MicroLEDs: from Headlamps to the Data Center
But let’s take a step back to understand how headlamps could possibly be relevant to the technology of AI data centers.
The steeply rising demand for AI – for both training and inference – is imposing huge pressure on equipment manufacturers to increase system compute capacity. Today, the main bottleneck is actually not in the compute function, thanks to the high performance of the accelerators (‘xPU’ devices such as GPUs and neural processing units, or NPUs) which process AI data inputs at astonishingly fast rates. The problem is the rate at which data can be shifted between xPUs, between an xPU and memory, and between servers in a rack.
Traditional copper electrical interconnects, which are familiar, cheap, and easy to integrate into electrical circuits, scale badly: over the distances required to scale up AI compute systems – sometimes up to 30m – copper interconnects require higher transmit energy, stronger equalization, and more complex signal conditioning, in part to overcome problems with electromagnetic interference (EMI). This in turn causes copper links to consume more power and generate more heat. This hamstrings equipment manufacturers’ efforts to increase bandwidth density (Gbits/s/mm) as well as system efficiency and reliability.
So, in order to increase network bandwidth, data centers have turned to optical interconnect technology. To do so, they have imported technology from the internet backbone: the companies which shuttle internet traffic between continents via a limited number of undersea cables – which are very difficult and expensive to install and maintain – have mastered the art of maximizing throughput per cable. Today’s intercontinental optical transport networks have links operating at high per-lane rates of up to 1.6Tbits/s.
This is the ‘fast-and-narrow’ approach to maximizing bandwidth: shoot as much data down each optical ‘pipe’ as fast as possible. But such high-frequency systems are complex, power-hungry and expensive. What is more, each link is a hugely consequential single point of failure which poses a high risk to system availability. And it is unclear how much further the fast-and-narrow architecture can be scaled, as every additional rise in per-lane speed and throughput becomes exponentially more difficult and expensive to implement.
In internet infrastructure, the high cost of cabling means that operators have no choice but to maximize data throughput per cable. But data centers do not face the same constraint.
This is why there is increasing interest in the data center world in a new approach: instead of going fast-and-narrow, why not go slow-and-wide? This means replacing a single, ultra-high speed link with hundreds or thousands of slower parallel optical channels, to give higher total bandwidth while using simpler, cheaper, lower-speed components.
In such a system, the data transmitters can be hundreds of microLEDs, replacing the single, high-power laser light source used in internet infrastructure. Data communications equipment manufacturers have not previously assembled parallel optical channels served by hundreds or thousands of microLED transmitters, so this slow-and-wide concept is unproved in data centers.
What a slow-and-wide architecture requires is the ability to deposit hundreds of optical emitters very close to a server’s data processors and data storage components. And because demand for AI never sleeps, data center operators need these miniature emitters to support reliable, 24/7/365 operation. But the deployment of chip-scale arrays of thousands of microLEDs has been validated in the demanding automotive market – and this is where the car headlamp can offer inspiration to data center equipment manufacturers.
It is precisely these capabilities which are in evidence in the microLED technology in ams OSRAM’s EVIYOS™ light source for adaptive beam headlights.
No comments: