CWWK CW-A9 Ryzen AI Max+ 395 for Local AI Workloads

CWWK CW-A9 is powered by the AMD Ryzen AI Max+ 395, integrates Radeon 8060S Graphics, and supports up to 128GB of LPDDR5X-8000 memory. It is positioned as an AI mini PC and compact workstation for users who need strong local computing performance.

CW-A9 features a 16-core, 32-thread processor with boost clocks up to 5.1GHz. The platform integrates Radeon 8060S Graphics, an AMD Ryzen AI NPU, and supports up to 128GB of LPDDR5X memory.

For users who want to run LLMs, AI coding tools, document processing, or other GPU-accelerated workloads locally, the key value of CW-A9 is not simply a higher TOPS figure. It is the combination of CPU, GPU, NPU, and high-capacity, high-speed memory within a single compact platform.

What Is the CWWK CW-A9?

CW-A9 is a high-performance compact PC based on the AMD Ryzen AI Max+ 395 platform.

Key specifications include:

• AMD Ryzen AI Max+ 395
• 16 cores / 32 threads
• Up to 5.1GHz
• Radeon 8060S Graphics
• 40 GPU compute cores
• AMD Ryzen AI NPU
• Up to 128GB LPDDR5X-8000
• Dual M.2 NVMe SSDs
• Dual 2.5GbE LAN
• Support for up to four displays
• Windows 11 / Linux support

These specifications allow CW-A9 to function not only as a conventional desktop computer, but also as a platform for local AI workloads that require significantly more compute resources and memory capacity.

Why Does Ryzen AI Max+ 395 Make Sense for Local AI?

Ryzen AI Max+ 395 is well suited to local AI because its resource architecture closely matches the main bottlenecks encountered in real-world local LLM workloads: unified memory bandwidth, a usable GPU compute path, and sufficient CPU resources for scheduling without becoming the bottleneck.

Unified Memory Architecture: Model Residency and Zero-Copy Advantages

The first advantage is the zero-copy model residency enabled by the unified memory architecture.

On a traditional PC, system memory and dedicated GPU memory are separate. Moving model data between the CPU and GPU creates additional memory copies and consumes bandwidth.

With Ryzen AI Max+ 395, up to 128GB of LPDDR5X is available as a unified memory pool shared by the CPU, GPU, and NPU. This allows large model weights to reside within the same memory space, reducing the need for offloading or CPU fallback when dedicated VRAM becomes insufficient.

This becomes particularly important with 30B+ models because it can determine whether a model can run reliably at all, rather than simply determining how fast it runs.

GPU Compute Role: From Auxiliary Acceleration to the Main Inference Path

The second advantage is that the GPU is not merely an auxiliary accelerator; it is part of the primary inference path.

Many AI PCs advertise strong NPU performance, but mainstream LLM inference frameworks such as vLLM, PyTorch, ROCm, and CUDA-like backends still rely primarily on GPU compute.

The value of the Radeon 8060S here is therefore not its graphics capability. It provides a general-purpose parallel compute resource capable of continuously handling workloads such as attention, matrix multiplication, and KV cache processing.

In other words, it is not simply an “accelerator”; it is part of the main compute plane.

This helps prevent workloads from falling back to CPU-only inference, which can cause a significant drop in performance.

CPU and NPU Roles: Scheduling Center and Edge AI Accelerator

The third advantage is that the CPU is better understood as a scheduler rather than the main bottleneck.

The 16-core Zen 5 CPU is important not simply because of its peak performance, but because it has enough resources to simultaneously handle tokenizer workloads, batching, API servers, container runtimes, and I/O pipelines without becoming a major constraint on the overall AI pipeline.

This matters in local LLM services because many latency problems are not caused purely by insufficient compute. CPU scheduling and data preparation can also leave the GPU waiting for work.

The NPU plays a different role.

The platform's 50 TOPS NPU is not primarily intended to run mainstream large language models. It is better suited to lighter inference workloads, embeddings, and speech or vision preprocessing.

Its value is therefore in offloading edge AI tasks, rather than replacing the GPU.

The real value of Ryzen AI Max+ 395 comes from addressing three core constraints of local AI:

• Can a large model remain fully resident in memory? — Unified memory
• Can inference continuously use the GPU as the primary compute path? — Radeon 8060S
• Can the system avoid CPU scheduling becoming a bottleneck under heavier workloads? — 16-core Zen 5 CPU

The combination of these three factors is the main reason the platform is well suited to local AI.

Why Does Up to 128GB LPDDR5X Matter for Local AI?

For local LLMs, the main role of memory capacity is to determine whether a model can be fully loaded and whether the system can maintain usable inference performance after loading it.

The value of 128GB LPDDR5X can be understood in three areas:

• Model loading capacity: 70B-class models, even with 4-bit or 8-bit quantization, can require roughly 30GB–80GB+ of memory. A 128GB configuration provides additional room for model weights and inference runtime overhead.
• KV Cache capacity: As context length increases, KV Cache consumption grows linearly. Long contexts such as 8K, 16K, and 32K can quickly consume tens of gigabytes of memory. Systems with less available memory may be forced to reduce context size or sacrifice performance.
• Concurrency: With multi-user, multi-session, or vLLM concurrent inference, each request consumes its own KV Cache resources. Insufficient memory therefore directly limits practical concurrency.

What Local AI Workloads Is CW-A9 Suitable For?

To understand how Ryzen AI Max+ 395 performs in real LLM environments, we tested Qwen3.5 models across different input lengths and concurrency levels.

The results indicate that CW-A9 is better suited to short- to medium-context local AI workloads.

Real-Time AI Conversation

For interactive workloads such as local AI chat and AI assistants, shorter contexts can significantly improve responsiveness.

For Qwen3.5-35B-A3B-AWQ, the recommended range is:

128–512 input length with 1–4 concurrent requests.

Short-Text Generation

This is suitable for tasks such as summarization, rewriting, and short-form content generation.

Recommended configuration:

128–512 input length with 4–8 concurrent requests, providing a relatively balanced tradeoff between throughput and latency.

Local Code Completion

Code completion and full-codebase analysis are very different workloads.

CW-A9 provides more stable response performance with shorter prompts.

Recommended configuration:

128–2048 input length with 8–16 concurrent requests.

Keep in mind that increasing both context length and concurrency can significantly increase ITL and TTFT.

For local AI coding workloads, it is therefore better to control context size rather than directly feeding an entire code repository into the model.

Document Processing and Summarization

Medium-length inputs remain practical, but latency increases significantly at 8192–16384 context lengths.

A more appropriate approach is:

Chunk the document first, then perform model inference, rather than sending extremely long contexts directly to the model.

We Tested Ryzen AI Max+ 395 with Local LLMs

Rather than relying only on theoretical AI specifications, we conducted actual LLM inference tests on the Ryzen AI Max+ 395 platform.

The test environment included:

• Ubuntu 24.04.1 LTS
• AMD Ryzen AI Max+ 395
• 64GB Memory
• ROCm 7.2.1
• Python 3.12
• PyTorch
• vLLM
• Qwen3.5-27B-AWQ
• Qwen3.5-35B-A3B-AWQ

These configurations were taken directly from the actual test environment.

One point is particularly important:

The benchmark was performed on a 64GB memory test system, not the 128GB configuration.

Different memory configurations, ROCm versions, model versions, quantization methods, context lengths, and concurrency settings can all affect the final results.

What Did the Test Results Tell Us?

One of the clearest trends in the test results was:

Ryzen AI Max+ 395 scales much better with concurrency at short context lengths than at long context lengths.

Using Qwen3.5-35B-A3B-AWQ as an example, increasing concurrency at shorter input lengths can significantly improve TPS. However, as the input length increases to 8K and beyond, TTFT rises rapidly while the throughput gains from additional concurrency begin to diminish.

At an input length of 4096, TPS was approximately 5.62 at concurrency 1 and increased to roughly 7 TPS at higher concurrency levels.

At an input length of 8192, however, TPS across different concurrency levels dropped to around 3.

This shows that the strength of CW-A9 is not “unlimited context length.”

It is better suited to:

Local inference, interactive AI, and development workloads with short to medium input lengths.

If you want to see the complete comparison covering TTFT, ITL, TPS, context length, concurrency, and Qwen3.5-27B vs. 35B, see the full benchmark analysis.

View the Ryzen AI Max+ 395 Local LLM Performance Benchmark →

CW-A9 Connectivity and Expansion

CW-A9 provides a relatively complete set of desktop-class connectivity and expansion options, including dual 2.5GbE Ethernet, support for up to four displays, high-speed USB-C, and NVMe storage.

Dual 2.5GbE Networking

CW-A9 is equipped with:

2 × 2.5GbE RJ45 ports.

Support for Up to Four Displays

CW-A9 supports:

• 2 × HDMI
• 1 × DisplayPort
• 1 × USB-C Display

Up to four displays can be used simultaneously.

High-Speed USB-C and NVMe Storage

CW-A9 provides two M.2 2280 NVMe PCIe 4.0 x4 storage interfaces.

It also includes:

• USB 3.2 Gen2 Type-A
• USB 3.2 Gen1 Type-A
• USB-C
• SD Card

CW-A9 Local AI Capability Overview

Is CW-A9 Right for Your Local AI Workflow?

If you are looking for a compact system that combines:

• AMD Ryzen AI Max+ 395
• Radeon 8060S
• Up to 128GB LPDDR5X
• Local LLM capability
• Dual 2.5GbE
• High-speed NVMe storage
• Multi-display output
• Windows 11 / Linux support

CW-A9 is a platform worth considering.

CW-A9 supports both Windows 11 and Linux, allowing it to function either as a high-performance everyday workstation or as a Linux-based local AI development environment.

What makes it particularly interesting is not Ryzen AI Max+ 395 alone, but the combination of a high-performance CPU, Radeon 8060S, high-capacity LPDDR5X memory, and a compact chassis.

For local AI, the most important consideration is never a single TOPS figure.

What matters more is:

Can your model be loaded? Is the context size appropriate? What does real-world response performance look like? And can the entire system reliably sustain your workflow?

Choose Your CW-A9 Configuration →