Author page description
November 1
- Gigabyte B343-C40-AAJ1 Review 10-Node AMD EPYC 4005 Goes High-Density
➀ The Gigabyte B343-C40-AAJ1 is a 10-node AMD EPYC 4005 (or Ryzen 9000) server housed in a 3U chassis, offering high-density computing with compact design;
➁ The system reduces costs through shared 2kW Titanium power supplies (4 for 10 nodes) and centralized management via two CMC ports, minimizing power and network infrastructure requirements;
➂ It utilizes three OCP NIC 3.0 slots for multi-host networking, enabling resource sharing and further lowering deployment costs.
October 30
- Meta IcePack Liquid-Cooled Networking Rack from OCP 2025
➀ Meta showcased its IcePack liquid-cooled networking rack at OCP 2025, targeting AI and high-performance networking applications;
➁ The rack uses pluggable optics (not co-packaged) and features 12x 64-port switches with partially populated columns for scalability;
➂ Despite liquid cooling integration, the design retains fans for redundancy, aiming to improve power density and thermal management in data centers.
October 27
- Broadcom CPO Reliability Testing with TH5-Bailly
➀ Broadcom demonstrated that its Tomahawk 5 TH5-Bailly CPO (co-packaged optics) solution offers higher reliability compared to traditional pluggable FR4 optics, with a 2.5M-hour MTBF;
➁ Lab stress tests showed significantly lower link failure rates and zero link flaps in the first 1 million device hours for CPO, alongside reduced power consumption;
➂ The results position CPO as a viable solution for AI clusters, potentially accelerating adoption in high-density networking environments.
October 25
- This is the Massive NVIDIA 800G OSFP to 2x 400G QSFP112 Passive Splitter DAC Cable
➀ NVIDIA introduces a high-speed 800G OSFP to 2x 400G QSFP112 passive splitter DAC cable, designed for connecting its SN5610 switches and ConnectX-8 NICs;
➁ The cable utilizes 112G PAM4 signaling and splits an 800G port into two 400G ports, supporting next-gen networking demands;
➂ Highlights the evolving complexity of 400G/800G connectivity standards, including OSFP and QSFP112 connector variations for high-performance computing environments.
October 23
- NextSilicon Maverick-2 Brings Dataflow and HBM3e to HPC Customers
➀ NextSilicon's Maverick-2 accelerator chip adopts a dataflow architecture, emphasizing FP64 compute performance for HPC applications, differentiating itself from NVIDIA's shift away from FP64.
➁ The chip utilizes HBM3e memory and optimizes ALU density by reducing traditional CPU overheads (e.g., branch prediction), while its software identifies computational hotspots for efficient offloading to the accelerator.
➂ Maverick-2 employs adaptive "Mill Cores" to reconfigure dataflow paths dynamically, enabling high parallelism and flexibility akin to FPGAs, with early adoption by Sandia National Lab for real-world deployment.
October 19
- The Massive Size of Broadcom Tomahawk 6 102.4T Switch Chips and Tomhawk Ultra Bonus
➀ Broadcom showcased its Tomahawk 6 102.4T switch chips and Tomahawk Ultra at OCP Summit 2025, highlighting significant size differences due to SerDes speeds (112G vs. 224G);
➁ The 112G SerDes variant features a larger package for additional I/O lanes, while the 224G version’s higher-speed SerDes enables a compact design;
➂ The Tomahawk Ultra 51.2T chip, distinct from Tomahawk 6 in architecture and package size, underscores Broadcom’s push for next-gen Ethernet scalability in AI/HPC clusters.
October 17
- NVIDIA DGX Spark and Partner GB10 Firmware
➀ The NVIDIA DGX Spark and Dell Pro Max with GB10 share similar hardware but differ in firmware, with NVIDIA and Dell providing separate signed updates;
➁ Firmware updates for Dell's system require Dell-signed packages, while DGX Spark uses NVIDIA-signed versions, impacting support responsibilities;
➂ Partner systems like Dell's may offer lower pricing (sub-$3,000) with customized storage configurations and support models compared to NVIDIA's direct offerings.
- AMD Helios MI450 Rack at OCP Summit with a Different Version from Meta
➀ AMD showcased its Helios MI450 AI reference rack at OCP Summit 2025, featuring power shelves, compute nodes, and networking trays, designed by the ZT Systems team acquired by AMD;
➁ Meta presented a modified version of the Helios rack at the event, with significant differences in power distribution, networking layouts, and component configurations;
➂ The rack highlights AMD's focus on AI/HPC infrastructure, including PCIe Gen6 support and EDSFF E1.S SSDs, with the system's design optimized for high-density AI workloads.
October 13
- Broadcom Tomahawk 6 – Davisson 102.4T Switch with Co-Packaged Optics Shipping
➀ Broadcom has started shipping its Tomahawk 6-Davisson 102.4T switch featuring integrated co-packaged optics (CPO), achieving a 70% reduction in power consumption;
➁ The switch employs TSMC's COUPE photonic engines and 3nm Condor SerDes, with field-replaceable laser modules for improved maintenance;
➂ This innovation addresses high-power networking demands in AI clusters, enabling higher port density and energy-efficient 1.6TbE connectivity.
October 11
- Cisco Silicon One P200 and Cisco 8223 for 51.2T Scale Across Networking
➀ Cisco unveiled its Silicon One P200 chip and Cisco 8223 switch, targeting 51.2T-scale networking for geographically distributed AI clusters and addressing power efficiency challenges.
➁ The Silicon One P200 emphasizes large buffers and a P4 programmable network processor for high-performance, flexible flow management, enabling multi-data-center connectivity critical for AI workloads.
➂ Competing with NVIDIA and Broadcom in scale-across networking, Cisco highlights security features and plans to support IOS XR and NX-OS, aligning with industry trends of colocating compute resources near power sources.
October 6
- AMD and OpenAI Ink Megadeal for 6GW of Future AI Compute
➀ AMD and OpenAI announced a 6GW AI compute mega-deal, marking a significant milestone in AMD's AI GPU market position with plans to deploy 1GW of MI450 GPUs in late 2026；
➁ The collaboration involves AMD's Venice server CPUs and Pensando networking solutions, providing infrastructure support for large-scale AI deployments；
➂ The agreement includes warrants for OpenAI tied to AMD's stock performance and technical milestones, signaling deep strategic alignment between the companies.
October 5
- KAYTUS MotusAI for Enterprise AI DevOps
➀ KAYTUS MotusAI is an enterprise AI DevOps platform designed to manage end-to-end AI workflows, including inference/training clusters, storage, models, and user roles;
➁ It provides multi-tier user permissions (general users, admins, supervisors), Kubernetes integration, and resource monitoring features for GPU utilization, storage, and cluster health;
➂ The platform supports NVIDIA MIG GPU partitioning, data synchronization, and cost management tools for enterprise AI infrastructure optimization.
October 4
- Dell N2224X-ON Review 24-Port 2.5GbE 4-port 25GbE 2-port 40GbE Switch
➀ The Dell N2224X-ON is a 24-port 2.5GbE switch with four 25GbE and two 40GbE ports, ideal for SMEs seeking high-density connectivity at lower costs.
➁ Available on eBay for $200-$400, it offers a user-friendly web GUI and CLI management, catering to both novices and professionals.
➂ Hardware highlights include hot-swappable 550W PSUs, modular fans, and flexible 40GbE port configurations for stacking or 10GbE breakout.
October 3
- What are 400G-SR8 Optics and Why Do They Matter
➀ 400G-SR8光模块采用8通道设计，每通道50Gbps，支持短距离（100米）400Gbps数据传输，适用于数据中心；
➁ 需搭配MPO-16 APC光纤，相比传统MPO-12或LC接口的线缆，成本更低但布线复杂度增加；
➂ 厂商示例包括NVIDIA MMA4U00 OSFP模块和FS 400G QSFP-DD模块，未来在MikroTik CRS812等交换机中将更普及。
October 2
- TP-Link 10Gbase-T PCIe Network Adapter Mini-Review
➀ The TP-Link TX401 10Gbase-T PCIe network adapter features a Marvell AQC113-B1-C controller, supporting 10GbE/5GbE/2.5GbE speeds with PCIe Gen3 x4 interface;
➁ Testing showed standard performance with iperf3 benchmarks, achieving expected throughput for a low-power (3-5W) consumer-grade solution;
➂ Designed for compact systems like mini-PCs and NAS units, it offers cost-effective 10Gbase-T connectivity but lacks advanced offload features found in datacenter NICs.
October 1
- Building New STH Studio Storage NAS
➀ The article details the construction of a high-capacity all-flash NAS for STH's video production studio to address growing storage demands from 8K/RAW camera footage and multi-editor workflows;
➁ Built around a QNAP TS-h1290FX chassis with 12x 30.72TB Solidigm D5-P5336 SSDs (360TB raw), the system leverages AMD EPYC 7002 CPUs, NVIDIA ConnectX-6 25GbE networking, and PCIe Gen4 expansion;
➂ The solution emphasizes SSD advantages in reliability, silent operation, rebuild speed, and future scalability to 1.5PB+ while maintaining a compact desktop form factor.
September 29
- MikroTik CRS812 DDQ 400GbE Switch Launched CRS812-8DS-2DQ-2DDQ
➀ MikroTik launched the CRS812-8DS-2DQ-2DDQ, a 1U switch featuring 2x400GbE QSFP56-DD ports, 2x200GbE QSFP56 ports, and 8x50GbE SFP56 ports, powered by Marvell's 98DX7335 chip and an Annapurna Labs AL52400 quad-core ARM CPU;
➁ Priced at $1,295, it targets high-performance networking with support for PAM4 56Gs signaling and includes low-cost 400Gbps QSFP-DD optics ($159);
➂ Designed for advanced users, the switch may appeal to AI/GPU clusters like NVIDIA DGX Spark, with 130-140W power consumption and dual PSU/fan modules for redundancy.
September 28
- STH Q3 2025 Letter from the Editor Trifecta Challenges
➀ STH navigates challenges balancing its roles in PR, analyst content, and influencer-driven media, driven by the growth of its YouTube channel and Substack platform;
➁ The channel reaches 1 million YouTube subscribers, while the Axautik Group's Substack surpasses 100K monthly views, focusing on monetizable industry analysis;
➂ STH Labs expands its short-form content strategy and plans to hire full-time staff to scale operations in 2026 amid evolving communication models with tech companies.
September 27
- PylePro PCO885 2U PDU Quick Look
➀ The article provides an internal overview of the PylePro PCO885 2U rackmount PDU, featuring 18 outlets and two USB-A ports;
➁ While lacking advanced features like USB-C PD and remote monitoring, its affordable price (~$89) makes it a cost-effective solution for basic power distribution needs;
➂ The detailed teardown reveals its internal wiring, breaker assembly, and construction, highlighting its simplicity and value-for-money proposition
September 24
- Touring MikroTik in Latvia to See How they Make Awesome Networking Gear
➀ The article details a tour of MikroTik's facilities in Latvia, dispelling misconceptions about the company's size and origin while showcasing its EU-based manufacturing and design processes;
➁ MikroTik's headquarters, housed in a renovated 19th-century factory, features modern offices, prototyping labs, and collaborations with contract manufacturer HansaMatrix;
➂ The visit highlights MikroTik's significant revenue scale (comparable to Netgear) and provides insights into Riga's historical landmarks alongside the company's production workflow.
- Microsoft Etches Complex Microfluidics Channels Directly into Silicon Chips
➀ Microsoft developed a new cooling technology by etching microfluidic channels directly into silicon chips, aiming to improve heat dissipation for high-power AI accelerators and future 3D-IC designs;
➁ The prototype used a 3rd Gen Intel Xeon Scalable processor (likely Ice Lake) with DDR4 RDIMMs, demonstrating leak-proof packaging and structural integrity challenges in channel depth optimization;
➂ This approach reduces thermal interface layers and increases surface contact, potentially revolutionizing liquid cooling for advanced chip packaging architectures.
- Crucial X10 2TB Portable SSD Review Compact and Neat
➀ The Crucial X10 2TB is a compact and cost-effective portable SSD with USB 3.2 Gen2 speeds, positioned between the X6 and X10 Pro models;
➁ It features IP65 dust/water resistance and a 3-meter drop rating, making it durable for mobile use;
➂ Despite its lower price than the X10 Pro, it delivers competitive performance, offering faster speeds than entry-level options like the X6.
September 20
- Why Enfabrica Has the Coolest Technology
➀ NVIDIA acquired Enfabrica for $900 million, leveraging its ACF-S chip that bridges PCIe and high-speed networking (32x112G lanes) to enhance scalability and efficiency in AI/HPC clusters;
➁ Enfabrica's EMFASYS system enables 18TB of shared CXL memory across GPUs, reducing reliance on costly HBM and improving workload optimization;
➂ The ACF-S architecture offers multi-path redundancy, minimizing bandwidth loss during failures and achieving 50% lower cost per token in large-scale AI inference.
- MikroTik CRS418-8P-8G-2S+RM Launched PoE Switch and More
➀ MikroTik launched the CRS418-8P-8G-2S+RM, a 16-port 1GbE switch with eight PoE+ ports (150W total), two 10G SFP+ uplinks, and dual redundant power supplies;
➁ The device integrates Qualcomm IPQ-8072 Arm CPU and Marvell switch chip, targeting small businesses needing an all-in-one networking solution with routing capabilities;
➂ ServeTheHome teased an upcoming factory tour video revealing MikroTik's manufacturing process and unreleased products.
September 18
- NVIDIA to Take 5 Billion Stake in Intel and Intel to Make Custom x86 NVIDIA Chips
➀ NVIDIA plans to invest $5 billion in Intel stock, pending regulatory approval, boosting Intel's share price significantly.
➁ Intel will manufacture custom x86 CPUs for NVIDIA's AI infrastructure and data centers, potentially integrating NVLink technology, impacting AMD's market position.
➂ Intel will also develop x86 SoCs with NVIDIA RTX GPU chiplets for consumer PCs, positioning NVIDIA to expand into both x86 and Arm-based consumer devices.
September 16
- Solidigm D7-PS1010 Review A Fast PCIe Gen5 NVMe SSD
➀ The Solidigm D7-PS1010 is a PCIe Gen5 NVMe SSD optimized for mixed to read-intensive workloads, available in a 7.68TB 2.5″ form factor suitable for AI systems;
➁ The drive features a distinctive black casing with passive cooling fins and is part of Solidigm's Gen5 lineup, leveraging SK Hynix's parent company technology;
➂ With a 1 DWPD endurance rating, the SSD balances performance and reliability, targeting enterprise and AI applications despite lower-rated endurance compared to older high-DWPD drives.
September 13
- SK hynix Produces HBM4 Faster than JEDEC Specs Entering Mass Production
➀ SK Hynix's HBM4 achieves over 10Gbps per-pin speed, exceeding JEDEC's 8Gbps standard by 25%, enhancing AI accelerator performance;
➁ Utilizes 1bnm process and Mass Reflow Molded Underfill (MR-MUF) technology to ensure production reliability and readiness for mass manufacturing;
➂ Positions SK Hynix as a key supplier for next-gen AI chips, benefiting clients like NVIDIA seeking premium memory solutions.
September 12
- QNAP QXG-25G2SF-CX6 25GbE NVIDIA ConnectX-6 Lx NIC Mini-Review
➀ The QNAP QXG-25G2SF-CX6 is a dual-port 25GbE NIC featuring NVIDIA's ConnectX-6 Lx chip, designed for NAS systems but adaptable to broader deployments;
➀ It features a low-profile design with an integrated cooling fan for improved thermal performance in compact spaces;
➁ Supports PCIe Gen4/Gen3 interfaces and SFP28 connectivity, enabling 25GbE or backward-compatible 10GbE network configurations.
September 9
- Ladies and Gentlemen We have a YouTube Channel 1 MILLION Subscriber Edition
➀ ServeTheHome's YouTube channel celebrates reaching 1 million subscribers, tripling its growth since surpassing 100,000 three years ago;
➁ The milestone highlights contributions from the team, sponsors, and the community, with special acknowledgment to Linus Sebastian for inspiring the channel's inception;
➂ Future content plans include studio productions, homelab builds, and international tech tours, partially funded by YouTube memberships.
- SiFive 2nd Gen Intelligence Family Launched
➀ SiFive launched its 2nd Gen Intelligence Family of RISC-V CPUs tailored for AI applications, featuring smaller sizes, lower power consumption, and enhanced accelerator IP integration;
➁ The new series includes X100/X200/X300/XM models with RVA23 standardization for improved software compatibility, and introduces dual interfaces (SSCI/VCIX) to optimize accelerator control;
➂ Architectural innovations like configurable cache hierarchy and optimized exponential functions aim to boost performance for AI accelerators while maintaining flexibility in embedded and large-scale systems.