Explainer: AI Accelerator

You are waiting at the kerb. The clock is ticking. Every second is making you late for that flight that you just cannot miss. The ride app shows your car is just around the corner. Another second passes. You fidget again. One more look at your phone screen, and your nails are in danger. Just as you are about to take a bite, the wheels screech and stop an inch away from your feet. A window rolls down, and out pops a face you drop your jaw at. There is no time for salivating, but come on, it’s Charles Leclerc. Wait! Is he going to drive you to the airport?

Confused, nervous, ecstatic, second-guessing, doing cartwheels in your mind- all at the same time- you slide into the backseat. His dashing smile is all the OTP you need. And just like that, the car flies. Now you are not worried about missing the flight. With this superhuman at the wheel, you can outrace the pilot now.

Why did the app send him, though? Whoelses, when the route involves navigating a complex flyover maze, and when many parallel lanes have to be cruised without missing a beat. Who else, when the traffic is insane, but the speedometer has to hit its max? Unless they were sending a Verstappen, no one is complaining about Mr. Leclerc.

A normal car needs these hands when acceleration has to be of a different level altogether. That too, on roads and maps that are complex, stress-dotted and ticking for time.

Your CPUs need F1 drivers too- and for the same reasons when the destination is about an AI workload. And they are getting these dapper, ultra-fast, extremely agile and ready-to-dash drivers in the form of AI accelerators now. If your computing hardware is a car engine and your driver is a processor, when it’s about handling an AI route, a normal driver will not work. At least, not alone. This erstwhile driver, or CPU, can still be there – but in the pillion seat. The wheel is better handed over to a specialised, made-for-this-speed driver- a special processor. We know it by the name of AI accelerators.

They can wear different jerseys, though. You may see them as GPUs (Graphical Processing Units) or NPUs (Neural Processing Units), or FPGAs (Field Programmable Gate Arrays). They basically do the same thing as specialised processors – helping the CPU handle the AI-level workload, manage parallel computing tasks and do it all in a reduced processing time. They fit like a T with the route of handling AI work related to machine learning, deep learning, modelling and a lot more.

If you have ever been to a gaming party, you must be aware of GPUs, which now double up as AI processors. FPGAs bring the uniqueness of being reprogrammed at the hardware level for the tasks we need them to perform. NPUs jump in when we need acceleration for deep learning applications. But every kind of processor does one thing – handle complex and time-crucial tasks like AI model training and inference in high-performance agility, which normal processors may not be up to. Especially with intensive calculations, large-scale computing work, matrix multiplications, low-latency tasks, high-throughput outcomes, and workloads that are too heavy or too complex or too fast for normal processors to handle. AI accelerators pack the purpose-built architecture, top-speed performance and special context that the whizzing-maddening map of AI needs.

And just like that, by the time you finish reading this page, your car whooshes at the airport. Fast. Smooth. Breezy. You reluctantly step out of the car and wave bye to the intolerably dash (ing) driver. As you step up to the boarding gate, you can’t help but wonder. What if the car were an F1 machine too- what would your ride have been like then? Hey, we are talking about AI PCs and AI supercomputers now! Let’s do that explainer too- after you land. Godspeed!