DEV Community

Cover image for Shuttle XPC nano NA20H: Zen 5, 50 TOPS NPU & 96GB DDR5
EmbedSBC
EmbedSBC

Posted on Originally published at embedsbc.com

Shuttle XPC nano NA20H: Zen 5, 50 TOPS NPU & 96GB DDR5

The Shuttle XPC nano NA20H series (NA20H2, NA20H5, and NA20H7) packs up to 50 TOPS of NPU compute, 96GB of DDR5, dual 2.5GbE, and dual 40Gbps USB4 into a 1.04-liter barebone chassis to eliminate I/O and inference bottlenecks at the edge, though its active dual-fan cooling and single M.2 2280 storage slot rule out dusty, sealed environments and internal RAID configurations.

na20h-front-angle

na20h front angle

Systems engineers deploying machine vision, multi-screen digital signage, or edge AI gateways frequently run into a mechanical and electrical dead end: low-power embedded SBCs hit a wall on compute and bus bandwidth, while tower IPCs with discrete accelerator cards simply will not fit inside tight control cabinets. The Shuttle XPC nano NA20H series targets that exact middle ground—fitting current-generation AMD mobile/embedded silicon and a dense array of high-speed interfaces into a 132 × 143 × 55mm (1.04L) enclosure rated for 24/7 continuous operation.

Compute Topology: How Do the NA20H2, NA20H5, and NA20H7 Split General-Purpose and Edge AI Workloads?

The Shuttle XPC nano NA20H series divides workloads cleanly between traditional multi-threaded compute and local AI inference by offering an all-big-core Zen 4 SKU (NA20H2) alongside two hybrid Krackan Point SKUs (NA20H5 and NA20H7).

Across the three models, the underlying silicon topology diverges sharply depending on whether your stack relies on dedicated matrix acceleration:

  • Shuttle NA20H2 (Deterministic Multi-Thread & Graphics): Built around the AMD Ryzen 7 H 255, this SKU uses a homogeneous 8-core / 16-thread Zen 4 layout boosting up to 4.9GHz, paired with an integrated Radeon 780M GPU and no NPU. For industrial SCADA HMI stations or OpenCL/Vulkan rendering pipelines that do not call a neural engine, avoiding asymmetric core scheduling eliminates thread-migration overhead.
  • Shuttle NA20H5 & Shuttle NA20H7 (Heterogeneous Edge AI Inference): Both models step up to AMD’s Krackan Point architecture, combining standard Zen 5 performance cores with high-density Zen 5c efficiency cores and a 50 TOPS NPU. The NA20H5 (Ryzen AI 5 PRO 340) configures 3 Zen 5 + 3 Zen 5c cores (12 threads) up to 4.8GHz with Radeon 840M graphics, while the flagship NA20H7 (Ryzen AI 7 PRO 350) scales to 4 Zen 5 + 4 Zen 5c cores (16 threads) at up to 5.0GHz with a Radeon 860M GPU.

Architectural Trade-Off: Compared to the homogeneous NA20H2, the NA20H5 and NA20H7 integrate Krackan Point silicon and a 50 TOPS NPU to offload quantized local models without starving the CPU or GPU, at the cost of requiring modern OS kernel schedulers to manage asymmetric Zen 5/5c core topologies and stepping down slightly on integrated GPU execution units.

na20h-side-views

na20h side views

Thermal Design: Can Dual Fans Keep a 54W TDP Processor Stable Inside a 1.04-Liter Enclosure?

The Shuttle XPC nano NA20H series relies on an active dual-fan heat-pipe assembly—pairing an 80mm CPU fan with a 60mm chassis fan —to sustain configurable thermal design power (cTDP) envelopes from 35W to 54W during 24/7 operation.

Dissipating up to 54W inside a 1.04-liter volume creates substantial heat flux density. Rather than relying on passive conduction through an extruded aluminum chassis, Shuttle opted for forced-air convection: an 80mm blower/axial fan coupled to heat pipes pulls thermal load directly off the SoC die, while a secondary 60mm chassis fan drives cross-flow ventilation through the enclosure.

Beyond keeping junction temperatures in check when the NA20H7 boosts to 5.0GHz, that secondary airflow path sweeps heat away from the stacked dual-channel DDR5-5600 SO-DIMMs and the PCIe 4.0 x4 NVMe SSD, preventing controller thermal throttling under sustained I/O. Power is supplied via a 19V DC barrel input backed by a bundled 120W external power adapter, leaving adequate transient headroom for peak 54W processor bursts alongside bus-powered USB4 peripherals.

na20h-front-ports

I/O & Bus Bandwidth: Where Are the Physical Bottlenecks for Machine Vision and Multi-Display Setups?

The Shuttle XPC nano NA20H series mitigates peripheral bus contention in multi-camera acquisition and control-room setups by exposing dual 2.5GbE LAN ports, dual 40Gbps USB4 interfaces, and four independent 4K/60Hz display outputs.

In production environments, bus saturation and interface lock-in cause far more deployment headaches than raw clock speeds. The NA20H chassis exposes a dense I/O layout across its front and rear panels:

  1. Segmented High-Speed Networking & Vision Acquisition: Dual Realtek 2.5GbE NICs allow physical network segmentation—dedicating one port to an isolated industrial camera subnet while routing telemetry over the second port to plant MES or cloud infrastructure. Flanking the network ports are two 40Gbps USB4 Type-C interfaces capable of driving external high-speed frame grabbers or PCIe-over-USB4 expansion docks.
  2. Quad-4K Independent Display Matrix: By combining two native HDMI 2.1 ports with two DisplayPort 1.4 alt-mode streams over USB4, the system drives up to four independent 4K displays at 60Hz straight off the integrated graphics engine. Peripheral connectivity is rounded out by five USB 3.2 Gen 2 Type-A ports, a single USB 2.0 port, a 3.5mm audio jack, and an integrated SD card reader.
  3. Memory Capacity vs. Storage Slot Constraints: Dual SO-DIMM slots support up to 96GB of DDR5-5600 memory, providing ample footprint for in-memory edge databases or local LLM weights. However, PCB real estate limits local high-speed storage to a single M.2 2280 slot (PCIe 4.0 x4 NVMe); the adjacent M.2 2230 E-Key slot is strictly keyed for wireless modules (two internal antennas are pre-routed inside the chassis, though no Wi-Fi card is included ).

na20h-rear-ports

OS & Driver Stack: What Should Firmware and Linux Engineers Expect from Barebone Delivery?

Because the Shuttle XPC nano NA20H series ships as an unpopulated barebone system with no preinstalled OS, engineering teams must provision storage, memory, and validated 64-bit Linux or Windows 11 images prior to field deployment.

Validating the software stack early will save integration cycles, particularly around kernel versions and network firmware:

  • Linux Kernel Readiness for Hybrid Cores & NPU: Shuttle officially lists support for 64-bit Linux and Windows 11. If you are deploying Ubuntu, Debian, or a custom Yocto/Buildroot image on the Krackan Point models (NA20H5 / NA20H7 ), ensure your kernel branch includes upstream AMD Zen 5/5c asymmetric scheduler optimizations (AMD P-State) as well as the AMD XDNA driver stack required to expose the 50 TOPS NPU. Conversely, the homogeneous Zen 4 NA20H2 runs cleanly on older, conservative enterprise LTS kernels without asymmetric scheduling dependencies.
  • NIC & Wireless Module Bring-Up: Both 2.5GbE ports rely on Realtek controllers, which are supported in mainline Linux via the r8169 driver; on stripped-down industrial distros, verify that the corresponding linux-firmware blobs are present in your initramfs to avoid link instability under heavy packet rates. Because the M.2 2230 E-Key slot ships empty, you can drop in an Intel Wi-Fi module with known mainline kernel support and snap onto the pre-installed dual antenna leads.

na20h-bottom-vesa-mount

Hardware Specification Matrix

Specification Shuttle NA20H2 Shuttle NA20H5 Shuttle NA20H7
Processor AMD Ryzen 7 H 255 AMD Ryzen AI 5 PRO 340 AMD Ryzen AI 7 PRO 350
Microarchitecture Zen 4 Krackan Point (Zen 5 / 5c) Krackan Point (Zen 5 / 5c)
Cores / Threads 8 Cores / 16 Threads (Homogeneous) 3 Zen 5 + 3 Zen 5c / 12 Threads 4 Zen 5 + 4 Zen 5c / 16 Threads
Max Boost Clock 4.9GHz 4.8GHz 5.0GHz
Integrated GPU Radeon 780M Radeon 840M Radeon 860M
NPU Compute None 50 TOPS 50 TOPS
Memory Up to 96GB DDR5-5600 (2× SO-DIMM) Up to 96GB DDR5-5600 (2× SO-DIMM) Up to 96GB DDR5-5600 (2× SO-DIMM)
Storage 1× M.2 2280 (PCIe 4.0 x4 NVMe) + 1× SD card reader 1× M.2 2280 (PCIe 4.0 x4 NVMe) + 1× SD card reader 1× M.2 2280 (PCIe 4.0 x4 NVMe) + 1× SD card reader
Networking Dual 2.5GbE (Realtek); 1× M.2 2230 E-Key (2× internal antennas, no card) Dual 2.5GbE (Realtek); 1× M.2 2230 E-Key (2× internal antennas, no card) Dual 2.5GbE (Realtek); 1× M.2 2230 E-Key (2× internal antennas, no card)
USB Interfaces 2× USB4 Type-C (40Gbps, DP 1.4 / PD); 5× USB 3.2 Gen 2 Type-A; 1× USB 2.0 2× USB4 Type-C (40Gbps, DP 1.4 / PD); 5× USB 3.2 Gen 2 Type-A; 1× USB 2.0 2× USB4 Type-C (40Gbps, DP 1.4 / PD); 5× USB 3.2 Gen 2 Type-A; 1× USB 2.0
Display Outputs 2× HDMI 2.1 + 2× DP 1.4 (via USB4), up to 4× 4K/60Hz 2× HDMI 2.1 + 2× DP 1.4 (via USB4), up to 4× 4K/60Hz 2× HDMI 2.1 + 2× DP 1.4 (via USB4), up to 4× 4K/60Hz
Power & TDP Configurable 35W–54W; 19V DC input, 120W external adapter Configurable 35W–54W; 19V DC input, 120W external adapter Configurable 35W–54W; 19V DC input, 120W external adapter
Cooling & Chassis Active dual-fan heat-pipe (80mm CPU + 60mm chassis); 132 × 143 × 55mm (1.04L) Active dual-fan heat-pipe (80mm CPU + 60mm chassis); 132 × 143 × 55mm (1.04L) Active dual-fan heat-pipe (80mm CPU + 60mm chassis); 132 × 143 × 55mm (1.04L)
OS & Duty Cycle 64-bit Linux / Windows 11 (not preinstalled, barebone); VESA mount; 24/7 rated 64-bit Linux / Windows 11 (not preinstalled, barebone); VESA mount; 24/7 rated 64-bit Linux / Windows 11 (not preinstalled, barebone); VESA mount; 24/7 rated

Deployment Boundaries: Where It Fits vs. Where It Breaks

  • Green-Light Deployment Scenarios:
    • Space-Constrained Quad-4K Control & Signage Nodes: With a 1.04L footprint, bundled VESA mount, and native four-display 4K/60Hz support, units mount cleanly behind large-format operator displays.
    • Edge Vision & Local AI Inference Gateways: Pairing the NA20H5 or NA20H7’s 50 TOPS NPU with 96GB DDR5 and dual 2.5GbE + dual 40Gbps USB4 handles multi-stream camera ingestion alongside local model execution.
    • Legacy LTS Linux Multi-Threaded Control Workloads: When NPU acceleration is unnecessary, the all-big-core Zen 4 NA20H2 avoids hybrid-core scheduler tuning on older enterprise Linux kernels.
  • Hard Constraints to Avoid:
    • High-Particulate, Oil-Mist, or Sealed Fanless Enclosures: Thermal dissipation relies on 80mm + 60mm active fans to manage 35–54W loads; airborne particulates will foul the heatsink fins and bearings.
    • Internal Dual-Drive Hardware RAID: The board provides only one M.2 2280 NVMe slot (the second M.2 slot is 2230 E-Key for wireless only). Fault-tolerant storage requires external USB4 arrays or network-attached storage.
    • Zero-Touch Turnkey Field Drop-Ins: Because units ship as barebones without RAM, SSD, Wi-Fi cards, or an OS, system integrators must budget bench time for assembly and image provisioning.

Engineering FAQ

*Q: Can the Shuttle XPC nano NA20H series house two M.2 NVMe SSDs internally for RAID 1 mirroring?*

A: No. The motherboard includes only one M.2 2280 slot supporting PCIe 4.0 x4 NVMe; the second M.2 2230 E-Key slot is dedicated exclusively to wireless modules.

*Q: Do NA20H barebone units include a pre-installed Wi-Fi card or operating system?*

A: Neither is included. Units ship as barebones without RAM, storage, an OS, or a wireless card, though two internal wireless antennas are pre-installed.

*Q: Which physical ports are required to drive four simultaneous 4K displays on the NA20H series?*

A: You must utilize both native HDMI 2.1 ports alongside both USB4 Type-C ports (via DisplayPort 1.4 alt-mode) to output four independent 4K/60Hz streams.

Availability & Ordering

The NA20H2, NA20H5, and NA20H7 are currently listed on the Shuttle Europe official website. Because the lineup ships in a barebone configuration, Shuttle has not published fixed MSRP pricing; exact part configurations, regional availability, and final unit costs depend on geographic market and components selected through Shuttle’s online configurator.


Originally published on EmbedSBC — embedded systems, SBC and edge-AI hardware analysis.

Top comments (0)