Nvidia's amplification move, Tesla's computing power chip, completely cold?
The advent of the era of artificial intelligence has promoted the rapid development of computing chips. As the two giants in the field of AI chips, the competition between Nvidia and Tesla has attracted much attention. Recently, NVIDIA has launched new AI chips such as B200 and GB200, which have greatly improved computing power and energy efficiency, which is regarded as a serious challenge to Tesla's self-developed computing power chip Dojo.
NVIDIA's new B200/GB200 computing power and energy efficiency have been greatly improved
At the GTC conference in March this year, Nvidia released new AI chips such as the B200 and GB200. These chips are manufactured on a 4nm process and have a single-precision floating-point operation capability of up to 1.5 PetaFLOPS, which is more than 3 times higher than the previous generation. The power consumption of GB200 has also been greatly optimized, with the power consumption of GB200 being only 300W, and the energy efficiency ratio has been improved by 2.5 times.
This leap in performance is mainly due to NVIDIA's innovations in technologies such as tensor core and streaming multiple data sampling. Industry insiders believe that the launch of Nvidia's new products may change the "rules of the game" for AI chips.
It poses a major challenge to Tesla's Dojo computing power chip
The outstanding performance of NVIDIA's new products undoubtedly poses a major challenge to Tesla's self-developed Dojo computing power chip. Tesla has been planning to gradually replace Nvidia's GPU with its self-developed D1 chip. According to published information, the performance of the D1 chip is at least 4 years behind the performance of Nvidia products of the same level, and the heat generation is also greater.
Tesla CEO Elon Musk has said that if Nvidia can provide enough GPUs, they may not need Dojo. And the emergence of Nvidia's new products has undoubtedly made Musk's idea more firm. Many believe that it is difficult for Tesla to compete head-on with Nvidia's new products in the short term.
Tesla's self-developed D1 chip lags behind
The performance is at least 4 years behind the same level of products
According to the published information, Tesla's self-developed D1 computing power chip is at least 4 years behind Nvidia products of the same level in terms of performance. The D1 has a single-precision floating-point power of about 9 PetaFLOPS, while Nvidia's new GB200 has reached a staggering 1.5 PetaFLOPS.
Not only that, but the design of the D1 chip also has some flaws. Due to the ring design, the memory bandwidth of the D1 is limited and the potential of computing power cannot be fully realized. The D1 also lacks a dedicated hardware acceleration unit like Nvidia's Tensor Core, which will further lag behind in performance in some AI application scenarios.
The fever is greater
In addition to its lack of performance, the D1 chip also has significant drawbacks in terms of heat generation. Due to the use of the older 16nm process, the heat generation of the D1 is about 30% higher than that of Nvidia products in the same period.
Excessive heat generation will not only affect the stability and service life of the chip, but also increase the cost of heat dissipation and cooling. This is undoubtedly a heavy burden for Tesla's self-driving system.
It is difficult to compete with NVIDIA's new products in the short term
Based on the above factors, many believe that it is difficult for Tesla to compete head-on with Nvidia's new products in the short term. Even if the D1 chip eventually goes into production, its performance and energy efficiency will lag far behind Nvidia's peers.
Tesla seems to be aware of this, too. Musk has said that if Nvidia can provide enough GPUs, they may not need Dojo. This shows from the side that Tesla does not have to develop its own chips.
Tesla prepares with both hands
Hoarding a lot of NVIDIA GPUs
To prepare for future demand, Tesla has stockpiled a large number of Nvidia GPUs. According to reports, Tesla had already ordered $1 billion worth of Nvidia GPUs at the end of 2022 for its self-driving system.
This approach is undoubtedly a kind of insurance for Tesla's prospects for self-developed chips. If the Dojo project ultimately fails, Tesla will at least have Nvidia GPUs to use.
The Dojo project leader was fired and the project was at an impasse
The outlook for the Dojo project is not encouraging. In March of this year, Ganesh Venkataramanan, a senior Tesla employee in charge of the Dojo project, was fired. This was seen by the outside world as a sign that Tesla had lost confidence in the prospects of the Dojo project.
According to people familiar with the matter, the Dojo project is currently at an impasse. Tesla's top management is divided on the future direction of the project, and the pressure of Nvidia's new products makes the prospects of the Dojo project even more uncertain.
Promote the application of relevant technologies
Liquid cooling technology
In addition to the impact on Tesla, the emergence of new NVIDIA products will also promote the application of some related technologies. Due to the high power consumption, NVIDIA's new products have more stringent requirements for heat dissipation and cooling. Liquid cooling technology is expected to be more widely used in scenarios such as data centers.
Liquid cooling technology can take away heat more efficiently than traditional air cooling, so that the chip can obtain a better heat dissipation environment. Liquid cooling systems are also more expensive to build and maintain, which will be a factor for data center operators to weigh.
Optical transceiver technology
In order to give full play to the computing power potential of new products, optical module technology will also become a development direction in the future. Optical modules can greatly increase the bandwidth between the chip and the memory, thereby accelerating the transmission of data.
NVIDIA has adopted optical-based NVLink technology in its new products to improve the communication capabilities between chips. Optical module technology is expected to be further promoted to help improve AI computing power.
NVIDIA's latest AI chip, with its excellent computing power and energy efficiency, poses a serious challenge to Tesla's self-developed Dojo computing power chip. Tesla's self-developed chip D1 is far behind Nvidia's new products in terms of performance and heat, and it is difficult to compete with it head-on in the short term.
Tesla seems to be aware of this, too, and has stockpiled a lot of Nvidia GPUs to prepare for both. The Dojo project leader was fired, and the project was at an impasse, leaving the future uncertain.
In addition to the impact on Tesla, the emergence of NVIDIA's new products will also promote the application of related technologies such as liquid cooling and optical modules, bringing new development opportunities to data centers.
NVIDIA's "amplification move" this time will undoubtedly change the development pattern of AI chips. Whether the prospect of Tesla's self-developed computing power chips is "completely cool" remains to be tested by time. But what is certain is that Nvidia has temporarily gained the upper hand in this battle for computing power.