Free to read. Sign up to save your progress and take knowledge-check quizzes.

Sign up free
5 min read·Updated May 4, 2026

AMD Ryzen AI

AMD logoBy AMD

AMD Ryzen AI brings on-device AI to laptops and mini-desktops via the XDNA 2 NPU. The flagship Ryzen AI Max+ 395 (Strix Halo) delivers 50 TOPS NPU, 126 TOPS system-wide, and up to 128 GB unified memory — enough to run open-weight LLMs up to roughly 70 billion parameters locally.

Listen to this lesson

Free preview · first 0:30
0:00 / 0:30

Audio & video lessons are paid features

Plus unlocks audio streaming. Pro adds downloadable audio, video, certificates, and more.

Plus adds:
  • Audio streaming
  • Downloadable PDFs
  • All AI Playbooks
  • Personalized content
Pro also adds:
  • Certificates of completion
  • Audio MP3 downloads
  • Video lessonssoon
  • & More…soon

Watch this lesson

Video coming soon

Learning Objectives

  • Understand what Ryzen AI is and how it competes for client-side AI workloads against Intel Core Ultra and Apple Silicon
  • Identify the flagship Ryzen AI Max+ 395 specs and what its 128 GB unified memory ceiling enables for local LLM inference
  • Recognize the Copilot+ PC certification and what the 50 TOPS NPU floor means in practice

What Is AMD Ryzen AI?

Ryzen AI is AMD's on-device AI platform for laptops, mini-desktops, and increasingly small-form-factor workstations. It pairs a Zen 5 CPU, an RDNA 3+ integrated GPU, and a dedicated XDNA 2 Neural Processing Unit (NPU) on a single package — the same three-engine pattern Intel Core Ultra and Apple Silicon use, but with a unified-memory ceiling that no other client-AI vendor matches in 2026.

The headline product is the Ryzen AI Max+ 395 — codenamed Strix Halo, shipped January 2025. The recently announced Ryzen AI 400 series (CES 2026, January 5) extends the lineup but ships in volume later in 2026.

💡Key Concept

Why an NPU? Modern AI client workloads — Copilot summarization, Recall search, video-call effects, on-device LLM inference — run continuously in the background. CPUs handle them with high power draw and battery cost; GPUs are faster but throttle laptop thermals. NPUs are a third silicon block tuned for sustained low-power AI inference: lower TOPS-per-watt cost than CPU or GPU, and they run alongside the other engines without competing for thermal budget.

Ryzen AI Max+ 395 — Headline Specs

The Ryzen AI Max+ 395 is the highest-end client AI chip AMD ships today and the centerpiece of the platform's local-LLM pitch.

SpecRyzen AI Max+ 395
ArchitectureZen 5 CPU + RDNA 3+ iGPU + XDNA 2 NPU
CPU cores16 Zen 5 cores
Integrated GPURadeon 8060S (40 RDNA 3+ Compute Units)
NPU TOPS50 TOPS (XDNA 2)
Total system TOPS~126 TOPS (NPU + iGPU + CPU)
MemoryUp to 128 GB LPDDR5X unified memory
Max local LLMUp to roughly 70 billion parameters
Copilot+ PCCertified — exceeds Microsoft's 40 TOPS NPU floor
LaunchJanuary 2025

The 128 GB unified-memory ceiling is the differentiator. Apple Silicon M4 Max maxes out at 128 GB (M4 Ultra higher); Intel Core Ultra client SKUs typically cap at 96 GB. For developers running open-weight models locally, 128 GB shared between CPU, iGPU, and NPU is the difference between running a 70-billion-parameter model and not.

Copilot+ PC Certification

In June 2024 Microsoft introduced Copilot+ PC as a hardware certification for Windows laptops capable of running on-device AI features (Recall, Live Captions, Cocreator, Windows Studio Effects). The threshold is 40 TOPS of NPU performance. Ryzen AI Max+ 395 hits 50 TOPS on the NPU alone and clears the bar comfortably.

Copilot+ certification matters less for the marketing badge and more because it gates Windows-side AI features. Non-certified laptops do not get Recall or the full Studio Effects suite at all.

Notable 2025-2026 Deployments

SystemForm factorNotable angle
Beelink GTR9 ProMini-PCFirst non-laptop Strix Halo deployment for home labs
ASUS ROG Flow Z13ConvertibleGaming + AI dev workload focus
HP ZBook Ultra G1aMobile workstationEnterprise CAD + on-device AI
Framework DesktopModular mini-PCDeveloper-friendly, repairable form factor

The mini-PC and modular-desktop deployments are quietly the strategic story here — AI developers who want to run open-weight LLMs locally without a discrete GPU farm now have a sub-1-liter desktop option that runs 70-billion-parameter models out of the box.

⚠️Warning

Ryzen AI 400 series details may have moved. AMD announced the Ryzen AI 400 series at CES 2026 in January, but SKU-level TOPS ratings, memory ceilings, and ship dates have continued to be revised through Q1 2026. The Ryzen AI Max+ 395 numbers above are verified shipping specs; treat 400-series claims as roadmap until production silicon is independently benchmarked.

Pricing

Ryzen AI Max+ 395 (laptop)Retail
  • Built into Copilot+ certified laptops
  • Typical street: 1,500-2,500 dollars
  • Includes Windows + AI features
Strix Halo mini-PC (Beelink, Framework)Retail
  • Bare-bone or barebones-plus-RAM
  • Typical street: 1,300-2,200 dollars
  • Add Windows or Linux
Mobile workstation (HP ZBook Ultra G1a)Enterprise
  • Volume procurement via OEM
  • Custom pricing
  • 3-year warranty + AMD Pro features

Strengths

  • 128 GB unified memory ceiling — Highest in the client-AI segment in 2026; lets developers run open-weight LLMs up to roughly 70 billion parameters locally without a discrete GPU
  • Copilot+ PC certified at 50 TOPS NPU — Above the 40 TOPS threshold; full access to Microsoft on-device AI features
  • Three-engine architecture — CPU, iGPU, and NPU on a single package; AI workloads can be scheduled to whichever engine matches the workload best
  • Mini-PC and modular-desktop options — Beelink, Framework, and ASUS ship form factors below 2 liters, matching what Apple Mac mini does for the M-series ecosystem
  • Open ROCm path — Same ROCm 7.x stack that runs on Instinct also runs (with framework-specific patches) on Strix Halo's RDNA 3+ iGPU — a unified developer experience from laptop to datacenter

Limitations and Considerations

  • NPU software ecosystem is still maturing — XDNA 2 is supported by Microsoft's DirectML and AMD's own Ryzen AI software, but the catalog of NPU-optimized models is smaller than what Apple's Core ML or NVIDIA's TensorRT serve up
  • No discrete-GPU performance — Strix Halo is a unified-memory client part. For training workloads or batch inference at production scale, an Instinct datacenter accelerator or a discrete RTX-class consumer GPU still wins
  • Memory bandwidth ceiling — LPDDR5X unified memory tops out around 256 GB per second of bandwidth; HBM-equipped datacenter GPUs ship 8 TB per second. Local LLM inference is bandwidth-bound, so token-per-second throughput is meaningfully below Instinct or consumer RTX cards
  • Premium pricing for the headline SKU — Sub-1,500-dollar laptops typically use lower-tier Ryzen AI 9 HX models with smaller memory ceilings; the full 128 GB Max+ 395 experience starts at roughly 2,000 dollars
  • Ryzen AI 400 series specs not yet stable — CES 2026 announcement specs have continued to revise; production SKUs land later in 2026

Key Takeaways

  • AMD Ryzen AI is the on-device AI platform behind Copilot+ certified laptops and a growing set of mini-PCs and modular desktops; the headline product is the Ryzen AI Max+ 395 (Strix Halo), shipped January 2025
  • Strix Halo combines 16 Zen 5 cores, a 40-Compute-Unit RDNA 3+ iGPU, and a 50 TOPS XDNA 2 NPU — totaling roughly 126 TOPS system-wide and clearing Microsoft's 40 TOPS Copilot+ threshold
  • The 128 GB unified-memory ceiling is the platform's defining advantage; it lets developers run open-weight LLMs up to roughly 70 billion parameters locally without a discrete GPU
  • The Ryzen AI 400 series, announced at CES 2026, extends the lineup but ships in volume later in 2026; verified shipping specs today are the Max+ 395 numbers above

Save your progress & take the quiz

Sign up free to bookmark lessons, track which modules you've completed, and lock in what you learned with a quick knowledge-check quiz at the end of each lesson.

🧭Recommended for you