Quick Verdict: The AMD Ryzen AI 9 HX 370 delivers a 50 TOPS XDNA 2 NPU with a strong integrated Radeon 890M, making it the stronger pick for sustained local inference and mixed GPU-NPU workloads. The Intel Core Ultra 9 285H rates its NPU at 48 TOPS and leads on single-threaded efficiency and platform maturity, so the right choice depends on your runtime stack and thermal budget.
Pros
- Compact mini PC form factor
- Ryzen AI 9 HX 370 processor
- 32GB DDR5 memory and 1TB SSD
- Wi-Fi 7, Bluetooth 5.4 and USB4 ports
- Windows 11 Pro preinstalled
Cons
- Compact chassis may limit internal upgrade options
- Heavy 8K editing may need more memory or storage than listed
The GEEKOM A9 Max suits buyers who want a compact Windows mini PC for productivity, multitasking and light-to-moderate creative work. Its Ryzen AI 9 HX 370, 32GB DDR5 memory and 1TB SSD form the core of the build, while the advertised AI and 8K editing figures should be confirmed through your own testing.
- Processor: AMD Ryzen AI 9 HX 370
- Memory: 32GB DDR5 RAM
- Storage: 1TB SSD
- Operating system: Windows 11 Pro
- Wireless: Wi-Fi 7 and Bluetooth 5.4
- Ports: USB4.0 and HDMI 2.1
Pros
- 24-core Core Ultra 9 285HX processor
- Dual 25G SFP+ ports for high-speed networking
- Four DDR5 SODIMM slots for memory expansion
- Four SSD slots for storage expansion
- WiFi 7 and multiple wired network ports
Cons
- Listed 32GB RAM and 1TB SSD may need upgrades
- Advanced networking ports may not suit basic home use
A compact workstation mini PC for users who want a powerful processor, advanced wired networking, and expansion slots in a small chassis. Buyers planning heavier work should review the listed memory and storage and plan for expansion.
- Processor: Core Ultra 9 285HX, 24 cores and 24 threads (24C/24T)
- Memory: 32GB RAM with four DDR5 SODIMM slots
- Storage: 1TB SSD with four SSD slots
- Networking: 2x25G SFP+, 10G RJ45, 2.5G RJ45, WiFi 7
- Video and USB: HDMI, USB4, and 2x USB4 V2 output
Architectural Overview: Two Different Philosophies for On-Device AI
The Ryzen AI 9 HX 370 and the Core Ultra 9 285H both target the premium Copilot+ class of thin-and-light and performance laptops, yet they reach the same goal through different silicon decisions. AMD’s approach places a 12-core, 24-thread Zen 5 and Zen 5c hybrid CPU alongside a Radeon 890M RDNA 3.5 integrated GPU and a dedicated XDNA 2 neural processing unit rated at 50 TOPS (INT8). Intel’s Core Ultra 9 285H, built on the Arrow Lake-H platform, relies on a tile-based design with a discrete NPU block rated at 48 TOPS, sitting alongside its Xe-LPG Arc graphics and a hybrid P-core and E-core CPU arrangement.
Both processors satisfy Microsoft’s Copilot+ baseline of 40 TOPS on the NPU, so the raw headline figures are closer than the underlying architectures suggest. The practical differences emerge in how each NPU is fed data, how memory is shared with the CPU and GPU, and how the scheduler distributes work across compute engines. For readers evaluating the broader category, our Copilot+ laptop selection guide outlines the minimum NPU thresholds and battery considerations that frame this comparison.
The XDNA 2 architecture in the Ryzen AI 9 HX 370 uses a spatial array of AI Engine tiles with local memory and a shared on-chip data fabric. AMD’s public documentation describes these tiles as arranged in a dataflow model, which reduces round-trips to system DRAM for convolution and matrix-multiply kernels. For a detailed breakdown of tile topology and dataflow scheduling, consult the AMD XDNA Architecture Technical Overview.
Comparative Specifications and Platform Data
The following table consolidates the publicly documented specifications for both processors. Values reflect manufacturer datasheets and official product briefs. Where vendors publish ranges rather than fixed values, the table reports the stated maximum.
| Specification | AMD Ryzen AI 9 HX 370 | Intel Core Ultra 9 285H |
|---|---|---|
| CPU Cores / Threads | 12 / 24 (4 Zen 5 + 8 Zen 5c) | 16 / 22 (6 P-core + 8 E-core + 2 LP E-core) |
| Max CPU Boost Clock | Up to 5.1 GHz | Up to 5.4 GHz |
| Integrated GPU | Radeon 890M (RDNA 3.5, 16 CUs) | Intel Arc Graphics (Xe-LPG, 8 Xe cores) |
| NPU Architecture | XDNA 2 (spatial dataflow) | Intel NPU 4 (discrete block) |
| NPU Rated Throughput | 50 TOPS (INT8) | 48 TOPS (INT8) |
| Total Platform AI Throughput | Up to 80 TOPS (combined NPU, GPU, CPU) | Up to 99 TOPS (combined platform) |
| Memory Support | LPDDR5X-7500, up to 128 GB unified | LPDDR5X-8533, up to 96 GB |
| Base / Default TDP | 28 W (configurable 15-54 W) | 45 W (configurable 35-65 W) |
| Process Node | TSMC N4P (compute tile) | TSMC N3B (compute tile) |
Two details in this table deserve emphasis. First, the Ryzen AI 9 HX 370 typically ships in configurations with a lower default power envelope, which shapes its sustained performance in thin chassis. Second, the Core Ultra 9 285H draws on a more advanced compute-tile process node, which is a meaningful factor in performance-per-watt on the CPU side. Neither difference settles the NPU question on its own, but both influence the thermal analysis later in this article.
NPU Throughput: Reading TOPS Figures Critically
TOPS (tera operations per second) is a theoretical peak metric that assumes perfect utilization of multiply-accumulate units at a specified precision, typically INT8. A 50 TOPS rating does not mean a model will run 50 trillion operations per second in practice. Real throughput depends on memory bandwidth, operator support, kernel fusion quality, and how efficiently the runtime maps a graph onto the hardware. When comparing a 50 TOPS part with a 48 TOPS part, the two-unit difference is well inside the margin that software optimization can swing.
This is why we analyzed manufacturer specifications, technical whitepapers, and verified user feedback from developer forums and independent benchmark aggregators, rather than relying on TOPS alone. The pattern across published sources is consistent: the XDNA 2 NPU tends to show stronger sustained throughput on transformer-style workloads when the model fits in its accessible memory path, while Intel’s NPU 4 shows competitive latency on smaller convolutional and vision models. Results vary significantly by framework version and quantization format, so the figures below should be treated as directional rather than definitive.
ONNX Runtime Throughput Considerations
ONNX Runtime is the most common cross-vendor path for NPU inference on Windows. Both platforms expose their NPUs through execution providers, and performance depends heavily on whether the model’s operators are fully supported on the NPU or fall back to CPU or GPU. Operator coverage is the first-order variable. Models built from standard attention blocks, layer normalization, and GELU activations generally compile cleanly on both NPUs, but unsupported custom operators can fragment the graph and erase much of the theoretical advantage.
For INT8 quantized vision models in the 3 to 10 million parameter range, reported throughput on both platforms tends to cluster closely, with the outcome often dictated by batch size and driver maturity. For larger language models, the constraint shifts from compute to memory capacity and bandwidth. A 7B-parameter model at 4-bit quantization occupies several gigabytes before accounting for the key-value cache, and the ceiling on context length becomes the binding limit. Readers planning larger deployments should work through the model memory footprint calculations before assuming any NPU can host the workload.
Execution provider selection also matters. The Ryzen AI software stack exposes the XDNA 2 NPU through a vendor-specific provider, while Intel’s OpenVINO toolkit targets NPU 4 alongside its CPU and GPU devices. Developers who want portable code can use DirectML as an abstraction layer, though this may leave some NPU capability unused depending on the operator set. The practical recommendation is to benchmark your actual model with the vendor provider first, then compare against a DirectML baseline.
DirectML and Local Assistant Inference
Local assistant workloads, such as on-device summarization, text rewriting, and retrieval-augmented chat, stress a different part of the system than raw vision benchmarks. These tasks involve repeated short bursts of inference interleaved with CPU work and frequent memory allocation. DirectML routes the work through the GPU or NPU depending on driver capability, and the scheduler’s decisions can make a substantial difference in perceived responsiveness.
On the Ryzen AI 9 HX 370, the Radeon 890M is a capable DirectML target in its own right, and its shared memory access to the LPDDR5X pool allows large models to be placed without copy overhead. This makes the iGPU a viable fallback for models that exceed NPU-friendly sizes. On the Core Ultra 9 285H, Arc graphics similarly participate in DirectML, and Intel’s driver stack has improved steadily in its handling of mixed precision. In both cases, first-token latency tends to be dominated by model load time and prompt processing, while steady-state token generation is bounded by memory bandwidth.
For users who expect their laptop to run a persistent local assistant, the relevant question is less about peak TOPS and more about how quickly the system can keep a quantized model resident without starving other applications. Our analysis of user-reported workloads suggests that the larger unified memory ceiling on the Ryzen platform offers flexibility for keeping bigger models warm, while the Intel platform’s more mature scheduling behavior in some applications can produce smoother interactive response. Developers choosing between these machines should also review choosing a laptop for machine learning guidance for storage and memory sizing advice that applies to both.
Thermal Efficiency and Sustained Performance
Thermal behavior determines whether a theoretical advantage survives a ten-minute inference session. Thin chassis designs commonly reach thermal equilibrium within minutes under sustained NPU and GPU load, after which clock speeds and NPU frequency are reduced. The Ryzen AI 9 HX 370’s configurable power range gives OEMs latitude to tune sustained output, and designs that allocate adequate cooling typically hold NPU frequency more consistently. The Core Ultra 9 285H’s larger default TDP can deliver higher peak CPU performance but places more demand on cooling hardware to maintain that output during long inference runs.
Efficiency per watt is where the process node difference becomes measurable. The N3B compute tile on the Intel part is designed for higher transistor density and improved switching efficiency, which benefits bursty CPU work. For sustained NPU inference, however, the NPU’s own power draw and the memory subsystem’s energy cost often dominate, and the gap narrows. Reported battery life under continuous local inference varies widely by OEM power profile, display brightness, and cooling design, which makes chassis-level testing data more informative than silicon-level figures alone.
Display choice also interacts with sustained workloads, particularly on laptops that run long sessions with static assistant interfaces. Users who pair these machines with OLED panels should review our workstation display maintenance recommendations to protect panel longevity during extended inference sessions with persistent UI elements.
Software Ecosystem and Driver Maturity
Hardware specifications rarely decide the outcome on their own. Driver maturity, framework support, and update cadence shape real-world results for months after launch. Intel’s NPU software stack has matured through successive driver releases and benefits from OpenVINO’s broad model zoo and tooling. AMD’s Ryzen AI software suite has expanded its supported model list and operator coverage, and its integration with popular inference frameworks has improved steadily. Developers should check the current release notes for their target framework, because operator support can change quickly between minor versions.
Platform considerations also extend to the wider OEM ecosystem. Many Copilot+ designs ship with firmware that exposes power profiles and thermal targets, and these settings influence NPU sustained throughput as much as the silicon does. Buyers should prioritize models with independent thermal reviews and clear documentation of their sustained power limits.
Verdict: Matching the Platform to Your Workload
For developers and power users whose priority is sustained local language model inference, large unified memory, and a capable integrated GPU that can share the workload, the Ryzen AI 9 HX 370 presents a compelling case. Its 50 TOPS XDNA 2 NPU, combined with a configurable power envelope and generous memory ceiling, suits transformer-heavy assistant workloads well. For users who value single-threaded CPU performance, the most mature Windows scheduling behavior in certain applications, and a more advanced compute process node, the Core Ultra 9 285H is the stronger choice, with a 48 TOPS NPU that remains fully capable of Copilot+ features.
Neither processor is a clear universal winner. The NPU TOPS gap is small enough that software optimization, thermal design, and memory configuration will matter more than the headline number. Buyers should verify operator support for their specific models, read chassis-level thermal and battery testing rather than silicon specifications alone, and budget memory capacity generously for any serious local inference work.
