Ryzen AI Max+ 395/495 (e.g. AMD Halo Developer Platform) vs NVIDIA DGX Spark vs RTX PRO Workstations

IRC-Mania Die Seite rund um Sicherheit und Programmierung

Internet, Sicherheit , Internet Relay Chat

Gesponsert

Requirement: Top-tier desktop PC for large LLMs (70B+), multi-agent systems, and desktop use within a €3,000–5,000 budget.

You get what you pay for is an old saying, and at this price point, I’m determined not to learn that lesson the hard way.

That’s why I’ve taken a close look at what it means to spend my savings on a mini PC that costs as much as a decent used car, instead of investing in an AI server that would run me the price of a brand-new Dacia Duster.

In this article, I’m leaving out all the “budget mini PCs” that may have their place in the market, and focusing exclusively on the top tier. I’ll walk you through what matters in an AI server and why TOPS and FLOPS are often just marketing noise rather than meaningful metrics.

If you’re unsure about the different processor types (CPU, GPU, TPU, NPU), check out our article CPU, GPU, TPU, and NPU: Why the GPU Matters for AI.

I’m currently planning to buy a solid AI home server that I can scale up, run large AI models effortlessly, and still keep within budget.

Right now, the main contenders are the Ryzen AI Max+ 395 / Ryzen AI Max+ 495 for desktop use and NVIDIA DGX Spark models.

If you want to run coding agents, RAG pipelines, or just a private 70B+ LLM locally without racking up a monthly cloud bill, this new category of devices deserves a serious look. The Ryzen AI Max+ 495 is coming soon, though with only minor tweaks (see below).

The local AI hardware market has shifted in 2026: Ryzen AI Max mini PCs, NVIDIA DGX Spark, traditional desktop PCs with dedicated GPUs, and even early server variants (Hetzner GEX44) are all competing for the same audience: developers and businesses who want to run 70B-parameter models and multiple AI agents simultaneously at home or in the office. Prices vary wildly, and marketing specs like TOPS numbers are often misleading. That’s why I’m breaking this down systematically with concrete buying criteria, actual devices, and my recommendation at the end.

Why Local AI Hardware Makes Sense in 2026

I work as an application developer for law firms, building numerous AI applications and APIs that comply with GDPR regulations. I keep running into frustrating API design choices: low token limits, no conversation history, APIs behind Cloudflare systems that impose their own restrictions.

When 50 employees use the AI simultaneously, you suddenly spend all your time on logging, analytics, and debugging—not because anything is truly broken, but because you have to work around API limitations. If you’ve read our API learning path: REST, GraphQL, gRPC, Security and Documentation — IRC-Coding, you know how to adapt API calls to fit rate limits: See also - API Rate Limiting Implementation

• Rate Limiting and Throttling for APIs

The desire for your own AI without token limits, without API rate caps, is immense. And the costs often pay for themselves within months.

Before we dig into hardware, a quick word on motivation if you’re still on the fence. Anyone regularly using coding agents or multi-agent systems quickly faces three-figure monthly bills from cloud providers. A local AI mini PC with enough unified memory often pays for itself within a year if you use it intensively. Add data privacy and data sovereignty—especially if you work with sensitive source code or customer data professionally—and local inference becomes the only practical option, regardless of cost.

Buying Criteria: What Really Matters in a Mini PC for Local LLMs

I’ve broken down the evaluation into several categories. The order roughly reflects relevance for multi-agent use and 24/7 operation.

If you don’t care about the detailed explanations, jump straight to the comparison table.

I deliberately included hardware slightly outside the budget range so you can make a proper comparison and an informed decision.

1. Memory Bandwidth Beats TOPS

The most important and most misunderstood point first. When an LLM generates each token, it essentially reads the entire model from memory once. Token generation speed—measured in tokens per second—depends primarily on memory bandwidth, not theoretical compute power. The NPU TOPS figures often advertised, like “126 TOPS,” are nearly irrelevant for LLM inference because Ollama, llama.cpp, and LM Studio run models through the GPU, not the NPU. My first rule: never buy an AI mini PC based solely on the TOPS number on the box.

The Ryzen AI Max+ 395, codename Strix Halo, delivers around 256 GB/s memory bandwidth via Quad-Channel LPDDR5X. This is the central bottleneck to keep in mind during every comparison.

2. RAM capacity, sizing for 70B models, and multi-agent setups

Strix Halo devices come with RAM soldered directly to the board—you cannot upgrade it later, so capacity must be finalized at purchase. For a setup running a single 70B model alongside multiple parallel coding agents, 128 GB is non-negotiable; anything below that isn’t worth considering. Here’s what you need:

• A capable MoE model (Mixture of Experts) like Qwen3-Coder-Next as the primary agent requires roughly 45 GB in Q4 quantization
• A reasoning model for debugging, such as DeepSeek-R1 32B, consumes about 18 GB
• Several smaller models for parallel single tasks—say, Qwen3 8B—at around 5 GB each
• Headroom for KV cache during multiple concurrent conversations

64 GB RAM handles a single 30B model comfortably, but becomes tight quickly in genuine multi-agent setups with multiple models loaded in parallel.

When you crunch the numbers, it becomes clear that a single 7,000-euro GPU cannot carry these models alone. Most NVIDIA graphics cards max out at 32 GB VRAM, which isn’t enough—and your budget is already spent. A mini PC can handle this. The insight: with smaller models, an NVIDIA GPU has the advantage because it delivers more tokens per minute, but with large models you get zero tokens because it cannot load them directly or completely.

Explanation:

• A MoE model (Mixture of Experts) is a neural network containing many specialized sub-networks (experts), but for each input it activates only a small subset. This keeps overall capacity large while keeping computation and memory per prediction low, since only a few experts run per token.

3. Cooling and sustained load, 24/7 readiness

Critical if your device runs continuously as mine does: ignore peak wattage in the spec sheet and focus on sustained power ratings. Consumer-oriented devices often throttle audibly and thermally under sustained load, while workstation-class machines bring more heatsink mass and more stable fan control. An integrated power supply sized appropriately—say, 320 W rather than a tight external supply—is also a good sign of thoughtful thermal design.

4. Network and cluster capability

For multi-agent setups across a local network or future cluster expansion, two factors matter:

• 10GbE instead of 2.5GbE when multiple services or containers need to be served across the network
• USB4 v2 with 80 Gbps for direct coupling of two devices into a cluster

Some manufacturers like Minisforum explicitly market cluster capability on their MS-S1-Max series with concrete numbers—more on that shortly. With most consumer mini PCs, “cluster” is pure marketing with no real validation behind it.

5. Software ecosystem, ROCm versus CUDA

One point often overlooked in buying advice: AMD’s ROCm support for the Strix Halo GPU architecture, internally called gfx1151, is still relatively young. Driver support improved significantly in 2026, but NVIDIA’s CUDA ecosystem remains more mature and broadly supported. If you work heavily with PyTorch, fine-tuning, or custom CUDA kernels, factor that in.

6. Expandability, brand, and support

A PCIe-x16 slot, not present on all devices, allows later upgrades such as an additional network card. With no-name brands lacking a track record of firmware maintenance, I’d be cautious about a 24/7 production device—downtime risk in continuous operation quickly outweighs savings versus an established brand.

One often-overlooked expandability point: an existing OCuLink port, or one retrofitted via M.2 adapter, allows attaching an external NVIDIA GPU later as a supplement. More on the hybrid approach in its own section below.

MoE beats dense: the critical insight for model selection on Strix Halo

Because memory bandwidth is the bottleneck, Mixture-of-Experts models with few active parameters per token run disproportionately faster on Strix Halo hardware than dense models of comparable total size. In practice on a 128 GB device, this means:

Practical tip: instead of loading five to seven different models in parallel, it’s more memory-efficient to run a single model through a server with parallel request handling—llama.cpp-server, vLLM, or Ollama with multiple slots—to serve multiple concurrent agent sessions. A second model makes sense only if you genuinely need a different model quality, say for reasoning-heavy tasks.

Even in AI agent teams, maximizing vRAM or shared memory pays off. You might want to run a dev team in parallel:

For example:

• Coding model
• General language model
• Embedding model
• Reranker
• Vision model

Here you definitely don’t want to settle for an extremely small model just because memory ran out.

Ryzen AI Max+ 395 mini PCs compared (128 GB class)

Here’s the concrete selection, rated against the criteria above. Nearly all devices use the same chip—Ryzen AI Max+ 395 with Radeon 8060S—with differences in cooling, networking, cluster capability, and brand support. For comparison, we’ve also included NVIDIA DGX Spark, AMD Halo Developer Platform, and an Apple Mac mini M4 Pro.

All recommended AI mini PCs and servers are also listed in our Amazon shop (affiliate link)

The technical alternative to Ryzen AI Max+ 395 / 495: NVIDIA DGX Spark (GB10 Grace Blackwell)

The NVIDIA DGX Spark is the obvious alternative to Strix Halo mini-PCs (AMD Halo Developer Platform) because it also offers 128 GB Unified Memory in a desktop form factor, but with the GB10 superchip—a Grace CPU with 20 ARM cores plus Blackwell GPU—instead of x86.

NVIDIA DGX Spark technical specs: 128 GB LPDDR5X Unified Memory, roughly 273 GB/s memory bandwidth—in the same ballpark as Strix Halo and not dramatically faster—up to 1 PetaFLOP FP4 compute performance with sparsity, ConnectX-7 networking to couple two units for 256 GB combined memory supporting models up to 405 billion parameters, DGX OS built on Ubuntu.

Pricing starts around €4,700 and up.

Advantages over Strix Halo: The CUDA ecosystem is considerably more mature than ROCm, there’s native NVFP4 support for 4-bit inference, official cluster coupling via high-speed networking instead of a USB4 workaround, and professional software support from NVIDIA.

Drawbacks: Despite the impressive PetaFLOP numbers, memory bandwidth barely exceeds Strix Halo’s, so a dense 70B model won’t be significantly faster on the DGX Spark either—decode speed is memory-bound, not compute-bound. The ARM architecture also comes with limitations for traditional desktop software. For dual-use scenarios combining general computing with AI work, an x86 mini-PC is more practical. Price-wise, the DGX Spark sits well above most Strix Halo alternatives at comparable RAM configurations.

NVIDIA DGX Spark:

If you specifically want to stay within the NVIDIA ecosystem or work with NVIDIA NIM and Blueprints, you get a mature software package here. For “desktop use plus local AI agents” like in our starting scenario, an x86 Strix Halo mini-PC is often the more pragmatic choice in practice.

The third option: Desktop PC with dedicated NVIDIA GPU

The classic alternative remains a desktop PC with an RTX card. Here, GPU VRAM capacity is the deciding factor—unlike Unified Memory systems, VRAM represents a hard ceiling:

• RTX 5090, 32 GB GDDR7, roughly 1,792 GB/s bandwidth: Noticeably faster than Strix Halo for everything that fits in 32 GB. A 70B model in Q4 already requires roughly 35 to 40 GB, exceeding capacity. For models up to about 30B, the RTX 5090 is the faster choice; for 70B, you need either aggressive quantization with quality loss or a different solution.
• RTX PRO 6000 Blackwell, 96 GB GDDR7 with ECC, matching the 5090’s bandwidth: A 70B model in Q4 fits comfortably with over 50 GB buffer for context and multiple concurrent users. ECC memory also protects against silent memory errors during continuous operation over many hours. A complete workstation with this card typically costs €20,000 to over €30,000 for dual-card setups—a different league from mini-PCs.

The key tradeoff: A desktop with dedicated GPU runs substantially faster than any Unified Memory system when models fit entirely in VRAM, because GDDR7 offers significantly higher bandwidth than LPDDR5X. The price is either a hard model size limit on the RTX 5090 or dramatically higher costs for the RTX PRO 6000. For coding agents with moderate model demands up to roughly 30B, where raw speed matters, a desktop with RTX 5090 is a serious alternative to a mini-PC. For 70B models within the Strix Halo budget range, Unified Memory remains the more practical path.

Tip: Strix Halo devices can also accommodate a dedicated GPU and benefit from both.

The server option: Rack systems with Ryzen AI Max+ 395

If you’re really planning for multiple users or a larger team, the question of a proper rack solution eventually comes up. It’s worth a brief overview here, because the market is still thin.

The Ryzen AI Max+ 395 is currently somewhat unusual because it’s really a workstation APU, not a classic server chip. AMD positions it as a compact AI workstation with up to 128 GB Unified Memory (with the PRO 495, potentially 192 GB in the future).

I took a closer look at ASRock Rack during my research.

Many have probably seen the ASRock Industrial AI BOX-A395. However, this isn’t a classic 19-inch rack server—it’s a compact small-form-factor workstation in an aluminum enclosure measuring 200 x 100 x 232 mm, designed for enterprise and edge deployments. Specs: the same Ryzen AI Max+ 395, up to 128 GB LPDDR5X, 10GbE plus 2.5GbE, two USB4 ports, TPM 2.0, and redundant BIOS for higher operational reliability. This device isn’t built for a classic server rack; it’s better understood as a particularly robust, enterprise-grade desktop variant.

There’s a solid reason for this: the Ryzen AI Max+ 395 is a soldered system-on-chip, not a socketed processor. ASRock Rack’s classic server mainboards for AM5 (with AMD EPYC 4004/4005 or Ryzen 7000/8000/9000) are built for socketed CPUs. Strix Halo simply doesn’t fit because there’s no matching socketed version of the chip. That’s why it’ll likely be a while before an established server vendor like ASRock Rack offers a proper 1U or 2U board specifically for this chip.

The only option I’m aware of that’s actually designed for 19-inch rack mounting is the Minisforum MS-S1 Max I mentioned earlier. The manufacturer explicitly markets it with four switchable operating modes, including a dedicated “Rack” mode, a pull-out chassis tray for maintenance, and a header for shared power control across multiple units in a cluster setup. For anyone actually running multiple Ryzen AI Max units in a rack, this is currently the most practical solution without resorting to classic EPYC or Xeon servers with significantly higher power draw and cost.

BUT no real server alternative: For a genuine server alternative, I’d look in these directions instead:

AMD Threadripper Pro

+

<AdSlot position="article-middle" />

256 GB or 512 GB RAM

+
NVIDIA GPU

Workstation

Threadripper Pro
|
+-- RTX 6000 Ada
|
+-- RTX 6000 Ada
|
+-- NVMe storage

Unified Memory

As we can see, a server is not a realistic option in this price range.

For a detailed cost-benefit analysis, check out AI Server Rental: Is a Dedicated AI Server Worth It?.

You’re thinking of pairing an RTX graphics card with a mini-PC featuring Ryzen AI Max, wondering if you’ve now got the best of both worlds?

Yes, but let’s dig into what OCuLink actually means…

The OCuLink Hybrid Approach in Detail: What the 4x Connection Really Means

As mentioned, these top-tier AI mini-PCs can indeed accommodate an NVIDIA graphics card, providing both sufficient VRAM and GPU power through that external GPU.

Since this causes confusion regularly, here’s a closer look at what OCuLink actually is and where the real limitation lies.

OCuLink is essentially just a cable that routes a PCIe connection to an external device. On most Strix Halo mini-PCs—whether built-in natively or added via an M.2-to-OCuLink adapter cable—you get four PCIe 4.0 lanes, referred to as “x4”. A standard desktop graphics card, by contrast, typically uses 16 lanes, or “x16”. This means the OCuLink port carries only a quarter of the available PCIe bandwidth: roughly 8 GB/s in each direction with PCIe 4.0 x4, compared to around 32 GB/s with a full x16 connection.

Whether this is a real bottleneck depends entirely on your workload, and this distinction gets muddled often: are you a gamer or an AI developer?

During training or fine-tuning, it’s genuinely noticeable. Gradients and activations must constantly shuttle between CPU memory and GPU, and the narrow x4 connection significantly throttles this data flow.

For pure inference—chatting with a local model or running a coding agent—the picture changes. Once the model is fully loaded into the external GPU’s VRAM, token generation requires almost no data crossing the PCIe link. Only relatively small intermediate results between layers and input/output tokens travel over the connection. Community benchmarks confirm this: for pure inference, OCuLink performs almost identically to an internally built graphics card within measurement accuracy.

It’s different when a model doesn’t fit entirely in the external card’s VRAM and must partially offload to CPU memory—called layer offloading. Then data actually moves between CPU and GPU with each token, and the x4 connection becomes noticeably restrictive. For your use case, this means: if a model fits entirely on the connected NVIDIA card, you lose almost no performance through OCuLink. But if you want to run a model larger than the card’s VRAM, requiring it to be split across devices, the APU with its own Unified Memory is usually the better choice than an OCuLink offloading setup.

A second often-overlooked constraint is a BIOS limitation on many Strix Halo boards: externally connected AMD graphics cards are hard-capped to 120 watts regardless of card model. NVIDIA cards aren’t affected by this according to current community reports, making the hybrid approach effectively NVIDIA-specific in practice. One documented real-world test with an OCuLink-connected RTX 5090 on a Strix Halo device showed roughly a third faster token generation and nearly half faster prompt processing compared to the APU alone—a real, tangible gain despite the narrow x4 connection.

This means: the 4x limitation sounds worse than it is in practice for pure inference. It only matters when you’re training, fine-tuning, or running models larger than your external card’s VRAM. For everyday use as a coding agent machine, the hybrid approach is a legitimate option, not a workaround. That said, you don’t spend this much on a graphics card just to leave performance on the table. Beyond the technical side, ego and confidence naturally play a role.

Direct Comparison: Mini-PC vs. NVIDIA DGX Spark vs. Desktop GPU PC

My Current AI Mini-PC Recommendation

For the use case of desktop work alongside local multi-agent workflows running 70B models and coding agents, I’d go with a single Ryzen AI Max+ 395 mini-PC equipped with 128 GB of RAM. A cluster setup only makes sense if you’re serious about stepping up to 235B+ models, which most multi-agent configurations don’t require.

That said, expanding a system is considerably more appealing than having to replace it down the line.

Among the Strix Halo devices, I favour the Minisforum MS-S1 Max for its thoughtful cooling design, 10GbE connectivity, and optional vendor-validated cluster and rack expansion options if you want to scale later.

The GMKtec EVO-X2 is a solid, widely available alternative when the Minisforum’s 128 GB variant is out of stock. The brand-new GMKtec EVO-X3 is worth considering if native OCuLink support for future eGPU expansion matters more to you than clustering capability. That said, I’d wait a few weeks for independent real-world testing before recommending it for round-the-clock production use—it was only unveiled days ago.

I’d avoid no-name brands without a track record when you’re looking at a device that runs continuously. I’ve included some Amazon options, but the price difference is marginal, so it’s worth sticking with the established manufacturers.

DGX Spark and RTX PRO 6000 workstations are the right choice for edge cases: mandatory CUDA support, enterprise requirements, or very large models. For the typical developer workflow with 70B models and coding agents, the price-to-performance ratio of a Strix Halo mini-PC is hard to beat right now.

If you’re torn between a mini-PC and an NVIDIA desktop, as I am, the OCuLink hybrid approach strikes me as the most compelling middle ground for 2026. Get the best of both worlds and decide confidently:

Start small with pure Unified Memory, then add a used RTX 4090 or a new RTX 5090 later for faster, smaller models, while your large 70B models continue running on APU memory. That’s my current setup.

Transparency note: This article contains Amazon affiliate links. If you purchase through these links, I may receive a commission from Amazon at no additional cost to you.

AMD Ryzen AI Max+ 395, AMD Halo Developer Platform, and the New Ryzen AI Max+ PRO 495—What’s the Difference?

NVIDIA unveiled the DGX Spark. AMD didn’t just respond with a new processor; they also introduced their own developer platform, the AMD Halo Developer Platform, built around the Ryzen AI Max+ 395.

One question keeps coming up.

Is There a Difference Between the AMD Ryzen AI Max+ 395 and the AMD Halo Developer Platform?

The short answer is yes.

The Ryzen AI Max+ 395 is the actual processor—specifically, an APU. It combines CPU, GPU, and NPU on a single chip.

The AMD Halo Developer Platform, by contrast, is a complete mini-PC that already includes this processor.

It’s comparable to an Intel Core processor. The processor itself is just one component. Only when you add a motherboard, memory, SSD, and case do you get a finished system.

Since the AMD Halo Developer Platform mini-PC is barely available in Europe, people typically opt for the Minisforum MS-S1 Max 128GB instead.

Ryzen AI Max+ 395

AMD Halo Developer Platform

The Developer Platform comes with everything needed for an AI workstation.

The actual compute power comes from the Ryzen AI Max+ 395. The Developer Platform is simply AMD’s complete reference system.

The New Ryzen AI Max+ PRO 495

Things got interesting with the announcement of the new Ryzen AI Max+ PRO 495.

At first glance, it looks like a new generation. Looking closer at the specs, you realize AMD has really just released a refresh.

Where’s the Real Difference?

The CPU and GPU have been clocked slightly higher.

In practice, depending on the workload, that translates to roughly three to five percent more performance.

The actual difference lies elsewhere entirely.

The Ryzen AI Max+ PRO 495 supports up to 192 GB of Unified Memory.

This is the point that could prove far more valuable for local AI models than marginally higher clock speeds.

Why Is Memory So Critical?

Most people initially think about more CPU power or a faster GPU.

With local language models, however, available memory often determines which models you can load at all.

As a rough guide:

For Ollama especially, this matters far more than an extra 100 MHz of CPU clock.

Are There Prices for the 192 GB Variant Yet?

Not yet.

AMD has announced the processor. Actual systems with 192 GB of Unified Memory are still in the announcement phase.

The known AMD Developer Platform with 128 GB currently sits around $4,000 USD.

Realistically, a system with 192 GB will likely land somewhere between €4,800 and €5,500, depending on the manufacturer.

Linux or Windows?

The conversation also touched on choice of operating system.

Although AMD positions the Developer Platform for both Windows and Linux, Linux is likely to be the more interesting choice for many developers.

The reasons include:

• better ROCm support
• Docker runs without additional virtualization
• simpler container usage
• native development environment
• less overhead

Windows remains worthwhile if you’re already using software that’s exclusive to it, or if WSL2 is already part of your workflow.

Who should consider this system?

A typical use case might look like this:

• Ollama
• Windsurf
• OpenWebUI
• Docker
• MCP Server
• Qdrant or Milvus
• multiple local agents

Agents in particular benefit when several models can sit in memory simultaneously.

For example:

• coding model
• general-purpose language model
• embedding model
• reranker
• vision model

With 192 GB of Unified Memory, you wouldn’t need to constantly swap between different models.

Can local models replace Claude?

An interesting question that came up was how local models stack up against Claude Sonnet or Claude Opus.

The answer isn’t straightforward.

Today, Claude Sonnet and Opus remain among the strongest models available.

However, local 70B models have caught up considerably.

*Personal opinion

The gap has narrowed significantly compared to a few years ago. But honestly, I get very solid results even with basic SWE1.7 models.

A realistic workflow

Most people probably won’t abandon cloud models entirely.

A hybrid approach makes sense right now.

• Local models handle the bulk of daily work.
• Claude or other cloud models come in only for particularly complex tasks.

This way you can balance privacy, cost, and quality effectively.

The Ryzen AI Max+ PRO 495 isn’t a completely new generation, but rather a measured refinement of the PRO 395.

The slightly higher clock speeds will barely register in day-to-day use.

The real advancement lies in support for up to 192 GB of Unified Memory.

For local AI applications running Ollama, agents, Windsurf, and multiple models simultaneously, this single point could make the decisive difference.

Personally, I’m less interested in whether the processor got three percent faster.

What’s more exciting is what new possibilities open up when you have 192 GB of Unified Memory available in a compact workstation for the first time.

Frequently asked questions

Further reading: If you want to dive straight into multi-agent systems with a local AI PC, Multi-Agent Frameworks: LangGraph, CrewAI and OpenClaw compared offers a practical overview.

Premium VPS Hosting

Schnelle und sichere VPS Server

Gesponsert