AI Wiki

AMD Kria

AI HardwareRobotics
17 min read
Updated
Suggest editHistoryTalk
RawGraph

Last edited

Fact-checked

Sources

23 citations

Revision

v2 · 3,336 words

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

AMD Kria is a family of embedded system-on-modules (SOMs) and matching carrier boards sold for machine vision, robotics, motor control, and industrial automation. Xilinx launched the line on 2021-04-20, about ten months before AMD closed its acquisition of the company, and the original modules were built around custom Zynq UltraScale+ MPSoC devices that put Arm CPU cores and FPGA fabric on a single die [1][2]. At Advancing AI 2026 in San Francisco (2026-07-22 to 2026-07-23) [19], AMD extended the brand onto x86 with the Kria AI SOM, a COM-HPC module based on the Ryzen AI Embedded X100 Series, and the Kria AI Robotics Developer Platform, which pairs that module with a carrier card carrying a Spartan UltraScale+ FPGA [3][4][5]. AMD calls the developer platform "the industry's first open, fully integrated platform for autonomous robotics" [4]. The developer platform was sampling with early-access customers at announcement and is targeted for general availability in Q4 2026; the Kria AI SOM itself is expected from ODM partners in the same quarter [3][6].

What a SOM is, and the problem Kria was built to solve

A system-on-module is a small board carrying the processor, memory, storage, power management, and boot firmware, exposed through one or two dense connectors. It is one of the standard ways to get edge AI compute into a shipping product, alongside single-board computers and full chip-down designs. The customer designs only the carrier board underneath it: connectors, sensors, power input, mechanical mounting. That splits the hard part of an embedded design, high-speed DDR routing and power sequencing on an advanced SoC, away from the part that differs between products.

For FPGA-based designs the split matters more than usual. Chip-down designs around a Zynq UltraScale+ device require board-level signal integrity work plus FPGA tooling expertise, and AMD has claimed the Kria approach typically saves up to nine months of time to deployment compared with designing the silicon onto a custom board [7]. That was the original pitch: buy a production-qualified module, skip the board bring-up, and deploy the same part in volume rather than respinning from an evaluation kit.

The second half of the pitch is lifecycle. Industrial customers buy on availability windows, not launch-day performance. The K26 shipped with a stated 10-year product availability and, for the industrial grade, a 10-year expected product lifetime and a temperature range of -40 C to 100 C [2].

The Kria family

Module or kitSiliconAnnouncedTarget use
Kria K26 SOMCustom Zynq UltraScale+ MPSoC (XCK26): quad-core Arm Cortex-A53, dual-core Cortex-R5F, Mali-400MP2, 256K system logic cells, 4 GB DDR42021-04-20Vision AI, smart city, machine vision, vision-guided robotics [1][2]
KV260 Vision AI Starter KitK26 SOM plus vision carrier card2021-04-20, $199 at launchEvaluation and prototyping for camera pipelines [1]
KR260 Robotics Starter KitK26 SOM plus robotics carrier card2022-05-17, $349ROS 2 robotics, industrial communication and control [7][8]
Kria K24 SOMCost-optimized custom Zynq UltraScale+ MPSoC, ECC LPDDR4 on the industrial part2023-09-19; AMD MSRP $250 commercial, $350 industrialMotor control and DSP-heavy edge workloads [9][21]
KD240 Drives Starter KitK24 SOM plus motor-control carrier card2023-09-19, under $400Drives, motor control, DSP [9]
Kria AI SOMAMD Ryzen AI Embedded X100 Series (16-core Zen 5, RDNA 3.5 iGPU, XDNA 2 NPU), 64 GB or 128 GB LPDDR5X, COM-HPC form factor2026-07-23Physical AI, autonomous machines, humanoids [3][4]
Kria AI Robotics Developer PlatformKria AI SOM plus Spartan UltraScale+ FPGA carrier card and thermal enclosure2026-07-23Robotics prototyping through production [5][6]

The Zynq generation, 2021 to 2023

The K26 is the module the family was built around, and AMD documents it as a production-ready hardware platform shipped without preloaded firmware in non-volatile memory [20]. Its custom XCK26 device pairs a 64-bit quad-core Cortex-A53 application processor with a 32-bit dual-core Cortex-R5F real-time processor and 256K system logic cells of programmable fabric, plus an integrated H.264/H.265 codec that handles up to 32 streams at a combined 4Kp60. Xilinx rated it at up to 1.4 TOPS of AI processing with a deep learning processing unit configured for INT8 and upgradeable to INT4, and at 245 I/Os; the product brief lists 11 four-lane MIPI or sub-LVDS camera interfaces plus one four-lane SLVS-EC interface, and claims support for up to 15 cameras across multiple interfaces. The board measures 77 x 60 x 11 mm with dual 240-pin connectors and includes a TPM 2.0 module for IEC 62443 security [2]. Commercial parts launched at $250 and industrial parts at $350, with the $199 KV260 starter kit as the entry point [1][10]. Prices have moved since: AMD lists the commercial K26 at $325 and the KV260 at $249 today [21]. AMD refreshed the KV260 on 2025-06-26, adding auto-focus through the IAS camera interface, and described it as its most popular Kria starter kit, or in the body text the highest-selling development platform built on any Zynq UltraScale+ MPSoC [11].

The KR260, announced 2022-05-17, moved the same module onto a robotics carrier with native ROS 2 support and the Kria Robotics Stack (KRS), a set of libraries that push ROS 2 nodes into the FPGA fabric [7]. Canonical certified Ubuntu images for the platform [8].

The K24, announced 2023-09-19, took the family down-market. AMD described it as half the size of a credit card, drawing roughly half the power of the K26 while remaining connector-compatible with it, and aimed it at motor control for the industrial robot market, surgical robotics, elevators, and EV charging [9][12].

Kria AI: the 2026 generation

The Kria AI SOM breaks with the Zynq lineage. It carries a Ryzen AI Embedded X100 Series processor: up to 16 Zen 5 CPU cores, an RDNA 3.5 integrated GPU that AMD rates at 60 FP16 TFLOPS, an XDNA 2 NPU rated up to 50 TOPS, and 64 GB or 128 GB of LPDDR5X shared as unified memory across all three engines [3][5]. AMD says the module can scale to 234 concurrent agents and make up to 8,000 control decisions per second; the newsroom release phrases the same figure as "more than 8,000 control decisions every second" [3][4].

Two details are easy to get wrong. First, there is no FPGA on the Kria AI SOM itself. AMD's own specification for the developer platform describes it as a "SOM based on Ryzen AI Embedded X199 processor (integrated CPU, GPU, NPU) and robotics carrier card with Spartan UltraScale+ FPGA", and CNX Software noted the same break with earlier Kria modules [5][6]. AMD's own materials are careful about this: the Advancing AI press release attributes the combination of "CPU, GPU, NPU and FPGA compute" to the Kria AI Robotics Developer Platform rather than to the SOM, and the Kria AI product page describes the SOM as combining CPU, GPU, NPU and unified memory with no FPGA in the list [4][13]. Coverage that credits the module with all four engines has collapsed the distinction. Second, the module is not sold by AMD directly in the way the K26 was: AMD lists the Kria AI SOM as available through partners, with Arbor, Congatec, iBase, IEI, Sapphire, and Seavo named as launch partners for X100-based modules [4][14].

AMD adopted the COM-HPC open standard rather than a proprietary connector, which it presented as an anti-lock-in decision. "AMD can build it, and any other company on the planet can adopt it and build it," said KV Thanjuvar Bhaaskar, robotics lead and senior manager at AMD [6].

Developer platform I/O

The Kria AI Robotics Developer Platform ships as SOM plus carrier card plus thermal enclosure [5].

Interface classProvided
Camera1x FAKRA x4 (GMSL2/3), 1x FAKRA x4 (GMSL1/2), so up to 8 inputs
Networking2x 1/2.5/5 GbE RJ45; 2x 1 GbE RJ45 for EtherCAT/TSN; 1x 10 GbE (x4) QSFP
Fieldbus and serial2x CAN-FD, 1x RS485
USB2x USB4 Type-C, 3x USB 3.2 Type-C, 2x USB 2.0 Type-A
Display2x Mini DisplayPort
Audio3.5 mm jack, A2B (Automotive Audio Bus)
ExpansionOculink PCIe, M.2 2280 Key M (NVMe), M.2 2230 Key E (Wi-Fi/Bluetooth), Pmod
SensorsInertial measurement unit on board

AMD has not published the specific Spartan UltraScale+ part number on the carrier [5][6]. The Robot Report reported that open-source baseboard schematics accompany the platform, which would let integrators fork the carrier design for production [14].

Why FPGA fabric sits next to CPU, GPU, and NPU

An FPGA is a fabric of lookup tables, flip-flops, DSP blocks, and I/O cells that a designer wires into a fixed circuit after the chip is manufactured. A CPU or GPU fetches and executes instructions; an FPGA implements the computation as hardware, so a pipeline built in fabric emits a result every clock cycle with timing determined at synthesis rather than at run time. Three jobs in a robot exploit that property.

The first is hardware I/O timing. EtherCAT and time-sensitive networking need synchronization inside a microsecond, and motor commutation loops need jitter bounded in single-digit microseconds. On a general-purpose core running Linux the same work is exposed to interrupt latency, scheduler jitter, cache contention, and power-state transitions. AMD's own figures show the size of the problem it is engineering around: with BIOS and Linux tuning it targets interrupt latency under 7 microseconds at six-nines, and for hard real-time it recommends a Xen hypervisor with a FreeRTOS virtual machine, cache coloring, and VM isolation [14]. Fabric logic sidesteps that stack entirely.

The second is sensor fusion and preprocessing at the wire. Deserializing GMSL camera links, correcting and rectifying frames for computer vision, and timestamping lidar and IMU samples into a common clock can all happen in fabric before anything reaches the CPU. That removes memory copies and, more importantly, keeps timestamps honest, since a timestamp applied in hardware at arrival does not drift with scheduler latency. AMD describes the carrier FPGA's role as "real-time I/O, sensor fusion, and safety features" [14].

The third is safety. Watchdog and interlock logic implemented in fabric is physically independent of the SoC running Linux, so an application that hangs cannot leave actuators energized. AMD has pointed to this heritage explicitly, noting that its adaptive computing parts have run "the sensing, safety and real-time control systems at the heart of industrial and surgical robots" for decades [3].

The complementary claim is about the memory hierarchy rather than the fabric. Perception, fusion, and planning traditionally shuttle tensors between CPU and discrete GPU address spaces; a unified memory pool shared by CPU, GPU, and NPU removes those copies. That is the architectural argument AMD makes for the X100 over a CPU-plus-discrete-GPU robot design [14].

Real-time and performance claims, and how they were measured

Every performance figure published for the Kria AI SOM as of 2026-07-27 originates with AMD or with a study AMD commissioned. No independent hands-on review or third-party benchmark of shipping hardware had appeared, which is unsurprising given that the platform was still sampling with early-access customers.

ClaimSource and method
Up to 3.4x better real-time results and 1.6x more free CPU capacity vs NVIDIA Jetson ThorOpenNav Robotics Workload Benchmark, commissioned by AMD, published by Open Navigation LLC on 2026-07-23. Measured on a GMKtec EVO-X2 mini PC with a Ryzen AI Max+ 395 configured to approximate X199 specifications, not on Kria AI hardware, against a Jetson AGX Thor Developer Kit [4]
Up to 2.3x more concurrent agents, up to 234 agentsmimik whitepaper "Architectural Fit for Production-Scale Agentic AI on Heterogeneous SoCs", commissioned by AMD, published 2026-07-23. A modeled sweep of 455 agentic workflows across two device classes rather than a hardware measurement [4]
8,000 control decisions per second (125 microsecond loop)AMD measurement of CPU real-time control loop latency running a Bosch Rexroth controller on X100 silicon [4]
Sub-100 ms vision-language-action reasoningembedL benchmarking of GPU VLA inference latency using a Pi0.5 model [4]
Up to 3x higher peak FP32 vs Jetson T5000AMD, framed for signal processing. It is a peak arithmetic rate and AMD publishes no endnote for it. The medical ultrasound beamforming result AMD announces in the same sentence is a different claim, an average 1.7x against a discrete NVIDIA RTX 4000 Ada, not against a Jetson [6][15]
1.5 microsecond control-loop closureAttributed to AMD by The Robot Report as an architectural figure for the SOM; it measures something different from the 125 microsecond controller loop above [14]

The proxy-hardware caveat on the headline 3.4x figure comes from AMD's own footnote, not from a critic [4]. It is a normal practice for pre-production silicon, but it does mean the most-quoted comparison against NVIDIA was not run on a Kria AI SOM.

Software

The Zynq-era Kria stack centers on AMD's FPGA tooling: Vivado and the Vitis software platform, Vitis AI for quantizing and deploying neural networks to the on-chip deep learning processing unit, Yocto-based PetaLinux, certified Ubuntu images, and PYNQ for driving the fabric from Python and Jupyter [1][12][16]. The Kria App Store distributes pre-built accelerated applications as Docker containers, so a software developer can run a working vision or motor-control pipeline without writing RTL [1][16].

The 2026 generation runs a different stack. The AMD Robotics Software Suite is built on AMD ROCm and ROS 2 and includes a Robotics Core SDK, a Physical AI SDK, accelerated ROS nodes, MoveIt 2 support, and Nav2 acceleration, with PyTorch, ONNX, and TensorFlow supported as frameworks [3][14][23]. AMD describes the suite as an open, portable software stack, releases open-source baseboard schematics and FPGA design files with the developer platform, and ships CUDA-to-ROCm migration tooling, claiming an average of 75 percent code preservation on existing CUDA sources [3][14][23]. Because the carrier FPGA is a discrete Spartan UltraScale+ device rather than fabric fused to the application processor, FPGA work on the new platform still runs through the Vivado and Vitis flow, separate from the ROCm path.

Competitive position

Kria AI is aimed squarely at NVIDIA Jetson, and specifically at Jetson Thor. AMD's arguments are longevity, determinism, and openness rather than raw inference throughput: 24/7 operation for up to 10 years and a -40 C to 105 C range on industrial X100 SKUs, against the five years of continuous operation and -25 C to 80 C that AMD attributes to the Jetson T5000; an open COM-HPC module standard instead of a proprietary form factor; and an FPGA on the carrier for deterministic I/O [6][14][15]. The temperature pair is not measured the same way, and AMD says so in its own blog: the -40 C to 105 C figure is a junction temperature range for select industrial X100 SKUs, while the -25 C to 80 C figure is the maximum operating range NVIDIA specifies at the T5000's thermal transfer plate. Junction always runs hotter than the plate that cools it, so the two numbers are not directly comparable [22].

The counterweights are real. NVIDIA has a much larger installed base in robotics, a mature CUDA ecosystem, and its own robotics stack in Isaac and GR00T. AMD's own inference-per-watt case is narrower than its signal-processing case, and Rob Bauer of AMD framed the FP32 advantage in aerospace and defense terms while conceding that NVIDIA parts are "pretty highly optimized around inferencing" [14]. The superlative in AMD's marketing, that this is the first open, fully integrated autonomous robotics platform, is AMD's own framing and should be read as such: integrated developer platforms for robots already exist from NVIDIA, and open ROS 2 stacks on commodity x86 have been shipping for years [4][13].

Adoption and partners

For the Zynq generation, AMD's published examples include Solectrix, whose SXVPU industrial vision box uses a Kria SOM to process four GMSL2 camera streams in parallel for mobile machinery [17].

For Kria AI, AMD named two early-access customers at launch: Castec International, which is building an autonomous mobile robot for semiconductor fabs, and Foundation Robotics, a humanoid developer that The Robot Report says is migrating from Intel and NVIDIA to the X100 and plans FPGA-based hand control [14]. Alongside the hardware, AMD launched the Robotics Partner Network, an ecosystem program spanning ODMs, ISVs, simulation and sensor vendors, and AI model providers; Analog Devices, SICK, and Voyant Photonics were named on the sensor side, and Open Robotics, Open Navigation, and OpenCV on the software side [14][18]. Robot OEMs are treated as customers of the network rather than members [18]. "No single company will build the future of physical AI," said Amey Deosthali, AMD's senior director for industrial, robotics, and healthcare [18].

Availability and pricing

X100 processors began customer sampling in June 2026 with production expected in Q4 2026 [15]. Kria AI SOMs are expected from ODM partners in Q4 2026, and the Kria AI Robotics Developer Platform was sampling with early-access customers at announcement with general availability also targeted for Q4 2026 [3][6].

AMD has not announced pricing for either product; the AMD product page carries a "Notify Me" button rather than a price [5]. CNX Software estimated above $5,000 for the developer platform on the basis that it competes with the Jetson AGX Thor Developer Kit, which is a reporter's expectation rather than a disclosed figure [6]. CNX Software also reported that Sapphire will sell a version of the same design as the EDGE+ Apex SOM/Carrier Robotics Platform, though no product page existed at the time of writing [6].

References

  1. "Xilinx Introduces Kria Portfolio of Adaptive System-on-Modules for Accelerating Innovation and AI Applications at the Edge." Business Wire, 2021-04-20. https://markets.financialcontent.com/stocks/article/bizwire-2021-4-20-xilinx-introduces-kria-portfolio-of-adaptive-system-on-modules-for-accelerating-innovation-and-ai-applications-at-the-edge ↩
  2. "Kria K26 System-on-Module Product Brief." Xilinx/AMD. https://www.xilinx.com/publications/product-briefs/xilinx-k26-product-brief.pdf ↩
  3. "AAI 2026: AMD Introduces Open, Turnkey Integrated Platform for Physical AI." AMD Newsroom, 2026-07-23. https://newsroom.amd.com/news/aai-2026-kria-robotics-dev-platform/ ↩
  4. "AMD Kria AI Solutions." AMD. https://www.amd.com/en/products/system-on-modules/kria/ai.html ↩
  5. "AMD Kria AI Robotics Developer Platform." AMD. https://www.amd.com/en/products/system-on-modules/kria/ai/robotics-developer-platform.html ↩
  6. Aufranc, Jean-Luc. "AMD launches Ryzen AI Embedded X100 processors, Kria AI SoM, and physical AI/robotics developer platform." CNX Software, 2026-07-24. https://www.cnx-software.com/2026/07/24/amd-launches-ryzen-ai-embedded-x100-processors-kria-ai-som-and-physical-ai-robotics-developer-platform/ ↩
  7. "AMD Robotics Starter Kit Kick-Starts the Intelligent Factory of the Future." AMD Investor Relations, 2022-05-17. https://ir.amd.com/news-events/press-releases/detail/1067/amd-robotics-starter-kit-kick-starts-the-intelligent ↩
  8. "Kria KR260, a scalable robotics platform powered with Ubuntu." Canonical, 2022. https://canonical.com/blog/kria-kr260-a-scalable-robotics-platform-powered-with-ubuntu ↩
  9. "AMD Accelerates Innovation at the Edge with Kria K24 SOM and Starter Kit for Industrial and Commercial Applications." GlobeNewswire, 2023-09-19. https://www.globenewswire.com/news-release/2023/09/19/2745717/0/en/AMD-Accelerates-Innovation-at-the-Edge-with-Kria-K24-SOM-and-Starter-Kit-for-Industrial-and-Commercial-Applications.html ↩
  10. Aufranc, Jean-Luc. "Xilinx Kria K26 SoM and vision AI devkit based on Zynq UltraScale+ XCK26 FPGA MPSoC." CNX Software, 2021-04-28. https://www.cnx-software.com/2021/04/28/xilinx-kria-k26-som-vision-ai-devkit-zynq-ultrascale-xck26-fpga-mpsoc/ ↩
  11. "AMD Refreshes Its Most Popular Kria Starter Kit with Sharper Vision for Edge AI." AMD Blogs, 2025. https://www.amd.com/en/blogs/2025/kria-kv260-vision-ai-starter-kit-refresh.html ↩
  12. "AMD Kria System-on-Modules." AMD. https://www.amd.com/en/products/system-on-modules/kria.html ↩
  13. "AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era." AMD Investor Relations, 2026-07-23. https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era ↩
  14. Oitzman, Mike. "AMD unveils Kria module for real-time control, unified memory for robots." The Robot Report, 2026-07-23. https://www.therobotreport.com/amd-unveils-kria-module-real-time-control-unified-memory-robots/ ↩
  15. "AAI 2026: AMD Delivers Leadership Heterogeneous Compute for Physical AI." AMD Newsroom, 2026-07-23. https://newsroom.amd.com/news/aai-2026-ryzen-ai-embedded-x100/ ↩
  16. "AMD App Store for Kria System-on-Modules." AMD Developer Resources. https://www.amd.com/en/developer/resources/kria-apps.html ↩
  17. "Solectrix Case Study." AMD. https://www.amd.com/en/resources/case-studies/solectrix.html ↩
  18. "AAI 2026: New AMD Open Robotics Partner Network for Physical AI Development." AMD Newsroom, 2026-07-23. https://newsroom.amd.com/news/aai-2026-robotics-partner-network/ ↩
  19. "AMD Advancing AI 2026, San Francisco, July 22-23." AMD. https://www.amd.com/en/corporate/events/advancing-ai.html ↩
  20. "Kria K26 SOM." AMD Adaptive Computing Wiki. https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/1641152513/Kria+K26+SOM ↩
  21. AMD Kria SOM product pages, MSRP as listed on 2026-07-27: K24 commercial $250, K24 industrial $350, K26 commercial $325, KV260 Vision AI Starter Kit $249, KR260 Robotics Starter Kit $349, KD240 Drives Starter Kit $399. https://www.amd.com/en/products/system-on-modules/kria/k24/k24c-commercial.html and https://www.amd.com/en/products/system-on-modules/kria/k26.html ↩
  22. Quenton Hall. "From Benchmarks to Behavior: Rethinking Performance in Autonomous Robotics." AMD Blogs, 2026-07-23. https://www.amd.com/en/blogs/2026/from-benchmarks-to-behavior-rethinking-performance-in-a.html ↩
  23. "AMD Advancing AI 2026: Ryzen AI Embedded X100, Kria AI Robotics Platform, and Robotics Partner Network." TechPowerUp, July 2026. https://www.techpowerup.com/351008/amd-advancing-ai-2026-ryzen-ai-embedded-x100-kria-ai-robotics-platform-and-robotics-partner-network ↩

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

1 revision by 1 contributors · full history

Suggest edit