“IIT'S THE the most advanced AI industry accelerator,” boasted Lisa Su, boss of Advanced Micro Devices (AMD), during the launch in December of its new MID300 tokens. Ms Su presented a series of technical specifications: 153 billion transistors, 192 gigabytes of memory and 5.3 terabytes per second of memory bandwidth. That is, respectively, approximately 2, 2.4 and 1.6 times more than the H100, the high-end artificial intelligence chip made by Nvidia. This rival chipmaker's semiconductor prowess is fueling the AI The boom has, over the past year, made it the fifth most valuable company in the United States, with a market capitalization of $1.5 trillion. Yet most experts agree that the numbers and Ms. Su were not lying: the MID300 in fact surpasses the H100. Investors also appreciated…AMDThe stock price jumped 10% the next day.
On January 30, during its quarterly results conference call, AMD announced that it plans to sell $3.5 billion worth of MID300 this year. The company also reported strong revenue of $23 billion in 2023, four times more than in 2014, when Ms. Su became chief executive. Its market value increased 100-fold under his leadership, reaching $270 billion. Compared to expected profits for the next 12 months, its valuation is even richer than that of Nvidia. Last year, it replaced Intel, which once led U.S. chip manufacturing, as the country's second-most valuable semiconductor company. Now he's tackling the bigger ones.
Such an ambition would have seemed fanciful ten years ago. At the time, Mark Papermaster recalls, AMDthe head of technology, AMD was facing an “existential crisis”. By 2008, it had spun off its chip manufacturing business to focus on processor design, outsourcing manufacturing to contract chipmakers such as TSMC from Taiwan. The idea was to be better able to compete on plans with Intel, whose vast manufacturing capacity AMD I couldn't hope to match.
It did not work. Several of AMDThe chips are down. Sales of its central processing units (CPUs), mainly intended for personal computers, were collapsing. In 2013, it sold and leased its Austin campus to raise funds. A year later, Ms. Su inherited net debt of more than $1 billion, a net annual loss of $400 million, and a market value of less than $3 billion, down from $20 billion. dollars in 2006.
She realized that the only way to AMD getting back into the game meant getting him away from the slowness PC market and focus on more promising areas like CPUs for data center servers and graphics processing units (GPUs, which make video game visuals realistic) for game consoles. She and Mr. Papermaster bet on a new CPU architecture designed to beat Intel not only in price, but also in performance.
When things got tough
The idea was to use a Lego-like approach to building chips. By dividing a chip into smaller parts, AMD could mix and match blocks to assemble different types of chips, at lower cost. When the first such composite chips were released in 2017, they were faster and cheaper than Intel's competing offerings, perhaps in part because Intel was distracted by its own problems (including manufacturing errors repeated as he moved to ever smaller transistors). Over the past ten years AMDMarket share of lucrative servers CPUs went from nothing to 30%, breaking Intel's monopoly.
After facing a giant, AMD now faces another. The competition with Nvidia is different. On the one hand, it's personal: Ms. Su and Jensen Huang, Nvidia's Taiwanese-born boss, are distant relatives. Unlike Intel, Nvidia is, like AMD, a chip designer and therefore less prone to production missteps. More importantly, the stakes are higher. Nvidia's market value, estimated at $1.5 trillion, is based on its dominance of the gaming market. GPUnot because of their usefulness in gaming, but because they also happen to be the best type of chip for training. AI models. Ms. Su expects global sales of AI chips to reach $400 billion by 2027, up from perhaps $40 billion last year. Does it have a chance against Nvidia?
Nvidia is a formidable rival. Its revenues and operating margins are almost three times higher AMD's. According to Jefferies, an investment bank, the company dominates the AI accelerator chips, which represent 86% of these components sold worldwide; before the launch of MID300, AMD barely registered. Nvidia also offers networking equipment that connects chip clusters and software, called CUDAto manage AI workloads. Nvidia dominated AI chip manufacturing because it offers the best chips, the best networking kit and the best software, notes Doug O'Laughlin of Fabricated Knowledge, a research firm.
AMDThe new processor shows that it can compete with Nvidia on semiconductor hardware. According to Mr. Papermaster, it is the result of a ten-year investment. AMD spends nearly $6 billion a year on research and development, almost as much as its biggest rival, and twice as much as a share of its sales (see table). This allowed him to adapt his Lego approach to GPUs. The combination of a dozen blocks (or “chiplets”) in a single chip allows AMD move processors and memory closer together, increasing processing speed. In December OpenAICat creatorGoogle Tag and the hottest in the world AI startup, said he would use the MID300s for part of his training.
To surpass Nvidia in networking and software, AMD teamed up with other companies. In December, it announced a partnership with networking equipment makers, including the two largest, Broadcom and Cisco. It also supports an open source initiative for chip-to-chip communication called the Ultra Ethernet Consortium as an alternative to InfiniBand, a rival championed by Nvidia.
Biting the byte
Nvidia's lead in software will be harder to chip away at. He invested in CUDA since the mid-2000s, well before the current AI wave. AI Developers and researchers love the platform, which allows them to fine-tune the performance of Nvidia processors. AMD hopes to draw customers away from Nvidia by creating its software, ROCKm, open-source and providing tools to facilitate the transition, by translating CUDA programs in ROCKm those.
Beating Nvidia at its own game won't be easy. Mr. Huang's business does not stand still. It recently announced plans to release a new chip every year instead of every two years. Tech giants with the biggest AI ambitions – Alphabet, Amazon, Meta and Microsoft – are busy designing their own accelerator chips. Despite AMDrobust sales, investors were disappointed by its forecasts for MID300 shipments. Its stock price fell 3% the day after the publication of its latest results.
Always, AMD has a big advantage. It's not Nvidia. AI Companies are desperate for an alternative to their biggest rival, whose dominant position allows them to charge high prices and, with demand outstripping supply, ration buyers of tokens. Despite their efforts to design their own hardware, big tech companies will rely on chipmakers for some time. AMD gives them options, notes Vivek Arya of Bank of America. Microsoft and Meta have already announced their intention to use AMDIt is GPUs in their data centers. And if Nvidia makes a mistake, AMD will be there to collect the Lego pieces. Just ask Intel. ■
To stay on top of the biggest business and technology news, sign up for Bottom Line, our weekly subscribers-only newsletter.