ServeTheHome Video Summaries

This is a hands-on look at the ASRock Industrial AI Box A395, an embedded-class mini PC built around the AMD Ryzen AI Max+ 395 with 128GB of LPDDR5X memory, 10GbE and 2.5GbE networking, multiple USB4 ports, and support for four displays. I opened the chassis up, walked through the port layout and internal design, ran CPU, GPU, and LLM benchmarks under Linux, and compared it against similar Strix Halo systems including the Framework Desktop.

■ Chassis, Ports, and Connectivity
- Exterior tour: handle, power button, top USB 2.0 Type-A ports, venting, and size comparison with the Framework Desktop
- Rear I/O: dual audio jacks, dual HDMI, USB 3.2 Gen 2 Type-A, two USB4 ports, and a USB 3 Type-C with DisplayPort output
- Networking: Marvel AQC113 10GbE plus Realtek RTL8125BG 2.5GbE, Wi-Fi 7 antenna, and clear port labeling

■ Internals and Hardware Design
- 400W internal FSP power supply, large heatsink and fan unit, and the resulting serviceability trade-offs
- M.2 SSD slot, Wi-Fi module placement, MCIO x4 and U.2 SPI connector, dual BIOS chips
- Reliability features: hardware watchdog and AC power-on behavior in BIOS

■ Benchmarking and Performance
- SPEC CPU, SPECrate, and FP rate at low optimization, plus an STH agent-style AI CPU benchmark
- SPECviewperf 15.01 GPU results in Linux, and LLM performance comparisons across systems
- Power consumption figures, acoustic measurements, and 10GbE topology mapping

■ Use-Case and Cost Perspective
- Positioning the box as an always-on small server or AI appliance, with shared storage and networking upsides
- Comparison notes on memory pricing, laptop alternatives, and where older DDR4 platforms still make sense

Best suited for home lab builders, small businesses, and developers who want a locally run AI system with high memory capacity and strong networking. Watching gives you a clear picture of how this class of Strix Halo hardware is laid out, what it can realistically do, and the practical factors to weigh before buying one.

多数のネットワークポートを備えながら、デスクトップとしても使えるCWWK 5155i UTというミニPCを徹底チェックします。磁気式カバーで手軽に開けられるメタルシャーシの内部構造、CPU・NICまわりの冷却設計、拡張スロット、そしてパフォーマンスとネットワーク速度の検証まで、ハードウェアを一つずつ確認していきます。

■ 外観とI/Oポート
- フロント・リアに配置されたオーディオ、USB 2.0/3.0 Type-A、USB Type-C、シリアルコンソールポート
- 6基の2.5GbE（Intel I226-V）、2基の10GBASE-T、2基のSFP+（Intel X710など）
- HDMIとDisplayPortによるディスプレイ出力

■ 内部構造と拡張性
- 磁気で固定されるサイドパネルと、メッシュ加工されたシャーシ
- NICを冷却するためのブロアファンと大型ヒートシンク
- M.2スロット2基＋5G用M.2スロット、SO-DIMMスロット、SATAアダプタ
- Intel Core Ultra 5 125H（Meteor Lake）と、Arc GPU・NPUの搭載

■ 検証結果と運用面
- ベンチマークの実施状況や10GbEペアリング時の帯域に関する確認
- アイドル時の消費電力と動作音の測定、19Vアダプタと内部電源入力の仕様
- VMware、Ubuntu、Proxmoxなどでの利用シーンの整理

ネットワーク機器としての使い方とデスクトップとしての使い方の両面を検証する内容で、ミニPCを購入・改造する前に仕様の見極め方や注意点を知りたい方、多数のNICを備えた小型ホストの選び方を学びたい方におすすめです。

This tour of Amazon Web Services' networking lab showcases the custom-built hardware that powers its data center backbone, from multi-terabit switches to long-haul optical systems. It provides a close-up look at how these components are designed, tested, and deployed at scale.

■ AWS Switch Portfolio
- Generational progression from 3.2T and 12.8T platforms to 51.2T and 102.4T designs
- Port platforms moving from 100G QSFP28 to 400G QSFP-DD and 800G/1.6T OSFP with PAM4
- Internal architecture: management processors, SSD storage, I2C monitoring boards, power distribution

■ Management, Timing, and Encryption
- 1G management switch and self-built serial console servers
- Satellite-disciplined timing appliance delivering microsecond-level precision
- Encryption switch (MaxE) supporting 400G ZR optics in a 2RU form factor

■ Optical Transport and Fiber
- DWDM modules carrying up to 64 wavelengths per fiber using 16QAM modulation
- Hollow-core fiber and its latency advantages over traditional single-mode fiber
- MMC connectors and the Firefly multi-fiber test and protection tool

■ Validation and Deployment
- Environmental chamber testing across temperature and humidity cycles

Network engineers, data center professionals, and infrastructure enthusiasts will gain a concrete understanding of the hardware choices behind hyperscale networking and the trends shaping upcoming generations of switches and optics.

STH's Patrick takes a detailed look at the GMKtec EVO-X3, a mini PC powered by the AMD Ryzen AI Max+ 395 (Strix Halo) with up to 128GB of memory and 2TB or 4TB NVMe storage. He covers its redesigned chassis, port layout, internals, performance behavior, and how it compares with other 128GB-memory systems in terms of pricing.

■ Design and Hardware
- Complete redesign compared to the EVO-X2, including a stand for vertical mounting
- Front, rear, and side port walkthrough: USB Type-A, audio jack, 2.5GbE Ethernet, USB-C, USB 4, HDMI, and an OCuLink port for an external GPU dock
- Internal layout: laptop-style cooling design, Wi-Fi 7 module, and M.2 NVMe drive options

■ Pricing and Market Context
- Current pricing around $3,599–$3,799, compared against price increases for the MacBook Pro, NVIDIA DGX Spark, and other 128GB systems
- Discussion of rising DDR5 memory costs and why 128GB systems represent a relatively affordable path to high memory capacity

■ Performance and Use Case
- High-performance mode up to 140W, power draw during LLM inference, and memory bandwidth considerations
- Running local AI models, including privacy and data-sovereignty benefits of open-weight models
- Key takeaways: platform advantages over the EVO-X2, missing 10GbE and limited USB/display ports, and the value proposition of the OCuLink option

Ideal for homelab builders, AI enthusiasts, and professionals evaluating local LLM inference hardware. Viewers will come away with a clearer picture of what the EVO-X3 offers, how it stacks up against competing 128GB systems, and which trade-offs to weigh before purchasing.

We take a close look at the GIGABYTE W775 V10L1, a desktop AI workstation built around the NVIDIA Grace Blackwell Ultra GB300 superchip, featuring a 72-core Grace CPU, Blackwell Ultra GPU, ConnectX-8 SuperNIC, and a fully liquid-cooled chassis. The video walks through the hardware design, internal architecture, connectivity, power and acoustics, then moves into real-world AI performance and practical deployment considerations.

■ Hardware Tour
- Front and side I/O, ventilation design, and rear I/O including QSFP112 400GbE ports
- 1600W 80 PLUS Platinum PSU, PCIe slots, and M.2 storage layout
- Liquid cooling loop, leak detection system, and BMC implementation

■ Architecture & Performance
- Grace CPU with LPDDR5x, Blackwell Ultra GPU with HBM3e, and NVLink C2C linkage
- Token throughput benchmarks with a 35B-class model and multi-system comparisons
- Power consumption and noise level measurements under load

■ Practical Use & Takeaways
- Business use cases for local AI infrastructure and agent workloads
- Cost considerations and where this platform fits

This is aimed at IT professionals, AI engineers, and business decision-makers evaluating on-premises AI hardware. Viewers will come away with a clear picture of how a workstation-class Grace Blackwell system is built, what it delivers in practice, and how it compares with alternative setups.