TSMC Starts Producing Tesla Dojo Chips, Computing Power to Increase 40-Fold in Three Years
Tesla’s path to autonomous driving requires significant computational power. TSMC has confirmed that Tesla’s next-generation Dojo supercomputer training chips have entered production, and computing power is expected to leap forward by 2027. (https://en.wikipedia.org/wiki/Tesla_Dojo)
Elon Musk’s grand plan involves much more than just selling cars. Software is the true key to profitability. To achieve this, Tesla requires both efficient algorithms and powerful computing. They are using a dual-pronged strategy: purchasing thousands of NVIDIA H100 GPUs and developing their own Dojo supercomputer chips. TSMC, the manufacturer responsible for these chips, has confirmed production has begun.
“Production has started for Tesla’s next-generation Dojo training modules. By 2027, we will offer a more complex wafer-level system, providing computing power over 40 times greater than existing systems,” TSMC explained during their North America Technology Symposium, where they detailed semiconductor technology and advanced packaging, allowing systems to be built on an entire wafer scale, enabling ultra-high computing performance.
01
Wafer Integration Offers 40x Computing Power
Tesla’s Dojo supercomputer revolves around its training module, arranging 25 D1 chips into a 5×5 matrix. Using 7nm process technology, each chip contains 50 billion transistors and provides 362 teraflops of computing power. It’s also scalable, allowing computation and power consumption to be adjusted as needed. This is just Tesla’s current version.
According to TSMC, Tesla’s new product differs from the wafer-scale system supplied to Cerebras. Simply put, Tesla’s Dojo training module (a 5×5 processor matrix) will be placed on a carrier wafer and filled with virtual chips in empty spaces. TSMC’s InFO (Integrated Fan-Out) will then cover this with a layer of high-density interconnects, maximizing data bandwidth between chips and operating like a giant chip.
By 2027, TSMC plans to pair these wafer-scale systems with CoWoS (Chip on Wafer on Substrate) advanced packaging technology, integrating them into wafers (SoIC). The full wafer will provide 40 times more computing power, with over 40 mask silicon layers and up to 60 high-bandwidth memory chips.
02
TSMC’s InFO_SoW Technology for High Performance
InFO_SoW technology aims to achieve high-performance connectivity, allowing Tesla’s 25 Dojo chips to work together like a single processor. TSMC uses virtual chips to fill gaps between processors to maintain consistency across the wafer.
Tesla’s wafer-scale Dojo processor contains 25 ultra-high-performance processors, consuming substantial power and requiring a sophisticated cooling system. Tesla uses advanced voltage regulation modules to supply 18,000 amps to the computing plane, dissipating up to 15,000 watts of heat, and necessitating liquid cooling.
Tesla hasn’t yet disclosed the performance of its Dojo wafer system. Still, despite facing challenges during development, it seems poised to become a powerful solution for AI training.
Elon Musk said that if NVIDIA had provided enough GPUs, Tesla might not have needed to develop Dojo. This batch of next-generation Dojo supercomputers, estimated to be part of Tesla’s new Dojo cluster, is based in New York with at least $500 million in investment.
Even with significant computing power, Tesla’s AI business remains challenging. In December last year, two senior engineers responsible for the Dojo project left the company. Now, Tesla continues to cut staff to save costs while needing more talent to ensure self-driving taxis launch on time and FSD improves.
The next-generation Dojo supercomputers will be in New York, while the Texas Gigafactory will build a 100 MW data center for autonomous driving software training, using NVIDIA solutions. Regardless of the location, TSMC is manufacturing these chips, playing a crucial role in the AI space.
03
Tesla Chair Urges Shareholders to Re-approve Musk’s High-Priced Pay Package
In a video released on Friday, Tesla board chair Robyn Denholm urged shareholders to re-approve CEO Elon Musk’s 2018 compensation plan. This comes after a Delaware judge overturned the plan in court. Musk’s pay package was initially worth up to $56 billion, but due to a drop in Tesla’s share price, it’s now valued at about $47 billion.
Denholm stated that a shareholder vote to reinstate Musk’s pay would be “critical to the company’s future.”
“We believe a judge’s opinion should not invalidate the millions of votes of the company’s shareholders. We urge you to make your voices heard by voting to approve the 2018 performance award,” she said.
Denholm argued that Musk deserves this reward because shareholders set this bonus to motivate the tech billionaire to achieve “ambitious financial and operational goals.”
She highlighted that under Musk’s leadership, Tesla’s revenue grew from $11.8 billion to $96.8 billion, turning a $2.2 billion loss into $15 billion in profit.
“These milestones have increased Tesla’s value from $53.7 billion to over $790 billion,” she said.
“His success is shared by all shareholders,” Denholm remarked about Musk, adding that investors saw a 1,100% increase in stock value.
Disclaimer: This article is created by the original author. The content of the article represents their personal opinions. Our reposting is for sharing and discussion purposes only and does not imply our endorsement or agreement. If you have any objections, please contact us through the provided channels.
