When Copper Meets AI: Redefining the Foundations of the Computing Power Revolution

Active Copper Cables: The Invisible Heroes Behind Large Models

Nowadays, everyone is focusing on large models, such as ChatGPT, xAI, especially Deepseek, which has recently become very popular!

These incredibly powerful AI models are trained in extremely large data centers, consuming a large amount of computing power to produce the super large models. A single GPU is far from satisfying the compute requirements of such large-scale training, requiring high-speed interconnect technology to link up GPUs, racks, etc., to achieve a multiplying effect on computing power.

As for the choice of materials for interconnects, the current principle is: "Use copper cables where possible, and only consider optical interconnects when copper cables are insufficient." Cost and power consumption are the main reasons.

Now, most supercomputing centers and Nvidia's next-generation GPUs mainly use copper cables. The NVIDIA GB200 NVL72 system uses 5,184 large copper cables, totaling over 2 miles in length. Compared to optical interconnects, this all-copper solution consumes less power. According to Nvidia co-founder and CEO Jensen Huang, compared with using optical transceivers and retimers, the copper interconnect solution can save about 20 kilowatts of electricity, reducing rack power consumption from 120 kilowatts to 100 kilowatts. In some special environments, such as data centers built on isolated islands or underwater, copper cables, as a passive transmission medium, are more durable and have lower maintenance costs. Especially in harsh environments (such as underwater or remote areas), the performance of copper cables is usually more stable. In these special occasions, despite the limited bandwidth of copper cables, they may still be the better choice.

The History of Copper

The history of copper can be traced back to the 1990s. At that time, the semiconductor industry was using aluminum wires, but as integrated circuits became smaller and smaller, the conductivity and reliability of aluminum became a bottleneck, and aluminum wires were more prone to breakage.

In 1997, IBM broke this bottleneck by announcing manufacturable copper CMOS technology. On September 1, 1998, IBM successfully launched the world's first microprocessor using copper interconnects - the PowerPC 750. Originally designed with aluminum, its operating frequency was 300 MHz. After adopting copper interconnects, the speed of the same chip could reach at least 400MHz, a 33% improvement. This breakthrough by IBM led many semiconductor manufacturers, such as Motorola, Texas Instruments, AMD, and Intel, to explore copper interconnect technologies. According to IBM's research, the electrical resistance of copper wires is about 40% lower than that of aluminum wires, which can increase the speed of microprocessors by an additional 15%, while also reducing the size. Reliability and durability also significantly improved, with long-term reliability increased by 100 times.

Furthermore, the lifespan of copper cables has become very long, reaching between 10 to 30 years. In addition, copper cables' extremely low heat generation better solves the heat dissipation problem. The more components on a chip, the greater the heat generation, and copper dissipates heat faster than aluminum, reducing the operating temperature of the chip. Starting from IBM's breakthrough, copper interconnect technology became the industry standard. Major chip manufacturers rapidly adopted copper interconnects, continuously pushing the speed and efficiency of processors.

The previous section mentioned NVIDIA's GB200 NVL72 servers, which contain more than 5,000 copper cables. Typically, a cabinet with 4,000 to 5,000 copper cables is almost standard.

How to Ensure the Stability and Reliability of Thousands of Copper Cables Working Together?

Copper cables must be tested. To improve testing efficiency, the industry uses a device called PXI network analyzers. Its purpose is to help us test multiple sets of copper cables in parallel more quickly. A fully configured PXI network analyzer is equivalent to 8 desktop 4-port network analyzers. Such a configuration provides testing speeds many times faster than using a single desktop network analyzer with a switch matrix.

If the testing volume is not particularly large, one can also choose a single network analyzer with a switch matrix, by switching to select different copper cables for testing, which is sufficient to handle small-scale testing. If more port testing is needed, one control computer can control 4 PXI chassis, testing 32 pairs of differential copper cables in parallel. This configuration is suitable for scenarios requiring fast, large-scale testing.

Future Challenges and Trends of Copper Interconnects

The current mainstream 224G transmission rate can indeed meet the needs of most data centers and high-performance computing. However, with the continuous increase in data volumes and the rapid rise in bandwidth requirements for new applications, pre-research on 448G becomes particularly important.

In the pre-research for 448G, technical challenges such as signal attenuation, noise management, and power consumption may be encountered. However, with advances in semiconductor processes and innovations in modulation and demodulation technology, these issues can be gradually overcome.

In the pre-research for 448G, the debugging and consistency testing of high-speed signals are particularly critical. As a member of the semiconductor industry chain, Keysight Technologies can provide reliable high-bandwidth testing environments for researchers to verify key performance indicators such as signal integrity, power integrity, noise management, and bit error rate.

Jintian Copper produces copper strips, copper bars, copper wires, oxygen-free copper tubes, and other copper materials, which are widely used in the semiconductor field. Feel free to contact us at 0574-83005999.