Infplane Hilbert 128GB AI Max+ 395 | Dev Review (Local AI & Game Dev)
Keywords
Summary
155 words
Critical Evaluation
Value of the Information & Strength of the Argument
The video provides substantial value through hands-on benchmarks and comparisons with existing hardware like the M4 Max and RTX 5090. The argumentation is persuasive because it includes specific numbers, real working examples, and acknowledges limitations (e.g., ROCm on Windows). The host demonstrates a strong technical background, explaining concepts like unified memory, quantization, and GPU offloading. The comparison between Vulkan and ROCm is informative, and the anecdote about running Nvidia’s own Nemotron on an AMD system is compelling. However, the review is somewhat promotional as the unit was provided by Infplane, and the host expresses clear bias toward the system’s strengths, though he does mention drawbacks such as Wi-Fi driver issues and fan noise during compilation.
Scientific Rigor, Source Quality, Title Accuracy
The scientific rigor is moderate: benchmarks are presented with context and the host is transparent about settings (e.g., default out-of-box, no tuning). Sources cited are limited to product links and companion videos from the description, with no external references to published papers or official specifications. The title accurately reflects the content, and the review covers a wide range of workloads, making it useful for technical audiences. The lack of control benchmarks from other systems aside from personal comparisons reduces impartiality, but the host does report failures and issues (e.g., ROCm not working), which lends some credibility. No comments section was provided for analysis.
233 words
Title / Content Match
The title accurately describes the content: a developer review of the Infplane Hilbert with AMD AI Max+ 395 focusing on local AI and game development.
Quality & Reliability
7/10
The review is based on hands-on testing with concrete benchmarks, but it is a promotional context (provided by manufacturer) and some claims are anecdotal.
Key Moments
Markers derived by PSI from the transcript: the creator did not define chapters.
- Introduction: overview of Infplane Hilbert with AMD AI Max+ 395 and 128GB RAM
- Unboxing and initial setup; Windows installation and driver issues
- Storage benchmark showing ~5000 MB/s write speed
- LLM inference test with llama.cpp: 1B model, 2000 tokens/s prompt processing
- Running Qwen 3.5 122B model with 73GB VRAM usage, 21 tokens/s generation
- Running Nemotron 3 Super on AMD system (sacrilege), noting it doesn't fit on RTX 5090
- Image generation with FLUX: 14.75s, slightly faster than Mac M4 Max (17.2s)
- Gaming test: Rise of the Tomb Raider at 4K max settings, smooth frame rates
- Unreal Engine 5 sample project running at 60 FPS, memory usage 28GB
- Conclusion: ports include 10GbE, storage 2TB, overall positive impression
Cited Sources
- Infplane Hilbert product page — Direct link to the workstation's official page
- Kickstarter campaign for Infplane Hilbert — Crowdfunding link for the product
- LG C2 42" Monitor (affiliate) — Monitor used in the video (affiliate link)
- QNAP NAS Drive (affiliate) — NAS drive mentioned as part of setup (affiliate link)
- Companion video: Qwen 3.5 Review — Related video about Qwen 3.5 model
- Companion video: Local AI Cluster — Related video about building a local AI cluster
- Companion video: M4 Max Review — Comparison with Apple M4 Max
Concurring Sources
- M4 Max Review (companion video) — Used as baseline comparison for performance; the AMD system outperforms in some tasks.
Contribution & Novelties
The video highlights the unique advantage of 128GB unified memory in an AMD integrated GPU setup, enabling running large quantized LLMs (up to ~100GB) and even Nvidia’s own model (Nemotron 3 Super) that doesn’t fit on many discrete GPUs. It also demonstrates competitive performance against Apple M4 Max in selective tasks like image generation. The practical comparison between ROCm and Vulkan on Windows is valuable for developers. The mention of an NPU that could be leveraged alongside CPU and GPU for sharding is a novel idea worth exploring.
Pour aller plus loin :
- AMD Ryzen AI Max+ 395 — Background on the processor architecture and unified memory.
- llama.cpp on GitHub — The tool used for LLM benchmarking; open-source and widely used.
- Unified memory concept — Explains the shared memory model central to this workstation’s design.
- ROCm documentation — AMD’s compute stack; the video notes its current limitations on Windows.
149 words
Radar Profile
The radar profile shows a well-rounded performance with high scores in information quantity and quality, reflecting the comprehensive benchmarking and real-world tests. Technical level is slightly lower due to some jargon being explained, and reliability is moderate because of the promotional nature of the content and lack of external verification.