About This English Edition
This is the English edition of “AIハードウェア図鑑” (AI Hardware Zukan), a Japanese-language site. The English articles are editorially reviewed English translations of our original Japanese articles, prepared by the same team. The Japanese originals are our primary source: all benchmarks and hands-on testing are conducted by us in Japanese first, and the English pages follow. If an English page and its Japanese original ever differ, the Japanese version is authoritative. We translate so that our first-hand measurement data is available to readers outside Japan.
Purpose of This Site
AI Hardware Zukan provides information based on hands-on measurement for anyone evaluating PCs, GPUs, and hardware for AI (artificial intelligence) workloads. We focus on the specifications, cost-performance, and operational caveats that actually matter for uses such as Stable Diffusion, ComfyUI, local LLMs, and AI image and video generation — based on results verified in real environments rather than on paper spec sheets.
Take GPU selection alone: the criteria for “gaming performance,” “AI inference performance,” “AI image-generation performance,” and “AI video-generation performance” are entirely different. The role of this Site is to fill the gaps that manufacturer spec sheets cannot address, using numbers we verified on actual hardware. For example, a GPU highly rated for gaming may be unable to run large models if its VRAM is insufficient, and so cannot show its true strength for AI. Conversely, a GPU that is only mid-tier for gaming but has ample VRAM can excel at AI workloads.
AI-related software also evolves quickly: improvements in drivers, quantization methods, and inference engines can change the effective performance of the same hardware within a matter of months. We continue to re-verify after publication, and update the text whenever something changes.
Topics We Cover
- GPUs & Graphics Cards: AI-workload comparisons of NVIDIA RTX 50/40 series, AMD Radeon RX, and Intel Arc
- Local AI Environments: Setup and measurement for Ollama, llama.cpp, LM Studio, ComfyUI, and Stable Diffusion
- AI Image & Video Generation: Hands-on comparisons of LTX, Stable Diffusion XL, FLUX, and various video-generation models
- PC Builds for AI: Choosing memory (DDR5), SSDs, CPUs, power supplies, cases, and cooling — and compatibility issues
- Fundamentals: Explanations of technical terms such as VRAM, NPU, quantization, and inference engines
- Power Consumption & Running Costs: Performance-per-watt evaluated with always-on operation in mind
- External GPUs (eGPU): Hands-on testing of GPU expansion via OCuLink and Thunderbolt
We aim to provide “choices that account not only for one-off benchmarks but also for life after purchase.” Specifically, we also consider heat and noise under long continuous operation, the stability of driver updates, and the outlook for support periods.
Test Environment
Our measurement articles are verified in the environment below. All benchmarks are first-party data obtained in our own environment, not reproduced from other sites. The detailed configuration, connection methods, and measurement procedures of our test machine are summarized on our verification environment page (in Japanese).
- CPU: Intel Core i7-14700F
- Memory: DDR5 96GB
- GPU: NVIDIA RTX 5080 (main) / RTX 5060 Ti 16GB (eGPU via OCuLink)
- OS: Windows 11 Pro (primary test environment)
- Main tools: Ollama, ComfyUI, Stable Diffusion WebUI, llama.cpp, LM Studio
- Metrics: inference speed (tokens/sec), VRAM usage, seconds per generated image, power consumption, temperature
Results vary with the driver version, model version, and prompt conditions at the time of measurement. We state the conditions as clearly as possible in each article, so that readers can reproduce the same environment. LLM inference speed in particular varies greatly with batch size, context length, and quantization format (GGUF Q4, Q5, Q8, and so on), so we strive for like-for-like comparisons under matched conditions.
Benchmarks are reported primarily as the average of multiple runs after warm-up, excluding the first launch. We avoid relying on a single one-off number and instead use reproducible metrics.
Editorial Stance
Our baseline is simple: we want you not to regret your purchase. We present realistic options by price range and use case, and we do not excessively hype specific products or provide sponsored, leading evaluations. For benchmark results, we report not only the good numbers but also the points that fell short of expectations.
We evaluate by combining the following perspectives for each use case: for AI image generation, seconds per image and VRAM requirements; for local LLMs, inference speed (tokens/sec) and the model sizes you can handle; for AI video generation, time per frame and practical resolution; and, as common factors, power consumption and heat. Rather than collapsing these into a single overall score, we make a point of stating clearly “for this use case, this is the choice.”
The AI field also advances quickly, and it is not unusual for the conventional wisdom of six months ago to no longer hold. We review past articles regularly and add notes or corrections when an evaluation changes due to a driver or model update. If you find a description that does not reflect the latest information, we would appreciate it if you would let us know via the contact form.
Operator
We are the AI Hardware Zukan editorial team. Our members have hands-on experience implementing AI, local LLMs, and AI image generation, and our policy is to provide first-hand information obtained through testing on real hardware. Rather than copying spec sheets, we make a point of reflecting our own impressions after actually running the code, generating the images, and loading the models ourselves.
For questions, correction requests, interview requests, and similar matters, please contact us via our contact form. We will respond to the extent we are able after reviewing your message. We welcome feedback that helps us improve our articles.
Affiliate Disclosure
Articles may contain affiliate links, including Amazon Associates. If you make a purchase through a link, this Site may receive a referral fee, but this is never added to your purchase price.
We do not change our product evaluations based on whether a referral fee is involved. We judge comprehensively from the standpoint of performance and cost-performance, and describe our findings candidly.
Disclaimer
The information on this Site is based on information available at the time of writing. Prices, stock, specifications, driver support status, and the like change over time. Please make your final purchasing decisions at your own responsibility.
If you try any settings, modifications, or overclocking introduced in our articles, please do so at your own risk, with a full understanding of the risks of equipment failure, loss of warranty, and data loss. Note that operating with a configuration that has insufficient power capacity or cooling may shorten the life of your equipment.
Our benchmark results are one example from a specific test environment and are not a guarantee that the same results will be reproduced in other environments. Before adopting anything, please consider it against your own use case, budget, and operating conditions.
For details, please see our Privacy Policy.