Rene Haas

No Priors: Artificial Intelligence | Technology | Startups · Redefining Chip Architecture with Arm CEO Rene Haas

• Why Does the United States Need More Chip Factories?

  Rene Haas says the United States needs more chip factories on its own soil, for national security and to diversify the supply chain. He answers as an American citizen. Arm's headquarters are in the United Kingdom. About 40 percent of employees are in the UK, about 30 percent are in the United States, and about 30 percent are in Asia, and he calls the US footprint huge. He remembers the 1980s, when the United States led semiconductors and Japanese competitors took memory share with aggressive pricing. He thought the push to rebuild the American semiconductor industry then was the right move. After the internet and software companies took over the conversation, he says, people forgot that semi

  ▶ at 25:43 Share Transcript

• Why Did Arm Start Selling Its Own Chips?

  Rene Haas says Arm started selling a physical chip because product cycles were not slowing down, manufacturing was taking longer, and Meta could not buy the general-purpose CPU it wanted from anyone else. Arm began by licensing pieces, including CPU IP, GPU IP, and system IP. A few years earlier it moved to compute subsystems, a blueprint for how to stitch those pieces together. Demand for the subsystems was strong because they cut time to market and cost. A physical product was the next step. Some customers already have everything they need to build an Arm-based chip. Many other companies want an Arm-based product, and not every Arm customer serves those markets. Meta was the first example.

  ▶ at 2:52 Share Transcript

• Why Do AI Systems Still Need Central Processors?

  Rene Haas says central processors remain the heart of every computing problem, including artificial intelligence. After ChatGPT took off, attention locked onto accelerators, yet no computing problem has been invented that can skip the microprocessor.

  As more work moved from training toward inference, training stayed important for recursive learning and reinforcement learning, while inference uses the tokens and the information. Accelerators act like a token factory. Something still has to orchestrate, arbitrate, and decide where the tokens go, then carry them to users. Central processors do that job. Until another mechanism is invented, demand for processors stays large right beside the acc

  ▶ at 34:24 Share Transcript

• What Is the Next Bottleneck for AI?

  Rene Haas says the next bottleneck for AI is building out the data centers. Not many projects being built today are ahead of schedule and also in need of less labor than they thought. Pressure in parts of the United States to slow data center development, or to put restrictions on it, adds a further constraint.

  He calls the infrastructure headwind possibly acceptable, because wafer capacity and memory capacity would be the limit if construction were not. Several constraints throttle growth at the same time. Questions on panels keep asking when the AI bubble stops. Setting aside valuation bubbles in the stock market, supply versus demand is not even close to an oversupply. Demand is insatiab

  ▶ at 13:47 Share Transcript

• How Long Will AI Chip Supply Stay Constrained?

  Rene Haas says the constrained environment for AI chips lasts three to five years at least, so long as the Transformer is the unit of energy for AI training and AI inference. By design, that work is very compute intensive and very memory intensive.

  Young companies building AI chips are getting large funding and bringing innovative designs into an industry where capital requirements are massive. Relationships with memory vendors and substrate vendors are critical, along with access to a three nanometer line, a six nanometer line, and an advanced packaging line. The constraint does not stop in 12 months and does not stop in 24 months.

  Heavy compute and memory demand drives a need for supply

  ▶ at 12:15 Share Transcript

• What Is the Slow Part of Chip Design?

  Rene Haas says AI is already strong at the slow part of chip design at Arm, which is verification, validation, debug, and documentation rather than drawing the architecture. A design takes 24 to 36 months depending on complexity. RTL generation and mapping the architecture are not the largest share of that time.

  Daily use covers probably 80 to 90 percent of engineers at Arm. Turning the tools off would feel like the 1990s, with the internet only between the hours of two and four and a library down the hall for everything after that. RTL generation and best-in-class physical design are still immature because models train on public material while a lot of the information stays proprietary. Wo

  ▶ at 7:23 Share Transcript