The artikel is taken down. I still have it.
Introduction
With the rapid development of AIGC(Artificial Intelligence Generated Content) technology, Nvidia's expansion in the data center market is driving the development of the related industry chain at an unprecedented speed. The chart below clearly reflects that Nvidia is already the king of the data center market (and good luck to Intel).
Data Center Market Share
The Critical Role of Copper Cable in the Data Center
High-speed interconnections between internal servers are essential to efficient computing when building large-scale data centers. Copper cables are ideal for connecting GPUs and other hardware devices due to their low latency, low cost, high bandwidth, and reliability. Copper cable is divided into DAC(Direct Attach Cable) and ACC(Active Copper Cable); DAC is a passive product that uses conductive copper wire to achieve a direct connection between the two ends; there is no internal chip, generally shielded double coaxial line (Twinax) structure.
DAC schema
ACC schema
ACC is based on the DAC, the cable inside the increase in the active signal driver or equalizer chip. These active chips can compensate for part of the loss caused by copper transmission, so the transmission distance can reach 2 to 3 times the DAC.
Changes in the application of DAC and ACC
Connection to in-cabinet servers via DAC
In traditional data center needs, such as 400G, the performance of DAC can basically meet the short-distance transmission between servers in the cabinet. Hence, the application of ACC is still tiny. However, with the significant outbreak of AIGC, the massive arithmetic demand has spawned the data center to move towards 800G. At the 800G rate, the limit of DAC transmission distance becomes only 2 meters, which can not meet the transmission needs of AI data centers. ACC instead of DAC has become necessary to achieve the transmission distance of 4~5 meters.
Nvidia ACC application
Especially for Nvidia's high-performance GPU products, such as the DGX H100 series, to meet the massive data transmission required for large-scale language model training such as GPT-4, a large number of high-speed active copper cable lines for internal interconnections, and here we can get the answer from Nvidia's product catalog.
Nvidia Networking Interconnect Product Catalog
Conclusion
With the 60-second video generated by OpenAI's Sora last month once again wowing the world, we can't help but wonder if the age of AGI(Artificial General Intelligence) is just around the corner. But there's no doubt that the construction of large, high-speed data centers is sure to be in full swing, and with Nvidia's dominance, the demand for ACC will likely grow exponentially as well.
Bullish
AI
