Last month, AMD introduced the Ryzen AI Halo Developer Platform based on the Ryzen AI Max+ 395 processor, delivering up to 126 TOPS, and pre-orders are now available for $3,999.99 at MicroCenter (US only, in-store pickup).
The computer ships with 128GB LPDDR5x memory and a 2TB NVMe SSD. It offers 10GbE and WiFi 7 networking, HDMI 2.1 and USB-C DisplayPort video output, and a few extra USB-C ports
Ryzen AI Halo developer kit specifications:
- SoC – AMD Ryzen AI Max+ 395
- CPU – 16 cores, 32 threads, “Zen 5” Architecture clocked up to 5.1 GHz
- GPU – AMD Radeon 8060S Integrated Graphics with 40 CUs, RDNA 3.5 architecture
- NPU – AMD XDNA™ 2 NPU
- TDP – 120W
- Process – TSMC 4nm FinFET
- Memory – 128GB LPDDR5x up to 8,000 MT/s speed, 256GB/s bandwidth
- Storage – 2TB M.2 SSD
- Video Output
- HDMI 2.1b
- DisplayPort via USB-C port
- Networking
- 10GbE RJ45 port
- WiFi 7 and Bluetooth 5.4
- USB
- 3x USB-C ports, including one with DP Alt mode support
- 1x USB-C for power input
- Misc – Power button
- Power Supply – USB PD via USB-C port
- Dimensions – 150 x 150 x 45.4 mm
- Weight – 1.2 kg
The system can run Windows 11 or Linux to run AI workloads, and the company points to the AI Developer Program and “AMD AI Playbooks” to easily get started with step-by-step guides to run AI workloads on AMD hardware, from inference to fine-tuning. However, everything in the AI Playbooks is shown as coming soon, and we should expect the documentation to be ready once the development platform ships.
The 126 TOPS (INT8) Ryzen AI Halo Developer Platform appears to mainly compete against the 1,000 TOPS (FP4) NVIDIA DGX Spark introduced last year. Both are designed for local AI workloads, and AMD shared a few (advantageous) benchmark results where they pitted the two platforms against each other for a range of AI models: GLM 7.4 Flash-30B-A3B, GPT-OSS-120B, Qwen 3.5-122B-A10B, and Qwen 3.6-35B-A3B.
The AMD AI computer delivers 4% to 14% extra performance in these specific benchmarks. The company also highlights support for both Windows and Linux operating systems, better token/sec/dollar metrics, and the presence of a 50 TOPS NPU, not that it matters that much.
For some reason, the company also compared the AMD Ryzen AI Halo to the Apple M4 Pro, and it’s 3.3 to 7.3 times faster on a range of workloads.
It’s unclear why AMD decided to launch its own Developer Platform, since several of its customers already provide Ryzen AI Max+ 395 systems with similar specs and a slightly lower price point on Amazon and AliExpress. For instance, the GMKtec EVO-X2 with 128GB LPDDR5, 2TB NVMe SSD, four display outputs, and 2.5GbE (no 10GbE) and WiFi 7 goes for $3,299.99.
A more powerful model based on the 133 TOPS AMD Ryzen AI Max+ PRO 495 processor, and paired with 192GB of memory, is also planned. More details can be found on the product page.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress. We also use affiliate links in articles to earn commissions if you make a purchase after clicking on those links.
at least this one will get more linux support than the nvidia box i guess..
DGX Spark has much higher compute, and can scale with additional units. Not running windows is true today, but probably won’t be true tomorrow (RTX is the same hardware and will be compatible) – not that running windows is really an advantage for AI work. Memory bandwidth, much of a muchness, but the compute side and NVFP4 support that allows for the Spark to do more with less. When there was a $1500 price difference Strix halo had immediate appeal, but as their prices have equalized, I find the Spark to be a much better deal now.
https://www.cnx-software.com/2025/03/19/nvidia-dgx-spark-a-desktop-ai-supercomputer-powered-by-nvidia-gb10-20-core-armv9-soc-with-1000-tops-of-ai-performance/
I quote “The price should be around $3,000 to $4,000” for a system with 128GB, prices estimated in march 2025.
https://marketplace.nvidia.com/en-us/enterprise/personal-ai-supercomputers/dgx-spark/
$4700 with a 90-days free DGX sparc license.
So the price difference halved, or should I say they went up with about $200/month?
If AMD can cherry pick AI benchmarks where it matches or beats DGX Spark, then clearly that memory bandwidth is doing a lot of lifting.
Strix Halo is like AMD’s beta test of a mega APU and introduces a fan-out interconnect that will be used for Zen 6 Ryzens and other products. If Medusa Halo comes out with RDNA5 and LPDDR6 within the next 2 years, it should raise the compute and bandwidth by a lot while adding smaller data types. But there’s no guarantee they will add something like DGX Spark’s 200 GbE for scaling, not that everyone needs it.
Maybe one of the cheaper third-party models is a better buy for someone than Spark. There’s also the incoming 495 + 192 GB models to consider, if you really need that.
Nvidia announced a roadmap with two new gens in 2028 and 2030, so they should have something better when Medusa Halo comes out. I’m pretty sure they could stick 192 GB LPDDR5X on the existing ones if they wanted to.
I picked up the GPD Win 5 handheld — the Ryzen AI Max+ 395, 128GB RAM, 4TB model — for $2,300 right before RAM prices started to soar.
I use it as a companion dev machine. When I’m not gaming, I run small to medium-sized local models on it. It’s a pretty nifty little device.
How confusing. No Hailo AI processor.