above from the datasheets listed in the sources, and get Microchip’s from Microchip.
Channel count, which is where designs actually run out
Line rate is a yes-or-no question and it is usually answered correctly. Channel count is an
arithmetic question and it is usually answered late. A four-drive NVMe recorder at x4 per drive
needs sixteen transceiver channels of the right class, plus whatever the rest of the system consumes
— a 10 GbE management link, a JESD204 converter interface, a display output. Transceivers
in an FPGA are grouped in banks with shared reference clocking and shared PLL resources, so
“the device has 24 channels” and “the device has 24 channels you can use the way
you want to” are different statements. Read the bank structure in the vendor’s transceiver
user guide before you commit, not after.
The three vendors’ transceiver-bearing families
All three vendors segment their catalogue the same way, even though the marketing names differ:
cost-optimised families with a modest number of transceivers at a modest rate class, mid-range
families that cover the mainstream, and high-end families carrying the fastest tiles. Rate classes
and channel counts vary by device, package and speed grade, so treat the table below as a map of
where to look rather than as a specification, and confirm the figures in the vendor documentation
linked in the sources. A dash means we have no sourced entry for that tier, not that nothing
exists.
| Tier | AMD | Altera | Microchip |
|---|---|---|---|
| Cost-optimised, SATA-class rates | Artix 7 (GTP transceivers) | Cyclone V GX/GT, Cyclone 10 GX | IGLOO2, SmartFusion2 |
| Mid-range, mainstream PCIe rates | Kintex 7, Artix UltraScale+, Kintex UltraScale and UltraScale+ | Arria V GX, Arria 10 GX | PolarFire, PolarFire SoC |
| High-end, fastest transceiver tiles | Virtex UltraScale+, Versal | Stratix 10, Agilex 5, Agilex 7 | — |
| With hard processor complex | Zynq 7000, Zynq UltraScale+, Versal | Arria 10 SoC, Agilex 7 SoC, Agilex 5 SoC | PolarFire SoC (RISC-V) |
| Space / high-reliability | Radiation-tolerant Kintex UltraScale line | — | RTG4, and the flash-based PolarFire line |
Read the shape of that table rather than its cells. Microchip positions PolarFire on power and on
configuration technology rather than on peak line rate, so a Gen5 NVMe host is an AMD or Altera
conversation; a SATA recorder that has to sit fanless in a sealed enclosure is a conversation that
starts with Microchip. Same method, different answers, which is what a method is for.
PCIe hard block generation, and the Root Complex trap
The PCIe hard block in the device defines the ceiling for an NVMe design, and it is a separate
question from the transceiver. A device can carry transceivers fast enough for Gen4 signalling and
still expose only a Gen3 hard block, in which case Gen4 is available to you only by building the
controller in soft logic — a project in its own right, not a configuration option.
Two consequences follow. First, the PCIe generation you can reach is the generation of the hard
block, and it varies within a vendor’s range far more than device families suggest: LDS’s own
published NVMe Host builds show the pattern clearly, with Gen2 and Gen3 variants on Kintex and Zynq
parts, a Gen4 variant on Artix UltraScale+, and Gen5 on Agilex 7 with the R-Tile. The device
family name tells you where to look; the hard block tells you what you get.
Second, and this is the one that costs schedules: hosting a drive is not the Endpoint use
case that most reference designs assume. An FPGA that talks to an NVMe SSD is the
host — it must operate its PCIe port as a Root Complex, enumerate the drive,
configure its base address registers and own the link. Almost every PCIe example design a vendor
ships, and almost every application note you will find, puts the FPGA in the Endpoint role plugged
into a server. Confirming that the specific hard block on the specific device supports Root Port
mode, at the generation and lane width you need, is a gate item and belongs on the checklist before
the board is laid out. Our companion piece on
PCIe on FPGA, endpoints, Root
Complex and NVMe works through the distinction and what it changes in the design.
The good news is that Root Port mode is broadly available and the vendors say so plainly, so this
is a check you close quickly rather than a risk you carry. AMD’s UltraScale overview states that
“the integrated blocks for PCIe can be configured for Endpoint or Root Port, supporting a
variety of link widths and speeds depending on the targeted device speed grade and package”.
Altera documents both the P-tile and the R-tile as “2×8 endpoint or 4×4 root
port”, and its Arria and Cyclone overviews state Root Port support at their own generations.
Read the sentence for your device, and note the qualifier in the AMD wording: speed grade
and package are part of what you are allowed.
Two traps in that last qualifier are worth naming, because both are the kind of thing found after
a board has been built. On Artix UltraScale+, AMD’s overview restricts Gen4 to specific devices in a
specific package — the AU10P and AU15P in the FFVB676 package — and describes the block
as operating “in compatibility mode” for 16.0 GT/s. Choosing the family is not the
same as choosing a Gen4-capable orderable part. And on Zynq UltraScale+, the four PS-GTR lanes in the
processing system serve both the PS PCIe controller, which AMD documents at
“PCIe at 5.0 GT/s (Gen 2) as a root complex or Endpoint in x1, x2, or x4
configurations”, and the PS SATA controller at
“1.5 Gb/s, 3.0 Gb/s, or 6.0 Gb/s data rates”. They compete for the same
four lanes. A design that assumed it could put SATA drives on the PS and an NVMe root port on the PS
at full width has a lane budget problem, not a licensing problem, and the PL blocks are where the
higher PCIe generations live in that device.
Why a storage data path is BRAM-hungry, not LUT-hungry
A storage data path consists mostly of buffering, and that changes which resource column you
should be reading in a device selector. The logic itself — a link layer state machine, a
transport layer, command and completion queue handling, CRC — is modest by modern FPGA
standards. What is not modest is memory. Every one of these eats block RAM:
- Elastic buffering between clock domains. The transceiver runs on a recovered or
reference clock; your application runs on a system clock. LDS’s published SATA figures show the
scale of the domain the link side works in — a system clock of 37.5 MHz at 1.5 Gbps,
75 MHz at 3 Gbps and 150 MHz at 6 Gbps — and every crossing needs a FIFO
deep enough to absorb the rate mismatch without dropping a sector. - Command and queue structures. An NVMe host maintains submission and completion
queues in memory. The LDS NVMe Host IP publishes an I/O queue buffer that is configurable at 32, 64,
128, 256, 512 or 1024 KB, with up to 16 independent I/O queues — that configurability is
precisely a memory-versus-performance dial, and it is the dial you turn when the device is tight. - Sector and payload staging. Sectors are 512 B or 4096 B, and a
recorder writing continuously needs enough staging to ride out the drive’s own latency excursions.
This is usually the largest single consumer, and it is the one that scales with your required
sustained rate rather than with the protocol. - Multi-drive striping. RAID0 across several drives multiplies the staging, not
the logic. A four-drive design is roughly four times the buffer and roughly the same state machine.
The practical consequence is that a device with plenty of logic and thin on-chip memory will
disappoint you, and one size down with generous block RAM will not. If the sustained rate is high
enough that on-chip memory runs out, the answer is external memory — and the memory
controller, its width and its own hard block then join the device decision. That is exactly where
the memory-on-package direction of the Versal Premium Gen 2 part in the news hook becomes
interesting, and equally where its 2027 production date becomes the deciding fact rather than a
footnote.
What none of this gives you is a number. LDS does not publish LUT, flip-flop, block-RAM or DSP
utilisation for any of its cores, and this article will not estimate them; the figures depend on the
device, the speed grade and the configuration. Ask for them against your exact target and treat any
figure you have not received in writing as unknown.
Power, configuration technology and SEU behaviour
Configuration technology splits the three vendors cleanly and it is the axis where the answer is
least ambiguous. AMD and Altera FPGAs are SRAM-based: the configuration is volatile, loaded at
power-up from an external flash device, and held in configuration memory cells for as long as the
part is powered. Microchip’s PolarFire and PolarFire SoC are flash-based: the configuration lives in
non-volatile cells inside the device. Three things follow, and together they are why Microchip
belongs on the shortlist of any storage programme that is not purely about peak throughput.
Static power and thermal design
Flash configuration cells do not leak the way SRAM configuration cells do, which is the physical
origin of the low static power Microchip markets the family on. For a storage design that shows up
as a system property rather than a datasheet line: a lower thermal budget means a smaller heatsink,
no fan, a sealed enclosure, a wider ambient range, and in a battery or solar-powered recorder it
means mission duration. If your enclosure is thermally constrained, this axis outranks line rate,
which is why we have a device-specific article on
NVMe Host IP on PolarFire SoC for
low-power designs.
Instant-on, and no external configuration flash
A non-volatile device does not wait for a bitstream to be shifted in at power-up, and it removes
that external configuration memory — with its own qualification, supply chain and failure
mode — from the bill of materials. For a recorder that must be capturing within a bounded time
of power being applied, that is a functional requirement rather than a convenience.
Single-event upsets
A single-event upset is a bit flip caused by an ionising particle. In an SRAM-based
FPGA, an upset can land in configuration memory and change the circuit itself, which is why
SRAM-based parts in radiation environments carry configuration scrubbing, ECC on configuration
memory and triple modular redundancy in the design. A flash-based configuration cell is not
susceptible in the same way, which is the structural argument behind Microchip’s position in space
and high-reliability programmes and behind the RTG4 line. This does not make user flip-flops or
block RAM immune — those are still SRAM and still need mitigation — and it does not
replace a radiation test campaign. Our note on
radiation-hardened and
space-grade FPGAs, and what they mean for IP covers what actually changes in the IP when this
constraint is real.
Control plane, lifecycle and toolchain cost
Three secondary axes are decisive often enough that they belong in the first pass rather than the
last, and each has a different answer.
Do you need a hard processor complex?
A storage data path does not need a CPU. A storage system often does: mounting a file
system, servicing a management interface, running a health monitor, handling configuration. All
three vendors sell that as a hard block rather than a soft core — AMD in Zynq 7000, Zynq
UltraScale+ and Versal; Altera in the SoC variants of Arria and Agilex; Microchip in PolarFire SoC,
whose processor complex is RISC-V rather than Arm. The distinction that matters is not which
architecture but where the boundary sits: everything on the critical data path belongs in fabric,
and everything episodic belongs on the processor. Our piece on
FPGA versus CPU, and when hardware
parallelism wins is the long version of that boundary argument, and
integrating NVMe Host IP on Zynq
UltraScale+ is what it looks like when the control plane and the data path share a device.
Longevity, and why it often outranks performance
For avionics, defence, rail, medical and instrumentation, the question “will I still be
able to buy this part in twelve years” outranks every performance criterion, because a
redesign at year eight costs more than the entire original development. The thing to obtain is not a
marketing page about product longevity but a written commitment, from the vendor or the distributor,
against the specific ordering part number, package and speed grade — together with the
end-of-life notification terms that come with it — and to obtain it before the schematic is
frozen rather than after. Ask all three; the answers differ. The corollary is the news hook:
sampling at the end of 2026 and production in the second half of 2027 is a strong reason for a
programme with a 2027 hardware milestone to select something mature instead, however good the new
part is.
Toolchain and licensing
The three flows are Vivado and Vitis for AMD, Quartus Prime for Altera, and Libero SoC for
Microchip, and the device you choose decides which one your team lives in for the next several
years. All three vendors tier their editions, and the tiering is by device as well as by feature
— which is why a family that looks cheap per unit can carry a per-seat tool licence that
another family from the same vendor does not.
This is also the axis where the received wisdom goes stale fastest, so treat any tool-cost figure
you were given more than a year ago as expired. AMD’s own licensing page states that
“AMD Vivado Design Suite is evolving to a new tiered licensing model starting with the 2026.1
release”, replacing the old edition split with a tier ladder. A free tier survives — the
page answers its own question, “Will the Vivado Design Suite still be free? Yes. The BASIC tier
remains free with annual renewal” — but what that tier covers is the part that
decides your budget. AMD lists BASIC as suitable for small designs across the 7 series, Spartan
UltraScale+, Artix UltraScale+, the smaller Kintex UltraScale+ parts and the lower Zynq UltraScale+
MPSoC range. Larger UltraScale+ parts and Versal sit above it. For a storage design that is a direct
coupling: pushing from a Gen3-class part up to a Gen4 or Gen5 part can move you into a paid tier at
the same time as it moves your BOM, and the two costs are easy to budget separately and then be
surprised by together.
We print no equivalent table for the other two flows, because we could not verify Quartus Prime’s
and Libero SoC’s current edition boundaries from a vendor source — and a coverage list copied
from an article is stale the moment a release ships. Check the licensing page for the exact ordering
part number, and budget tool cost per seat per year alongside device cost per unit: on a small team
at low volume, the tool line can be the larger of the two.
A decision table for choosing an FPGA for storage IP
The table below pairs a concrete requirement with the family class to shortlist across all three
vendors, and with the LDS article that covers that route in device-specific detail. It is a routing
table, not a recommendation: the shortlist is where to start reading datasheets, and the final part
number comes out of your own channel count, thermal budget and availability window.
| Requirement | Family class to shortlist | Where to go next |
|---|---|---|
| NVMe Gen4 host, 2 drives, industrial temperature | Eight transceiver channels at Gen4 class with a Gen4 hard block in Root Port mode. AMD: Artix UltraScale+ is where LDS publishes its own Gen4 NVMe build, but AMD restricts Gen4 there to the AU10P and AU15P in one package — verify the ordering part, not the family; Virtex UltraScale+ and Versal sit above it. Altera: Agilex with the P-tile, documented at PCIe 4.0 x16. Verify the hard block generation device by device. | Gen4 vs Gen5 NVMe host IP and the PHY |
| NVMe Gen5 host, maximum sustained rate | High-end tiles only: Altera Agilex 7 with R-Tile, or AMD Versal class. Verify Root Port support at Gen5 explicitly. | NVMe Host IP on Agilex 7 R-Tile |
| NVMe host with a control plane on the same device | SoC families: AMD Zynq UltraScale+ or Versal; Altera Agilex SoC or Arria SoC; Microchip PolarFire SoC. | NVMe Host IP on Zynq UltraScale+ |
| SATA recorder, low power, sealed enclosure or radiation environment | Flash-based configuration first: Microchip PolarFire or PolarFire SoC. Compare against a mid-range SRAM part only if the thermal budget allows it. | Low-power storage IP on PolarFire SoC |
| SATA III host, multi-drive, mid-range budget | One 6 Gbps channel per drive: AMD Kintex 7 or Kintex UltraScale; Altera Arria 10 GX or Stratix; Microchip PolarFire. | SATA Host IP on Kintex UltraScale |
| SATA host on an Altera-standardised platform | Cyclone GX for cost, Arria GX for mid-range, Stratix GX for the high end — matched to the SATA generation you need. | Altera SATA Host IP on Arria, Cyclone and Stratix |
| SATA device emulation, lowest cost | Cost-optimised families with a single transceiver channel at the target SATA rate: AMD Artix 7; Altera Cyclone GX; Microchip IGLOO2 or SmartFusion2 at the lower SATA generations. | SATA Device IP: an FPGA that presents itself as an SSD |
| Storage exported over Ethernet rather than attached locally | The gate moves to the network transceiver and the MAC; the drives may not be on this device at all. | NVMe SAN IP for storage networks |
| Still deciding between SATA and NVMe | Decide this before opening a device selector — it sets the transceiver requirement that eliminates most candidates. | SATA vs NVMe for FPGA storage |
What does not depend on the FPGA you choose
Some of the work is identical whichever of the three vendors you pick, and budgeting for it is
part of choosing honestly. The IP has to be integrated into your fabric, connected to your bus,
constrained, and timing-closed on the device you selected — at the speed grade you can buy,
with your other logic already competing for the same routing. A core that closes timing comfortably
on a fast-grade evaluation board is not automatically a core that closes on a mid-grade production
part with 70 % of the fabric already occupied. That work is the subject of our note on
timing closure for high-speed
storage interfaces, and it is the reason a validated starting point is worth more than a
datasheet.
Two facts about how LDS delivers its cores are directly relevant to that portability question,
and both come from the company’s own published product pages rather than from anyone’s estimate.
First, the cores are supplied in source: the product pages state that an IP is
“Available in VHDL source code format for ease of customization” and that it
“Can be customised by Logic Design Solutions”. Source delivery is what makes the same
core deployable on a Kintex and on a PolarFire without a redesign at the boundary, and what makes it
possible to change a filter, add a counter or adapt a bus interface when your system demands it.
Second, the published catalogue names validated targets and evaluation boards explicitly, across
all three vendors. The SATA Host line is published on AMD parts from Spartan 6 and Artix 7
through Kintex, Virtex and UltraScale+, and on Altera Cyclone GX, Arria GX and Stratix GX
devices. The NVMe Host line is published on Kintex 7 with the KC705, Virtex 7 with the
VC707, Zynq 7 with the ZC706, Zynq UltraScale+ and Kintex UltraScale, Artix UltraScale+ for
Gen4, Agilex 7 R-Tile for Gen5, and PolarFire and PolarFire SoC. A named board is a stronger
statement than a named family, because it implies a specific transceiver type, a specific reference
clock arrangement and a constraint set that already exists. If you want the catalogue view rather
than the device view, the complete guide
to NVMe Host IP for FPGA and the
SATA host and device IP selection
guide list them.
What the published material does not contain is performance numbers per device. The only
throughput figure LDS has ever published is roughly 500 MB/s on a single 6 Gbit SATA link,
in a 2019 news item. There is no published Gen4 or Gen5 throughput, no latency, no power figure and
no resource utilisation, and this article invents none of them. Those depend on the device, the
speed grade and the configuration — ask for them against your exact target.
Tell us the storage interface, the number of drives, the temperature and power envelope and the year you have to ship, and we will tell you which families survive that filter and what the IP needs on each of them.
Where this stops applying
This is a method for narrowing a catalogue, not a substitute for the vendor documentation or for
a bring-up. Its boundaries are worth stating:
- Family names are not part numbers. Transceiver count, transceiver class, PCIe
hard block generation and on-chip memory all vary within a family by device, package and speed
grade. Everything here tells you where to look; the datasheet tells you what you get. - No device figures are quoted or estimated. This article deliberately gives no
line rates per device, no LUT or block-RAM counts, no power figures and no throughput. Where a
number would be expected, the answer is that it depends on device, speed grade and configuration,
and should be obtained in writing for your exact target. - Availability changes faster than the analysis. Allocation, end-of-life notices
and new tiles move the answer. A shortlist built more than a few months ago should be re-checked
against current lead times before the schematic is frozen. - Free tool editions move between releases. Device coverage of the no-cost
editions of Vivado, Quartus Prime and Libero SoC changes; verify coverage for your exact device on
the vendor’s current download page rather than from any article, including this one. - Certification is a separate axis entirely. If you are working to DO-254, ECSS
or a comparable regime, the design assurance level, the documentation package and the vendor’s
qualification data drive the choice in ways this method does not model.
In summary
Choosing an FPGA for storage IP works best as an elimination in a fixed order.
Fix the interface, because it sets the line rate. Check the transceiver class and, more importantly,
the channel count at the speed grade you can actually procure. Check the PCIe hard block generation
and confirm Root Port support, because hosting a drive is not the Endpoint case the reference
designs assume. Size the on-chip memory against the buffering, not the logic against the state
machines. Only then weigh power and configuration technology, the control plane, the lifecycle
commitment and the tool licences — and expect a different vendor to win on each.
Applied honestly, that method produces AMD for the widest mid-range and high-end choice, Altera
for the leading PCIe tiles and a strong mid-range, and Microchip for anything where power,
non-volatile configuration or radiation behaviour is the binding constraint. What it never produces
is one right family for everyone, which is why this page exists to route you to the device-specific
articles rather than to answer for you. And whichever part you choose, the same integration work
follows: the IP still has to be connected, constrained and timing-closed on your device, at your
grade, next to your logic.
Frequently asked questions
Which FPGA vendor is best for storage IP?
None of them, in the abstract. The method decides it: the interface sets the line rate, the transceiver class and channel count eliminate most candidates, and the PCIe hard block generation decides whether NVMe Gen4 or Gen5 is available at all. After those gates, AMD offers the widest mid-range and high-end choice, Altera carries the leading PCIe tiles, and Microchip wins wherever low power, non-volatile configuration or radiation behaviour is the binding constraint. Different requirements produce different winners from the same method.
How many transceiver channels does an NVMe design need?
Lane count times the number of drives, plus whatever the rest of the system consumes. A drive is normally x4, so a two-drive host is eight channels and a four-drive recorder is sixteen, before you add a management Ethernet link or a converter interface. Channel count, not line rate, is where designs usually run out. Transceivers are also grouped in banks with shared reference clocking, so read the bank structure in the vendor’s transceiver user guide rather than trusting a single headline number.
Why does Root Complex support matter when choosing an FPGA?
Because an FPGA that talks to an SSD is the host, not the peripheral. It has to operate its PCIe port as a Root Complex — enumerate the drive, configure its base address registers and own the link — whereas almost every PCIe reference design a vendor ships puts the FPGA in the Endpoint role plugged into a server. Confirm that the specific hard block on the specific device supports Root Port mode at the generation and lane width you need before the board is laid out, not after.
Should I choose a device on logic capacity or on memory?
On memory, for a storage data path. The state machines — link, transport, command and completion queue handling, CRC — are modest by modern standards, while the buffering is not: elastic FIFOs across clock domains, I/O queue structures, sector staging at 512 B or 4096 B, and multiplied staging when several drives are striped. A device with plenty of logic and thin on-chip memory will disappoint; one size down with generous block RAM often will not.
What makes Microchip PolarFire different for storage designs?
Its configuration is flash-based rather than SRAM-based, which is a structural difference rather than a marketing one. Flash configuration cells do not leak the way SRAM configuration cells do, which is the physical origin of the low static power the family is sold on; the device needs no external configuration flash and is configured at power-up; and a configuration upset from an ionising particle cannot change the circuit the way it can in an SRAM-based part. User flip-flops and block RAM still need mitigation, and none of this replaces a radiation test campaign.
Is the newest FPGA family usually the right choice for a storage programme?
Rarely. Availability window, not peak specification, is what selects a device. A part sampling at the end of 2026 and reaching production in the second half of 2027 cannot carry a programme that has to ship hardware in 2027, whatever its memory bandwidth. For avionics, defence, rail and instrumentation the reverse question dominates: a written longevity commitment against the specific ordering part number, package and speed grade, obtained before the schematic is frozen, is worth more than a generation of headline performance.
- SATA-IO — Serial ATA International Organization specifications: the body that defines the SATA 1.5, 3 and 6 Gbps generations quoted as the line-rate requirement for a SATA design.
- PCI-SIG — PCI Express Base Specification: the body that defines the PCIe link rates per lane that set the transceiver requirement for an NVMe design.
- NVM Express, Inc. — NVM Express Specifications: the command, queue and namespace model an FPGA NVMe host implements above PCIe.
- AMD — 7 Series FPGAs Data Sheet: Overview (DS180), v2.6.1: the GTP, GTX and GTH line rates quoted in the transceiver section, the package restriction on Artix 7 GTP rates, and the Endpoint and Root Port capability of the 7 series integrated blocks for PCIe.
- AMD — UltraScale Architecture and Product Data Sheet: Overview (DS890), v4.10, 21 May 2026: the UltraScale and UltraScale+ transceiver types and their maximum and minimum data rates, the Endpoint or Root Port configurability of the integrated PCIe blocks, the Artix UltraScale+ Gen4 device and package restriction, and the Zynq UltraScale+ PS-GTR lanes shared between the PS PCIe and PS SATA controllers.
- AMD — Versal Architecture and Product Data Sheet: Overview (DS950), v2.11, 3 August 2026: the Versal transceiver rate classes and the Endpoint or Root Port configurability of the PL PCIE and CPM blocks referenced in the high-end tier.
- AMD — Vivado Design Suite licensing options: the tiered licensing model introduced at the 2026.1 release, the statement that the BASIC tier remains free with annual renewal, and the device coverage of that tier quoted in the toolchain section.
- Altera — Agilex 7 FPGAs and SoCs Device Overview: the E-tile, P-tile, F-tile and R-tile rate classes, the PCIe 4.0 x16 and PCIe 5.0 x16 hard IP figures, and the endpoint and root port configurations quoted for the P-tile and R-tile.
- Altera — Agilex 5 FPGAs and SoCs Device Overview: the GTS transceiver rate classes and PCIe 4.0 hard IP lane widths referenced for the Altera mid-range and high-end tiers.
- Microchip Technology — FPGAs and PLDs product range: the PolarFire, PolarFire SoC, RTG4, IGLOO2 and SmartFusion2 families referenced in the family map, the flash-based non-volatile configuration and low static power positioning, and the RISC-V processor complex named in the control-plane comparison. Confirm the current transceiver, PCIe and power figures for your candidate part against the device datasheet before you commit to it.
- Logic Design Solutions — NVMe Host IP product pages: the published NVMe and SATA variants, their validated targets and evaluation boards, the I/O queue buffer sizes, sector sizes and SATA system clock frequencies quoted here.
- CNX Software — AMD Versal Premium Gen 2 MoP adaptive SoC integrates 32GB LPDDR5X, 6 July 2026: the July 2026 announcement, its PCIe Gen6 and CXL 3.1 figures and its sampling and production timeline, cited in the opening.
