THIS ARTICLE IS SYNTHESIZED BY THE SEMICONDUCTOR INDUSTRY (ID: ICVIEWS).
TSMC began producing Tesla Dojo chips, and three years later the computing power increased by 40 times.
Tesla's road to self-driving must have enough computing power to support, TSMC confirmed that Tesla's next-generation Dojo supercomputer platform training chip began to be produced, and the computing power will make a huge leap in 2027.
Elon Musk's great plan, selling cars is only a small part, software is the key to real profits, to achieve the goal, in addition to having efficient calculations, but also indispensable powerful computing power, in terms of hardware, Tesla has adopted a two-pronged stack, on the one hand, it has bought tens of thousands of NVIDIA H100 GPUs, on the other hand, it has built its own supercomputing chip Dojo, TSMC, which is responsible for the production of supercomputer chips, has now confirmed that the product has begun production.
"Tesla's next-generation Dojo training module has begun production, and by 2027 we will be able to deliver more complex wafer-level systems with more than 40 times the computing power of existing systems. "TSMC's North American Technology Symposium detailed semiconductor technologies and advanced packaging to enable systems to be built at the scale of a wafer to create ultra-high computing performance.
Wafer integration can provide 40 times the computing power
Tesla's Dojo supercomputer is designed by Tesla, the core of which is a training tile, which arranges 25 D1 chips into a 5×5 matrix, uses a 7nm process, can hold 50 billion transistors, provides 362TFLOPs of computing power, and most importantly, has the ability to scale, can be continuously stacked, and allocate computing power and power consumption according to software needs. And that's just the mid-use version of Tesla.
According to TSMC, the new product for Tesla is different from the wafer-level system supplied to Cerebras, in simple terms, the Dojo training module (5×5 processor matrix) is placed on a carrier wafer, after filling all the blanks, TSMC's fan-out (InFO) is used to cover a layer of high-density interconnect, so that the data bandwidth between chips can be maximized, making it operate like a giant chip.
By 2027, TSMC plans to integrate these wafer-level systems with CoWoS advanced packaging technology into a wafer (SoIC), which will provide 40 times the computing power of the complete wafer, more than 40 mask silicon, and up to 60 high-bandwidth memory chips.
According to IEEE Spectrum, InFO_SoW technology is designed to enable high-performance connectivity, allowing Tesla Dojo's 25 chips to work like 1 processor, while TSMC fills in the gaps between chips with virtual chips to keep wafer-level processors consistent.
Tesla's wafer-level Dojo processors actually contain 25 ultra-high-performance processors, which consume a very high amount of power and therefore require a complex cooling system.
To meet the power needs of the Dojo processors, Tesla uses a complex voltage regulation module that provides 18,000 amps of power to the computing plane, dissipating up to 15,000W of heat, so it needs to be water-cooled.
Tesla has yet to reveal the performance of its Dojo wafer system – however, given all the challenges it faced during its development, it seems poised to be a very powerful solution for AI training.
According to Musk, if NVIDIA provides enough GPUs, Tesla probably does not need to develop Dojo on its own, and it is initially estimated that this batch of Dojo supercomputers will become part of Tesla's new Dojo cluster, located in New York, with an investment of at least $500 million.
Tesla's AI career is still full of challenges, despite the strong investment in computing power, two senior engineers in charge of the Dojo project left last December, and now Tesla continues to lay off employees to save costs, and more talented people need to dedicate their brains and livers to have a chance to launch self-driving taxis as scheduled, and try to take FSD to the next level.
Tesla's Dojo next-generation computer will be located in New York, and the headquarters of the Texas Gigafactory will build a 100MW data center, train self-driving software, and use NVIDIA supply solutions for hardware. No matter which side they are, these chips have the same goal, they are basically produced by TSMC, and it is not an exaggeration to call them AI promoters.
Tesla's chairman urged shareholders to re-approve Musk's sky-high compensation package
In a video released Friday, Tesla's board chairman, Robyn Denholm, urged shareholders to reapprove CEO Elon Musk's 2018 compensation package, which was overturned in court by a Delaware judge. Musk's compensation package was once worth as much as $56 billion, but is now worth about $47 billion due to the decline in Tesla's stock price.
Danholm said in the video that the shareholder vote to restore Musk's compensation would be "very important for the company's future."
"We believe that the opinion of one judge should not invalidate the will of all shareholders of the company to vote millions," the chairman told investors. Therefore, we once again ask you to make your voice heard by voting to approve the 2018 Performance Awards. ”
Danholm argued in the video that it was fair for Musk to receive the reward. Shareholders set the bonus to incentivize the tech billionaire to achieve "ambitious financial and operational goals."
She noted that under Musk's leadership, Tesla's revenue grew from $11.8 billion to $96.8 billion and turned a $2.2 billion loss into a $15 billion profit.
The milestones have increased Tesla's value from $53.7 billion to more than $790 billion, she said.
"His success is shared by all shareholders," Danholm said of Musk, adding that investors have seen their stock value rise by 1,100 percent.
*Disclaimer: This article was created by the original author. The content of the article is his personal point of view, and our reprint is only for sharing and discussion, and does not mean that we agree or agree, if you have any objections, please contact the background.
THIS ARTICLE IS SYNTHESIZED BY THE SEMICONDUCTOR INDUSTRY (ID: ICVIEWS).
TSMC began producing Tesla Dojo chips, and three years later the computing power increased by 40 times.
Tesla's road to self-driving must have enough computing power to support, TSMC confirmed that Tesla's next-generation Dojo supercomputer platform training chip began to be produced, and the computing power will make a huge leap in 2027.
Elon Musk's great plan, selling cars is only a small part, software is the key to real profits, to achieve the goal, in addition to having efficient calculations, but also indispensable powerful computing power, in terms of hardware, Tesla has adopted a two-pronged stack, on the one hand, it has bought tens of thousands of NVIDIA H100 GPUs, on the other hand, it has built its own supercomputing chip Dojo, TSMC, which is responsible for the production of supercomputer chips, has now confirmed that the product has begun production.
"Tesla's next-generation Dojo training module has begun production, and by 2027 we will be able to deliver more complex wafer-level systems with more than 40 times the computing power of existing systems. "TSMC's North American Technology Symposium detailed semiconductor technologies and advanced packaging to enable systems to be built at the scale of a wafer to create ultra-high computing performance.
Wafer integration can provide 40 times the computing power
Tesla's Dojo supercomputer is designed by Tesla, the core of which is a training tile, which arranges 25 D1 chips into a 5×5 matrix, uses a 7nm process, can hold 50 billion transistors, provides 362TFLOPs of computing power, and most importantly, has the ability to scale, can be continuously stacked, and allocate computing power and power consumption according to software needs. And that's just the mid-use version of Tesla.
According to TSMC, the new product for Tesla is different from the wafer-level system supplied to Cerebras, in simple terms, the Dojo training module (5×5 processor matrix) is placed on a carrier wafer, after filling all the blanks, TSMC's fan-out (InFO) is used to cover a layer of high-density interconnect, so that the data bandwidth between chips can be maximized, making it operate like a giant chip.
By 2027, TSMC plans to integrate these wafer-level systems with CoWoS advanced packaging technology into a wafer (SoIC), which will provide 40 times the computing power of the complete wafer, more than 40 mask silicon, and up to 60 high-bandwidth memory chips.
According to IEEE Spectrum, InFO_SoW technology is designed to enable high-performance connectivity, allowing Tesla Dojo's 25 chips to work like 1 processor, while TSMC fills in the gaps between chips with virtual chips to keep wafer-level processors consistent.
Tesla's wafer-level Dojo processors actually contain 25 ultra-high-performance processors, which consume a very high amount of power and therefore require a complex cooling system.
To meet the power needs of the Dojo processors, Tesla uses a complex voltage regulation module that provides 18,000 amps of power to the computing plane, dissipating up to 15,000W of heat, so it needs to be water-cooled.
Tesla has yet to reveal the performance of its Dojo wafer system – however, given all the challenges it faced during its development, it seems poised to be a very powerful solution for AI training.
According to Musk, if NVIDIA provides enough GPUs, Tesla probably does not need to develop Dojo on its own, and it is initially estimated that this batch of Dojo supercomputers will become part of Tesla's new Dojo cluster, located in New York, with an investment of at least $500 million.
Tesla's AI career is still full of challenges, despite the strong investment in computing power, two senior engineers in charge of the Dojo project left last December, and now Tesla continues to lay off employees to save costs, and more talented people need to dedicate their brains and livers to have a chance to launch self-driving taxis as scheduled, and try to take FSD to the next level.
Tesla's Dojo next-generation computer will be located in New York, and the headquarters of the Texas Gigafactory will build a 100MW data center, train self-driving software, and use NVIDIA supply solutions for hardware. No matter which side they are, these chips have the same goal, they are basically produced by TSMC, and it is not an exaggeration to call them AI promoters.
Tesla's chairman urged shareholders to re-approve Musk's sky-high compensation package
In a video released Friday, Tesla's board chairman, Robyn Denholm, urged shareholders to reapprove CEO Elon Musk's 2018 compensation package, which was overturned in court by a Delaware judge. Musk's compensation package was once worth as much as $56 billion, but is now worth about $47 billion due to the decline in Tesla's stock price.
Danholm said in the video that the shareholder vote to restore Musk's compensation would be "very important for the company's future."
"We believe that the opinion of one judge should not invalidate the will of all shareholders of the company to vote millions," the chairman told investors. Therefore, we once again ask you to make your voice heard by voting to approve the 2018 Performance Awards. ”
Danholm argued in the video that it was fair for Musk to receive the reward. Shareholders set the bonus to incentivize the tech billionaire to achieve "ambitious financial and operational goals."
She noted that under Musk's leadership, Tesla's revenue grew from $11.8 billion to $96.8 billion and turned a $2.2 billion loss into a $15 billion profit.
The milestones have increased Tesla's value from $53.7 billion to more than $790 billion, she said.
"His success is shared by all shareholders," Danholm said of Musk, adding that investors have seen their stock value rise by 1,100 percent.
*Disclaimer: This article was created by the original author. The content of the article is his personal point of view, and our reprint is only for sharing and discussion, and does not mean that we agree or agree, if you have any objections, please contact the background.