SECTIONS
EDITIONS
NETWORK
AIXX Pedia The encyclopedia of compute hardware

AIXX Pedia › Model Entry › AMD Ryzen AI Max+ 395

AMD Ryzen AI Halo Development Platform (official image via ServeTheHome)
Vendor render · AIXX Pedia (AMD official blog, platform demo image)

AMD Ryzen AI Max+ 395

Strix Halo · 16-core Zen 5 · 40CU iGPU · Up to 128GB unified memory · Model Entry

Overview OVERVIEW

AMD Ryzen AI Max+ 395 (Codename Strix Halo ) is AMD's flagship AI APU announced at CES on as a January 6, 2025 (edge SoC): 16-core Zen 5 (32 threads, 3.0–5.1GHz) + Radeon 8060S iGPU (40 RDNA 3.5 compute units) + XDNA 2 NPU 50 TOPS three compute engines co-packaged, with up to 128GB LPDDR5X-8000 quad-channel unified memory (256GB/s, 256-bit) — CPU/GPU/NPU share a single memory pool, allowing hundred-billion-parameter-class models to be fully resident on-device (per AMD's official statements and CES announcements[1]).

It is the flagship representative of the "on-device LLM" category: of the 128GB unified memory, up to 96GB can be allocated to the GPU — community tests show it can fully keep GPT-OSS 120B-class models resident (per AMD's official blog demo[2]); the headline three-engine benchmark claim is "up to 68% faster than an RTX 4070 laptop GPU" (1080p, AMD marketing[1]). TDP is 45W, configurable up to 120W — the same chip powers everything from thin-and-light laptops to desktop mini PCs (official figures[1]).

Three things buyers should know upfront : ① The LPDDR5X memory is soldered and non-expandable (choosing 128GB or 64GB is a one-time decision); ② The iGPU has no dedicated VRAM, and its 256GB/s unified memory bandwidth is an order of magnitude below discrete GPUs' GDDR7 (~1.8TB/s) — don't benchmark throughput-sensitive workloads against data center cards; ③ It is a consumer/workstation-class chip with no server-grade warranty or full ECC suite — for mass deployment, choose OEMs based on overall system reputation.

Use Cases FIT

AIXX Index (0–100): estimated by this library using a weighted average of public specs—memory 40% + interconnect 30% + ecosystem 30%, with weights adjusted by use case. The methodology is transparent and verifiable, and does not constitute a performance guarantee.

On-device LLM inference 88AIXX Index
128GB unified memory fully resident with GPT-OSS 120B-class (AMD official demo[2]): a private inference/Agent host where data never leaves the machine, with desktop-class power costs.
Light fine-tuning/LoRA 72AIXX Index
The large memory fits training sets and small-model LoRA — throughput is constrained by the 256GB/s bandwidth, so training efficiency is not its strength (physical bandwidth limit).
Content creation/gaming (iGPU) 82AIXX Index
40CU RDNA 3.5 iGPU: per AMD's marketing, up to 68% faster than an RTX 4070 laptop GPU at 1080p[1] — one chip covers both creation and 1080p gaming.
NPU on-device AI applications 78AIXX Index
XDNA 2 NPU at 50 TOPS (official figures[1]): extremely low power for local transcription/imaging/small-model residency, positioned for the Copilot+ ecosystem.
High-concurrency production inference 15AIXX Index
The 256GB/s bandwidth and consumer positioning cap concurrency — for serving external traffic, step up to data center cards (see our B300/L4 entries).

History HISTORY

January 6, 2025 , CES launch of the Ryzen AI Max series (Strix Halo): 16-core Zen 5 + 40CU iGPU + up to 128GB unified memory (launch figures[1]).

First half of 2025 , OEM systems hit the market: ASUS ROG Flow Z13 tablet, HP Z2 Mini G1a workstation, Framework Desktop, etc. (public reporting[3]).

2025–2026 , the "128GB on-device LLM" narrative exploded: AMD's official blog demonstrated local GPT-OSS 120B inference[2]; the community benchmarked LLM throughput against peer devices like DGX Spark (community reporting[4]).

2026 , one of the benchmark SKUs for on-device inference and consumer AI systems; available through domestic channels as complete systems (market details in the table below).

Technical Specifications SPECIFICATIONS

ItemSpecifications (AMD official figures, verified 2026-10[1])
Codename/ProcessStrix Halo · TSMC 4nm (chiplet multi-die packaging, public reporting[1])
CPU16-core Zen 5 · 32 threads · 3.0–5.1GHz (official figures[1])
iGPURadeon 8060S · 40 RDNA 3.5 CUs · up to 2.9GHz (official figures[1])
NPUXDNA 2 · 50 TOPS (official figures[1])
Unified memoryUp to 128GB LPDDR5X-8000 · quad-channel 256-bit · 256GB/s (official figures[1]; soldered, non-expandable)
Memory allocationShared by CPU/GPU/NPU; up to ~96GB allocatable to the iGPU (UMA+ dedicated allocation, official/community reporting[2])
PowerTDP 45W configurable up to 120W (official figures[1])
Graphics benchmarkAMD marketing claim: up to 68% faster than an RTX 4070 laptop GPU at 1080p (marketing[1])
Platform featuresUSB4 · PCIe 4.0 · Wi-Fi 7 (platform figures[3])
Typical systemsASUS ROG Flow Z13 / HP Z2 Mini G1a / Framework Desktop (public reporting[3])

Specifications are subject to AMD's official releases and blog (verified 2026-10[1][2]); the "68% faster" claim is AMD marketing and limited to 1080p gaming — rely on third-party tests for cross-comparisons.

36GBM4 Pro-class unified memory 64GBStrix Halo base model Up to 128GBRyzen AI Max+ 395 128GBDGX Spark On-device AI unified memory comparison · vendor official figures (LPDDR5X soldered, non-expandable)

Selection comparison · top three for on-device LLMs CONTEXT

ItemRyzen AI Max+ 395NVIDIA DGX SparkApple M4 Pro-class
Unified memoryUp to 128GB LPDDR5X128GB LPDDR5X24–48GB class
Memory bandwidth256GB/s (official figures[1])273GB/s (official documentation)~273GB/s (vendor page)
AI engines40CU iGPU + 50 TOPS NPUBlackwell GPU(6,144 CUDA)Apple GPU + Neural Engine
Software stackROCm (progressive Windows/Linux support, public reporting)Full CUDA stackMLX/Core ML
Form factor & powerLaptops/mini PCs 45–120WAll-in-one desktop 240W externalLaptops/mini PCs, model-dependent
One-line buying adviceLarge memory + AMD ecosystemCUDA-native developmentApple ecosystem, silent

Comparisons are based on public figures; actual LLM throughput is heavily affected by framework/quantization/context — community tests (Framework/Level1Techs, etc.[4]) serve as horizontal reference; make purchase decisions based on your own model testing.

Family Members FAMILY

Ryzen AI Max series lineup (Strix Halo):

Ryzen AI Max+ 395 (this entry · 16 cores + 40CU + 50 TOPS · series flagship)Ryzen AI Max 390 (12 cores + 32CU configuration, public reporting[1])Ryzen AI Max 385 (8 cores + 32CU configuration, public reporting[1])Host systems: ROG Flow Z13 / HP Z2 Mini G1a / Framework Desktop (public reporting[3])

Cooling & Deployment COOLING

The Ryzen AI Max+ 395 is a SoC chip ; cooling is defined by the system: laptop form factor (ROG Flow Z13 tablet) and desktop mini workstations (HP Z2 Mini G1a, Framework Desktop) share the same 45–120W configurable power range (official figures[1]) — the same chip runs at 45W in a thin laptop and 120W in a workstation, a performance gap of a full tier : before buying a system, check the OEM's rated TDP and cooling specs, then set performance expectations. Sustained-load performance (thermal throttling) should be verified via system-level testing.

Model Compatibility MODEL FIT

ScenarioFit-for-purpose guide for 128GB unified memory (estimated: weights × precision + KV/framework overhead)
Hundred-billion-class quantized inferenceGPT-OSS 120B-class fully resident (AMD official demo[2]): ~60–70GB at 4-bit, ample headroom on 128GB models
70B-class FP16 residencyWeights at ~140GB exceed the limit — FP8/INT4 quantization fits (estimated)
30B-class full residencyFP16 ~60GB class: even 64GB models can run it (estimated, size to fit)
LoRA fine-tuningSmall models feasible; throughput constrained by 256GB/s (estimated)
Not applicableHigh-concurrency external serving (capped by both bandwidth and positioning)

Method: VRAM usage ≈ parameter count × bytes per parameter (FP16=2B/FP8=1B/INT4≈0.5B), plus KV and framework overhead; same methodology as our selector's "VRAM accounting" for verification.

AMD Ryzen AI Halo Development Platform (official image via ServeTheHome)
AMD Ryzen AI Halo Development Platform · Official image via ServeTheHome (demonstrating 128GB unified memory capability)

Chip-level official photos (die/wafer) are available in AMD's official materials library; this image is taken from AMD's official blog and accurately labeled[2].

Video Overview VIDEO

Software & Ecosystem ECOSYSTEM

Each engine's software stack: iGPU/NPU go through ROCm and DirectML/ONNX Runtime (Windows home turf); Large-model inference the mainstream route is llama.cpp/ollama/LM Studio with Vulkan/ROCm backends (community reporting[4]). The gap with the CUDA ecosystem is an honest discount: test-run your own model stack first (ROCm's operator coverage on Windows is still catching up), while the Zen 5 CPU side also serves as usable inference capacity (CPU offload as fallback).

Ecosystem positioning vs. DGX Spark: Spark sells "CUDA parity with the data center," while Max+ 395 sells "large memory + performance-per-watt + Windows ecosystem" — developer habit favors Spark; dual-use on-device deployment and creation favors Max+ (community reporting[4]).

Purchasing & Inspection PROCUREMENT

1
Memory capacity comes first
64GB and 128GB are two different product lines (soldered, non-expandable) — running hundred-billion-class models requires 128GB; quotes must state capacity explicitly.
2
Verify system TDP
Same chip, 45W/120W tiers, a full tier apart in performance — the system's rated power delivery and cooling specs belong in acceptance criteria (note the power tier in benchmark conditions).
3
Channel & warranty
Consumer systems go through OEM warranty (clarify China mainland vs. international version differences); beware engineering samples on the parts market.
4
Test with your own models
Benchmark throughput and first-token latency with your own model's quantization level and context length — community scores are reference only.
5
Second-hand inspection
The iGPU has no memory chips to check — verify via the system serial number + OEM warranty status + actual memory capacity (BIOS/OS).

Common risks

Capacity mismatch risk
A 64GB unit bought for hundred-billion-class models won't cut it — calculate capacity directly from your target model's precision before deciding.
Bandwidth expectation risk
256GB/s is the physical ceiling of LPDDR5X — for throughput-sensitive workloads, review community tests before deciding (reporting[4]).

For complete-system purchases, on-site inspection or third-party verification is recommended; serial number and warranty status must match — Marketplace Search ›

How to Stress Test BENCH TEST

Judgment tools: llama.cpp/ollama benchmarks (measured tokens/s, with Quantization level and context length specified), 3DMark/gaming tests (iGPU graphics tier), HWiNFO (power/temperature limit logs), Measured Bandwidth (AIDA64 memory bandwidth vs. the 256GB/s official figure[1]). Acceptance for an edge inference machine comes down to: sustained throughput of the target model at the target Quantization level + long-context stability. Dedicated case-study clips in production.

Shipping & Delivery LOGISTICS

StageKey Points
Form FactorSold as complete systems (laptops/mini workstations) — verify original packaging and full accessories
Cross-borderConsumer compute tier, less export-control sensitive than data center cards — still subject to current listings and customs classification
InsuranceInsure at full system value (RMB tens of thousands tier); screen/portable-parts transit insurance recommended as an add-on

OEM & Variants VARIANTS

Variants / OEMForm Factor & Positioning
AMD official designStrix Halo SoC (FP11 package, official figures[1])
ASUS ROG Flow Z13Tablet form factor · 128GB tier mainstream (public reporting[3])
HP Z2 Mini G1aMini workstation form factor · content creation and edge inference (public reporting[3])
Framework DesktopDesktop modular form factor (public reporting[3])

Limitations & Controversies LIMITS

  • 256GB/s bandwidth physical ceiling : Unified memory buys capacity at the cost of bandwidth — LLM decode throughput is orders of magnitude behind discrete-GPU platforms (official figures for comparison[1]), and training suffers most.
  • Soldered memory : Capacity is fixed at purchase with no later upgrades — getting 64GB vs. 128GB wrong is a sunk cost.
  • ROCm Windows ecosystem still catching up : Long-tail operator/framework support trails CUDA (community consensus[4]) — a trial run of your own stack is a prerequisite before buying.
  • The “68% faster” marketing claim : Limited to 1080p gaming scenarios (AMD's claim[1]) — align scenarios and precision before any comparison.
  • Consumer positioning : No server-grade warranty/full ECC — for volume deployment, choose OEMs based on system reputation and support.

Related Articles NEWS

Editor-reviewed coverage of this machine and related platforms.

Models That Won't Fit Locally / Run Too Slowly —no need to pay for peak demand; call them on demand from the Model Hub
AIXX Cloud · Model Hub →

Market Notes MARKET

DateTypeMarketPrice BandBasis
2026-10Complete-system retailoverseas retailSystems priced in the RMB tens of thousands tier depending on configuration (128GB tier as the mainstream)Retail figures
2026-10System-level marketMainland ChinaChina domestic/parallel imports coexist · subject to system model and warrantyMarketplace Search ›

Note: system prices vary widely with configuration (memory/brand/form factor); no standing transaction anchor for this SKU in our database. Market and inventory: AIXX Exchange (Zero-commission Matching; not a party to transactions).

Timeline MILESTONES

TimeEvent
2025-01-06CES launch of Ryzen AI Max series (Strix Halo debut)[1]
First Half of 2025Systems like ROG Flow Z13/HP Z2 Mini hit shelves (public reporting[3])
2025–2026AMD official demo of GPT-OSS 120B running locally; edge LLM narrative explodes[2]
2026-10One of the benchmark SKUs for edge inference; mainstream config for consumer AI systems (market details below)

FAQ FAQ

How large a model can the Ryzen AI Max+ 395 run?
128GB models: GPT-OSS 120B-class fully resident (AMD official demo[2]); 4-bit handles hundred-billion-class models in general. 64GB models are comfortable at 30B-class full precision (estimated). Mind the difference between “can run” and “runs fast” — the 256GB/s bandwidth determines throughput.
How does it compare to the NVIDIA DGX Spark?
Both are 128GB edge-tier machines: Spark sells full-stack CUDA homogeneity (zero migration cost for developers), Max+ 395 sells big memory + power efficiency + Windows/creator ecosystem (comparison table above). Developer inertia picks Spark; edge deployment, dual creative focus, and budget sensitivity pick Max+. Decide with measured throughput on your own models.
Can it serve as a production inference server?
Not recommended: the 256GB/s bandwidth and consumer positioning cap concurrency — use data center cards for serving. It suits private inference where “data never leaves the machine,” personal Agent hosting, and development validation (official positioning[1]).
Buy 64GB or 128GB?
Size it directly to your target model: 30B-class FP16 full precision / hundred-billion-class 4-bit Quantization → 64GB suffices; GPT-OSS 120B-class full precision, multiple models resident, long-context large KV → go straight to 128GB (soldered, non-upgradable, get it right the first time). The quote must state capacity.
Is “68% faster than RTX 4070” true?
It's AMD's marketing claim, limited to 1080p gaming scenarios[1] — a comparison conclusion under specific games/settings. For general performance, rely on third-party tests and your own workloads; never base decisions on marketing numbers.

References REFERENCES

[1]AMD official/TechPowerUp · CES 2025 Ryzen AI Max series launch (16-core Zen5/40CU RDNA3.5/128GB LPDDR5X-8000/50 TOPS/45-120W) · https://www.techpowerup.com/330548/amd-debuts-ryzen-ai-max-s (verified 2026-10)
[2]AMD official blog · Demo of GPT-OSS 120B running locally on Ryzen AI Max platform (image source) · https://www.amd.com/en/blogs/2026/amd-ryzen-ai-max-ai-pcs-de (verified 2026-10)
[3]Public reporting · OEM system launch activity (ROG Flow Z13/HP Z2 Mini/Framework Desktop) · https://www.reuters.com/technology/ (per public reporting)
[4]Community benchmarks · Framework/Level1Techs LLM throughput benchmark collection · https://forum.level1techs.com/t/strix-halo-ryzen-ai-max-395- (third-party figures)
[5]This site · VRAM budgeting methodology in the selector (parameter count × precision + KV/framework overhead) · /wiki/xuanxing.html (this site's methodology)

Specs anchored to AMD official figures; this database does not vouch for “marketing claims” or “community consensus” entries; edge performance should be verified with your own models.

This Library
Editorial PolicyAbout This Library (Disclosure of Interests)Corrections Channel
Editorial Constitution Summary
Every claim sourcedDrawbacks always stated · corrections publicly loggedNo paid entries acceptedZero traffic-funneling or sales elements within entriesMarket data attached only as data modules
AIXX Family
AIXX PediaAIXX ExchangeAIXX News
Market overview: AIXX Exchange · © 2026 AIXX Pedia · An AIXX Product