-
Is liquid cooling for AI servers done by immersing them directly in liquid
In two-phase immersion cooling, a server is dunked into a vat of liquid. The liquid actively boils next to the heat-producing components, cooling them in the process. Liquid cooling is becoming a. Liquid cooling is a thermal management technology that directly addresses the immense heat generated by high-power AI servers like NVIDIA DGX systems. Cold Plate Liquid Cooling, often referred to as Direct-to-Chip (DLC), remains the most mature and widely deployed liquid cooling approach. The Cray-2 supercomputer, deployed in 1985, was famously immersed in. A single server rack packed with the latest NVIDIA GPUs can now consume over 100,000 watts of power—equivalent to the air conditioning load of 30 homes running simultaneously. Trying to cool this with traditional fans is like pointing a small desk fan at an erupting volcano; it's simply no longer. To address these issues, there has been a shift toward liquid cooling solutions, which offer better heat dissipation by applying coolant directly to heat-generating components or immersing them in a conductive liquid.
[PDF Version]
-
Does the power consumption of AI servers account for a large proportion
AI-optimized servers already account for 21% of data center energy use in 2025. Big Tech is spending tens of billions quarterly on AI accelerators, which has led to an exponential increase in power consumption. The rise of generative AI and. According to recent research, AI energy consumption is now dominated by inference and driven less by individual model runs than by scale, deployment patterns, and system inefficiencies. 29 GWh of electricity, whereas the electricity consumption for training the larger-scale GPT-4 rose dramatically to an estimated over 50 GWh [142, 37], equivalent to nearly 0. 1% of New York City's annual electricity use. AI at Work Research and insights powering the intersection of AI and business, delivered monthly. AI's rapid expansion also drives higher water usage, emissions, and e-waste, raising urgent sustainability concerns, according to Mahmut Kandemir, a distinguished professor in the Department of Computer.
[PDF Version]
-
Selection Guide for Low-Loss QSFP28 Optical Modules for Mining Applications
This guide provides a systematic selection process to help you choose the right QSFP28 module every time. The correct choice depends on matching fiber type, reach distance, switch compatibility, power budget, breakout requirements, and overall architecture. He had processed $12,000 worth of RMA'd optics in just two weeks. His 100G spine links kept dropping with CRC errors, and the system showed a frustrating mix of interface flapping and unexplained downtime. He had verified all. With so many different QSFP28 optical transceiver modules available for 100G connections, it can sometimes be overwhelming to decide on which module is the right one. Define the Application What are you. When you pick a 100G QSFP28 transceiver, think about what your network needs. 3 standard for 100G transmissions. Unlike older CFP. Selecting the wrong 100G optical module is a silent killer of data center ROI, leading to cascading failures in port density, thermal headroom, and cabling lifecycle. Technically speaking, while all three deliver 100Gbps, their underlying physical layers—ranging from 850nm parallel VCSELs to 1310nm.
[PDF Version]
-
Selection Guide for 800G Passive Optical Networks for Data Center Interconnection
This is the unified comparison that covers all five 800G interconnect types across the metrics that drive real deployment decisions. Zero power, lowest cost, lowest latency (~5 ns/m). 3ck specifies 2m. DAC · ACC · AEC · AOC · Optical Transceivers — the complete engineer's framework for choosing the right interconnect for every link in your AI data center. 800G · AI Interconnects · NVIDIA · Updated February 2026. For short-reach connections under 3 meters, 800G Passive Direct Attach Copper (DAC) is the superior choice, offering zero power consumption, the lowest possible latency, and. Generative AI data centers require ten times more fiber than conventional setups to support GPU clusters and low-latency interconnects. The transition to 800G networking has brought two competing form factors to the forefront: QSFP-DD (Quad Small Form Factor Pluggable Double Density) and OSFP (Octal Small Form Factor Pluggable).
[PDF Version]
-
IoT-grade 1 6T optical module PAM4 selection guide
Broadcom's Optical Module PHY portfolio spans multiple technology nodes — 16nm, 7nm and now 5nm, with data rates from 100 Gbs to 1. Comprising five flagship platforms, Centenario, Jesko, Portofino, Gemera, and Cygnus, Broadcom's DSP PAM-4 portfolio covers 100G, 400G, 800G, and 1. 6T PMDs. ts for data communications applications. 6T-2xDR4H can convert 8x212Gb/s electrical data to 8x212Gb/s optical signals. 6T 2×DR4 TRO OSFP transceiver delivers ultra-high-speed optical connectivity for AI and cloud data centers requiring the highest density and energy efficiency. The module offers very high functionality and feature integration, accessible via a two-wire serial interface.
-
High-Precision Selection Guide for Metropolitan Area Network-Grade ONU Optical Network Units
This report provides a comprehensive buyers guide for Optical Network Units (ONUs), focusing on market trends, leading manufacturers, and technical specifications anticipated for 2026. Understand what an ONT really does, how it differs from a router or modem, and how to select the right ONT class for FTTH, enterprise and campus fiber projects – with clear decision rules for engineers and procurement. Choosing GPON vs. Learn how to choose the right GPON/EPON/XPON ONU or ONT for your FTTH network. Complete guide covering PON standards, port configuration, WiFi, VoIP, management features and more. The shift from outdated electrical copper systems to optical fiber is driven by the immutable demands for. While GEPON OLTs (Optical Line Terminals) manage data flow in Ethernet-based networks, GPON ONUs (Optical Network Units) serve as customer-end devices in GPON systems. 0 DEED Optical networks are known for their incredible data transmission rates and the throughput achievable from a given fiber optic cable.
[PDF Version]