The artificial intelligence (AI) hype is real. Stock price increases aside, it has become clear that tech giants are in a race to deploy lots of AI semiconductor computing systems to their data centers to unleash the power of new breakthroughs -- embodied by the generative AI chatbot ChatGPT. It looks like tens of billions of dollars in chip sales are up for grabs in the coming quarters and years.
Advanced Micro Devices (AMD 1.14%) is in a race of its own to come up with AI chip-system designs to scoop up a portion of this new market, which is currently dominated by Nvidia. And AMD has another entry point into the AI market via a transformative acquisition made last year.
Emulation (not imitation): the sincerest form of flattery?
Before a powerful Nvidia GPU gets plugged into a data center to handle AI work, it needs to be manufactured. And before it gets manufactured, it needs to be designed and tested.
It can't be stressed enough that the design and testing phase is incredibly important. Errors in chip design can be costly, running up eight to nine figures in unnecessary expenses, not to mention delays in manufacturing.
This is why emulation is a big deal. In very simple terms, hardware and system emulation is the process of imitating the functionality of a chip design, using another programmable special-purpose chip, in a semi-virtual (software based) testing environment. These special-purpose chips used in emulation are often field programmable gate arrays (FPGAs).
AMD acquired what was the largest stand-alone FPGA company, Xilinx, in early 2022. The deal transformed the old AMD -- a company that was focused on the PC market and locked in a two-way battle with Intel for decades -- into an enterprise computing powerhouse.
One year after the Xilinx takeover, AMD reported its data center and embedded chips (predominantly FPGAs) accounted for just over half of its total sales in the first quarter of 2023.
That embedded segment is where the emulation chips reside. As AI systems have sent computing hardware complexity through the roof, fellow chip companies, data center equipment manufacturers, and data center and cloud infrastructure operators need AMD FPGAs to emulate their designs before manufacturing begins in earnest.
These FPGAs are programmed (thus the "programmable" part of the FPGA acronym) to behave like the ultimate circuitry that will be manufactured, and are used to ensure AI chips behave as intended, and that AI system hardware is debugged before it is installed.
A big market unto itself
Many investors are looking to AMD to provide some competition for those top-notch Nvidia GPUs. Some think AMD chips like the MI300 Instinct computing accelerator can do the trick, although AMD is an underdog in this department.
Its chips might be powerful, but AMD lacks the in-house software and AI training library that Nvidia has, let alone a cloud-based AI training software subscription (Nvidia DGX Cloud was a game changer when announced early this year).
But FPGAs are a big bet on AI, too, and still get little love from the investor community. The addition of Xilinx remade AMD's embedded segment into a high-growth powerhouse that raked in over $1.5 billion in sales last quarter.
A new generation of FPGA, the Versal Premium VP1902, was recently announced, effectively doubling the capabilities of the previous-gen Xilinx designs. AMD says these FPGAs are ready to handle design and emulation of the most advanced AI chips and systems, and fully integrate with the three top chip design software providers: Synopsys, Cadence Design Systems, and Mentor (owned by Siemens).
Besides generative AI and data center emulation, these FPGAs are used in a myriad of other AI applications -- including automotive computing, manufacturing automation, and communications network infrastructure.
As the AI market has heated up, Wall Street's outlook for AMD has been rising, too. Profits sunk like a stone in recent quarters due to a widespread downturn in the semiconductor market, but things are looking up for the back half of 2023 and into 2024. Its stock now trades for just over 27 times expected 2024 earnings.
Data by YCharts. TTM = trailing 12 months.
This might not be the timeliest of semiconductor stocks to buy right now after the big rally in recent months, but don't sleep on AMD. It will make lots of hay from the AI hype, and in unique ways not fully understood yet by many investors.
