Two New Product Lines Go Live: AI Compute Servers and Junhangclaw Inference Workstations
JH Semiconductor's website now hosts two new product lines: AI compute servers and Junhangclaw AI inference workstations. The compute spectrum runs from 50 TOPS on a personal desktop to rack-scale liquid-cooled training clusters, joining the existing enterprise storage, memory, and GPU supply-chain businesses into one-stop compute-plus-storage supply.
50 TOPS or 2,070 TFLOPS: A Selection Guide to the Four Junhangclaw Inference Workstations
The W50, W120, W280, and W2000 map to four user profiles: individuals, small teams, government/enterprise xinchuang, and desktop flagship. This guide compares them by compute tier, chip route, and data boundary, and explains why the W2000's 128GB LPDDR5x unified memory matters for running large models locally.
After 600W per Card: Where Air Cooling Ends and Cold-Plate Liquid Cooling Begins for 8-10 GPU Servers
With GPUs drawing up to 600W per card, cooling selection for 8-10 GPU servers is no longer a matter of preference. This post maps the boundaries of JH Semiconductor's three cooling routes: the preconditions under which the RTX 6000D platform's independent CPU/GPU airflow ducts and the Thor T5000 platform's high-volume air cooling hold up, the three things cold-plate liquid cooling (full CPU/GPU coverage, CDU connection, supply/return water monitoring) buys in high-power-density and low-PUE scenarios, and which three kinds of users the 8-GPU air-cooled platform — positioned to lower the liquid-cooling retrofit barrier — actually fits.
PC-Farm 4U4H/6U6H: Pluggable PC Nodes in a Standard Rack — Maintenance Granularity, the 750W Backplane and Redundant Power Design
JH Semiconductor's PC-Farm distributed node servers pack four (4U4H) or six (6U6H) pluggable PC nodes into a standard rack chassis. This post explains the workload shape this form factor targets: why node hot swap plus fully independent power control cuts maintenance granularity and the failure domain down to a single node, what the two-layer power design — a 750W-peak-per-node backplane and four redundant CRPS 1100W supplies — actually means, and why one GPU card per node happens to match count-scaled businesses such as render farms and batch inference.
The Three Ledgers of J960C: Capacity Gains, Latency Costs, and Model Boundaries of In-Drive Transparent Compression
The transparent-compression model J960C places its compression engine below the FTL, trading host-invisible pipeline stages for effective-capacity and write-amplification gains. This article breaks down the latency and throughput costs, the boundary between compressible and incompressible data, and why it exists as a dedicated TITAN 9 model rather than a line-wide default.
A Second Expansion Path Beyond DDR Channels: How CXL 2.0 Memory Expansion Lands in E3.S Bays
DDR channel counts lock a server's memory capacity and bandwidth into a slowly growing product, while CXL 2.0 memory expansion inserts a byte-addressable capacity tier between local DRAM and SSDs. This article gives a qualitative account of the memory wall, the engineering logic of memory tiering, and the deployment convenience of reusing storage bays through the E3.S form factor.
Component Procurement or Full-Rack Delivery: Five Links a Rack-Level Liquid-Cooled AI Cluster Must Close, and the Spares Chain After Handover
For the same batch of GPU compute, the line between component procurement and turnkey delivery is who carries the integration risk and integration time. This post breaks down the five links a rack-level liquid-cooled AI cluster must close in sequence — cabinet, CDU, water loop, coolant supply/return monitoring and rack commissioning — how this differs from the delivery boundary of an air-cooled single node, and how the delivery process connects to spare-parts and expansion services afterwards.
Machines Still in Service, Drives Already EOL: The Long-Supply Positioning of LEGACY 3 (J360/J380) and Change Control
Rail transit, industrial control and vertical-market systems stay in service far longer than the generational cadence of storage products, and the EOL of legacy-interface drives forces a三-way replacement dilemma of stockpiling, redesign or spot-market sourcing. This post explains the positioning of the LEGACY 3 series J360/J380 as long-supply Gen3 models — installed platforms need drives that stay the same, not drives that get faster — and how the four-layer SKU/PN/Revision/ECN change-control framework behind the frozen model list underwrites multi-year BOM stability.
How AI Data Centers Are Changing Storage Selection: From Peak Bandwidth to End-to-End Data Paths
AI storage planning cannot rely on sequential peak performance alone. This guide maps data preparation, checkpoints, inference retrieval and operations into a practical evaluation framework.
PCIe 4.0 vs PCIe 5.0 Enterprise SSDs: Six Questions to Answer Before Upgrading
PCIe 5.0 raises raw per-lane signaling rate, but benefits depend on CPU lanes, backplane, thermals, I/O behavior and software. Use this server-side upgrade framework.
U.2, E1.S, E3.S or M.2: Enterprise SSD Form Factors Are More Than Dimensions
SSD form factor affects serviceability, thermal path, bay density and future expansion. Compare U.2, M.2 and EDSFF with this engineering selection checklist.
DWPD, TBW and Write Amplification: How to Read Enterprise SSD Endurance
DWPD and TBW are useful endurance metrics, but cannot be compared without warranty term, capacity, write profile and workload. Build an actionable endurance model with this guide.
135 articles · Page 1/12