Three Mini Workstations That Run 70B Models Without a Discrete GPU

Nobody has compared these three yet. As of March 2026, the ASRock AI BOX-A395 was announced literally three days ago, the Acer Veriton RA100 has been shipping for weeks, and the Minisforum MS-03 exists only as a teaser shown at an Intel event in Shanghai. So this comparison is being written in real time — and the answer to "which one should I buy" is going to depend heavily on whether you know what problem you're actually solving.

Let's be direct about what this category even is. A 70B parameter model in Q4_K_M quantization weighs roughly 40GB. For years, running that locally meant either owning a pair of 3090s or waiting several seconds between sentences on CPU. What changed is that AMD's Strix Halo APU — the Ryzen AI Max+ 395 — ships with a Radeon 8060S that has 40 compute units and, critically, access to a shared LPDDR5x-8000 pool that can hit 215 GB/s of real-world bandwidth. That bandwidth is what feeds the model weights. More bandwidth means faster tokens. Simple as that.

Two of the three machines here use that exact chip. The third one doesn't.

See also: GPU Comparison: AMD vs NVIDIA for Local LLMs →

The Two AMD Machines: Same Brain, Different Body

The ASRock AI BOX-A395 and the Acer Veriton RA100 are running identical silicon. AMD Ryzen AI Max+ 395, 16 Zen 5 cores at up to 5.1 GHz, Radeon 8060S with 40 RDNA 3.5 compute units, XDNA 2 NPU at 50 TOPS, and support for up to 128GB of LPDDR5x-8000 in quad-channel configuration. Both max out at 256 GB/s theoretical memory bandwidth (actual measured is ~215 GB/s in the real world). Both can technically hold a 120B parameter model in RAM at 128GB config.

So why does the choice between them matter at all? Packaging, philosophy, and port count.

Acer's Veriton RA100 is the mainstream pick. Announced at CES 2026, available in Q1 2026, and already shipping in markets like Taiwan at 159,000 TWD (around $4,800 USD at current rates). That's not cheap, but it's in line with other Strix Halo machines. The RA100 is a clean consumer-facing box with Acer Sense Pro software for managing performance modes (Silent, Balanced, Performance), a good selection of ports — 2x USB4 at 40 Gbps, 3x USB 3.2 Gen 2, an SD card reader, HDMI 2.1 plus DisplayPort, WiFi 7, Bluetooth 5.4. It's designed for the developer or content creator who wants a known brand and doesn't want to think about it too hard.

ASRock's AI BOX-A395 came from the industrial division, announced March 17, 2026, and it shows. The aluminum chassis measures 232 × 200 × 100mm, comes with an optional carrying handle, and weighs 3kg. There's a 6-heatpipe copper cooling system with an internal 400W FLEX power supply — not an external brick. The standout spec that actually separates it from the Acer is the 10GbE port. Most other Strix Halo boxes have 2.5GbE or 5GbE. If you're planning to run this as a shared inference server that multiple machines on your network hit, 10GbE matters. The RA100 has a single LAN port and doesn't specify 10G. Pricing for the ASRock hasn't been announced.

The Intel Wild Card: Minisforum MS-03

Here's where the comparison gets complicated. The Minisforum MS-03 was teased at Intel's Core Ultra Series 3 "Panther Lake" launch event on March 12, 2026. It runs an Intel Core Ultra 7 356H — 16 cores (4 Cougar Cove P-cores, 8 Darkmont E-cores, 4 LP E-cores), up to 4.7 GHz, built on Intel's 18A process, and rated for up to 70W sustained. The NPU 5 hits 50 TOPS, same as the AMD machines on paper.

The GPU situation is very different. Instead of 40 AMD compute units, the MS-03 has 4 Intel Xe3 GPU cores — Intel Arc Battlemage integrated. And instead of 256 GB/s theoretical bandwidth, you're looking at LPDDR5x-9600 in dual-channel, which lands at around 76.8 GB/s theoretical, maybe 100 GB/s with Intel's memory-side cache compression factored in.

For Llama 3.3 70B inference, that gap is not subtle. On a Strix Halo machine, a Q4_K_M Llama 70B model generates tokens at roughly 35–50 tokens per second via the iGPU with proper backends (hipBLASLt or Vulkan on llama.cpp). On the MS-03, that same model — assuming you configure it with 64–96GB RAM so it actually fits — is going to generate somewhere in the neighborhood of 5–10 tokens per second, likely via CPU or the limited Xe3 iGPU. That's the difference between "usable in real time" and "possible if you're patient."

So why does the MS-03 exist on this list at all?

Because it's the successor to the Minisforum MS-01, and the MS-01 was the go-to machine for home lab clustering. The 195 × 195 × 42.5mm form factor is identical to the MS-01 chassis — flat, stackable, and designed with PCIe expansion in mind. If you're a Proxmox user running three of these in a cluster, or you want to build a 4-node setup where each node handles different parts of a pipeline, the MS-03 is interesting in a way the AMD machines simply aren't.

Also worth saying: the Intel 18A process story is genuinely compelling for efficiency at 70W. The MS-03 will likely run quieter under sustained loads than the AMD boxes. And Intel's OpenVINO stack for local AI workloads is more mature than AMD's ROCm ecosystem, which still requires configuration effort on Linux.

Specs Side by Side

Minisforum MS-03

Core Ultra 7 356H

16 (4P+8E+4LP)

Xe3 (4 cores)

~96GB LPDDR5x-9600

~100 GB/s (w/ compression)

50 TOPS

Up to 70W

TBD

TBD

195×195×43mm

TBD

TBA

Teased 3/12/26

The Real Question: What Are You Running This For?

For pure 70B inference throughput, both AMD machines eat the MS-03 alive. The memory bandwidth gap is the whole game. Running Llama 3.1 70B, DeepSeek-R1 70B, or Qwen2.5 72B at 40+ tokens per second in a usable local setup requires the Strix Halo's memory architecture. That's not an opinion — it's what 215 GB/s versus 100 GB/s does to LLM generation speed.

Between the two AMD machines, the choice comes down to use case. The Acer Veriton RA100 is the buy for someone who wants a polished, available-now workstation with good software, good port selection for daily use, and a name-brand warranty. The ASRock AI BOX-A395 is the buy if you're deploying this in a production environment, need 10GbE for network-attached inference, or want the industrial build quality and portable form factor. If pricing puts the ASRock near or below the Acer, it's the stronger machine on paper.

The Minisforum MS-03 is for a different buyer entirely — someone who ran MS-01 clusters in their home lab and wants to upgrade those nodes with Panther Lake's efficiency gains. For that person, the slower 70B speeds are a tradeoff worth making for the stackability and expansion options. For everyone else who wants to run a 70B model and talk to it in real time, it's the wrong machine.

Which One to Actually Buy Right Now

Want maximum 70B performance in a box you can buy today? Acer Veriton RA100 at 128GB. It's shipping, it's capable, and the software ecosystem is solid enough for production deployment.

Want the better AMD machine once pricing drops? Watch the ASRock AI BOX-A395. The 10GbE, internal PSU, and industrial cooling give it an edge over the Acer for anyone serious about using this as a local AI server rather than a personal workstation.

Building a home lab cluster or running Proxmox? Wait for the Minisforum MS-03 pricing. If it comes in under $1,200 and supports 64GB+ configs, it has a real argument as a cluster node. Just don't expect it to match the AMD machines on raw LLM inference throughput.

One last note: the GPU-less framing in the marketing for all three machines is technically accurate — no discrete GPU required — but it obscures a real difference between the AMD and Intel approaches. Strix Halo's iGPU is effectively a discrete-class GPU that happens to share a memory bus with the CPU. Panther Lake's Xe3 is not. They're both "no discrete GPU" in the same way a sports car and a family sedan are both "no racing slicks."