Wednesday, February 4th 2026

AMD Announces Kintex UltraScale+ Gen 2 FPGAs to Capture the Middle of the Market

AMD today unveiled the Kintex UltraScale+ Gen 2 FPGA family. The company's latest announcement is a strategic play for the middle-of-the-market in the FPGA segment that's dominated by the likes of Intel Agilex 5 series. Targeting sectors that demand longevity and reliability, such as medical imaging, industrial automation, and broadcast video production, the Gen 2 family is not just a lithography shrink, but it represents a significant architectural modernization of the existing 16 nm Kintex platform, infusing it with the high-speed connectivity and memory standards required for the data-heavy workloads of 2027 and beyond.

The original Kintex UltraScale+ series from the Xilinx era is popular in the industry, for a performance/Watt balance that made it the suitable for everything from ultrasound machines to 5G radio heads. However, as edge devices have become hungrier for bandwidth, the original platform's connectivity options began to show their age. The Gen 2 update addresses the I/O modernization need. The most striking upgrade is the memory subsystem, replacing DDR4 with newer LPDDR4X, DDR5 and LPDDR5X memory standards. LPDDR tends to be popular for compact, edge devices in the IIoT (industrial IoT) space. This should even pave the way for a new crop of 8K or 4K broadcast-tier cameras. Since these are produced in small batches, FPGAs are preferred over ASICs or elaborate SoCs.
Memory is a key part of the I/O AMD updated with Kintex UltraScale+, the company also improved the chip's serial interconnects. To begin with, AMD is implementing PCI-Express Gen 4, an important update from the previous-gen chips that were limited to Gen 3, which should improve storage subsystem performance. The company also added dual 100G Ethernet MAC blocks, which should prove crucial in IIoT applications that rely on real-time data aggregation.

AMD is competing not just with Agilex 5 series, but also Lattice Semiconductor's Avant series. Intel's chips are built on the Intel 7 foundry node, while Lattice uses the older TSMC 16 nm FinFET node. AMD, interestingly, is building the Kintex UltraScale+ Gen 2 family on the same exact 16 nm node that the first generation was built on. By sticking with 16 nm for Kintex Gen 2, AMD is betting that customers value platform stability, lower NRE (Non-Recurring Engineering) costs, and supply assurance over the marginal density gains of a newer node. AMD is offering higher performance, a newer I/O, and LUT6 logic compared to LUT4 by Lattice. It also offers a higher maximum frequency (Fmax).

Tech and I/O are one aspect of FPGAs, but security is quite a different thing. You program FPGAs with your IP, but what if someone decides to copy that IP from the FPGA? AMD has integrated CNSA 2.0-grade cryptography, bitstream encryption, and anti-cloning protections directly into the silicon. This allows for secure, authenticated operation in "zero-trust" environments.

AMD is committing to making the series available till 2045, which should make the product line particularly appealing to industries with extremely long product lifecycles, such as aerospace and defense. There are still airplanes in the sky that need navigation data fed to them using 5.25-inch floppy disks. That's how long product lifecycles can be in this segment.

AMD is planning to make Kintex UltraScale+ Gen 2 support available to Vivado and Vitis tools in Q3-2026. The first engineering samples of the chip should reach target audience by Q4-2026. Mass production of the chips should commence in the first half of 2027.
Source: AMD
Add your own comment

8 Comments on AMD Announces Kintex UltraScale+ Gen 2 FPGAs to Capture the Middle of the Market

#1
TumbleGeorge
Why total RAM are in Mb on image with table? Megabits, mebibits? Why no Megabytes or Gigabytes? There is not so small LPDDR4X/5/5X IC. Maybe marketing team TYPO failure?
Posted on Reply
#2
Bwaze
TumbleGeorgeWhy total RAM are in Mb on image with table? Megabits, mebibits? Why no Megabytes or Gigabytes? There is not so small LPDDR4X/5/5X IC. Maybe marketing team TYPO failure?
Ram is expensive nowadays, so they opted for a 4.2 MB to 6.4 MB memory capacity, but written in Megabytes, so it looks more impressive.

:p
Posted on Reply
#3
TheLostSwede
News Editor
TumbleGeorgeWhy total RAM are in Mb on image with table? Megabits, mebibits? Why no Megabytes or Gigabytes? There is not so small LPDDR4X/5/5X IC. Maybe marketing team TYPO failure?
I guess you don't understand FPGAs? That is on chip memory, that is used internally by the FPGA, it's not meant for applications, but rather for soft core to use.
Even AMD's highest-end FPGA's only have a few hundred Megabits of RAM.
www.amd.com/en/products/adaptive-socs-and-fpgas/fpga/virtex-ultrascale-plus.html#product-table
Posted on Reply
#4
Wirko
TumbleGeorgeWhy total RAM are in Mb on image with table? Megabits, mebibits? Why no Megabytes or Gigabytes? There is not so small LPDDR4X/5/5X IC. Maybe marketing team TYPO failure?
RAM on FPGAs looks like an exotic animal if you're only familiar with RAM for CPUs and GPUs. It's not even organised as 8-bit units but is more flexible, for example:
Each block RAM has two write and two read ports. A 36 Kb block RAM can be configured with independent port widths for each of those ports as 32K x 1, 16K x 2, 8K x 4, 4K x 9, 2K x 18 or 1K x 36 (when used as true dual-port (TDP) memory). If only one write and one read port are used, a 36 Kb block RAM can additionally be configured with a port width of 512 x 72 bits (when used as simple dual-port (SDP) memory). An 18 Kb block RAM can be configured with independent port widths for each of those ports as 16K x 1, 8K x 2, 4K x 4, 2K x 9 or 1K x 18 (when used as TDP memory). If only one write and one read port are used, an 18 Kb block RAM can additionally be configured with a port width of 512 x 36 bits (when used as SDP memory).
Apart from that, it can also function as FIFO buffer instead of being addressable memory. That applies to block RAM. An FPGA also has distributed RAM, the one that actually makes it an FPGA, and defines what the logic circuits do.
(source - AMD)

From another point of view, even RAM and flash manufacturers always state the chip capacity in bits, not bytes.
Posted on Reply
#5
r00tcz
As others pointed out, the on chip RAM is for use in buffers and FIFOs between different clock domains etc.
It's all very fast static RAM blocks, similar to cache on CPU. It's also spread all over the chip because many IP blocks require a bit of RAM to work.
For anything else like big buffers, work space or embedded CPU RAM there are the external DRAM ports.
But that's all optional, in many applications you don't need external DRAM at all and that simplifies design and frees the pins for another use as I/Os.

It's specified in megabits because the word width is flexible, it can be arranged in many different ways as noted above, so bits make sense. 8 bit words are only one of many ways how to use it...
Posted on Reply
#6
Wirko
r00tczIt's specified in megabits because the word width is flexible, it can be arranged in many different ways as noted above, so bits make sense. 8 bit words are only one of many ways how to use it...
One thing I don't understand is why we see these unusual word lengths (9/18/36/72 bits). Does that include ECC? Is it some floating point format that DSPs can process?
Posted on Reply
#7
serpretetsky
WirkoOne thing I don't understand is why we see these unusual word lengths (9/18/36/72 bits). Does that include ECC? Is it some floating point format that DSPs can process?
yeah, 1 bit per every 8 bits to fit a 1 parity bit per byte. But you can just use 9 bits however you want. Or 18,36,72.
Posted on Reply
#8
r00tcz
WirkoOne thing I don't understand is why we see these unusual word lengths (9/18/36/72 bits). Does that include ECC? Is it some floating point format that DSPs can process?
The DSP blocks are often using 9 or 18bit precision (at least in Altera world), but in general FPGA logic is very different from any CPU DSPs. In FPGA it's very common to optimize math so only minimum number of bits is required for the job. For example when doing classic FIR filter the samples have one width (as required by other processing) and coefficients have different one (just enough for target noise floor). And ideally you do fixed point math for the performance reasons. These multiplier blocks can be combined in different ways for required number of bits in and out. And ECC is usually not implemented directly, but might be part of the chip logic transparent to the user (for example on space grade hardened FPGAs) and both types of chips require the same bitfile so it's possible to develop and test the design on cheap $10k FPGA before purchasing $200k hardened one for final design.
Posted on Reply
Nov 20th, 2026 16:03 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts