Review of AI Chips from AMD, Google, and Tesla

This article reviews the AI chips produced by AMD, Google, and Tesla.

Recently, there have been discussions about Tesla's potential to surpass AMD in AI chip performance and volume of chips, particularly with the launch of the Dojo 3 chip.

Tesla's Dojo Chips

The top line of Tesla's Abundance slide emphasizes AI computation alongside projects like Optimus, Robotaxi, and Full Self-Driving (FSD). This highlights the significance of the Dojo 2 and Dojo 3 AI training chips, which play crucial roles in perfecting FSD capabilities and enhancing the training of the Optimus robot.

It is expected that the Dojo 3 chip will significantly enhance Tesla's AI capabilities, possibly placing it ahead of AMD in terms of both performance and production volume.

AMD's Performance in AI Chips

In 2024, it is estimated that AMD shipped around 300,000 to 400,000 units of its Instinct MI300 chips, generating about $5 billion in revenue for the year. To put this into perspective, the average selling price (ASP) per chip can be calculated as follows:

• $5 billion ÷ 300,000 units = ~$16,667 per chip
• $5 billion ÷ 400,000 units = ~$12,500 per chip

Looking forward, AMD aims to sell around 500,000 AI chips in 2025, which could bring in approximately $7.5 billion in revenue.

Nvidia's AI Dominance

Nvidia's data center revenue acts as a good indicator of its AI chip sales. Analysts project Nvidia's data center revenue to reach $110.36 billion in 2024, with expectations of growing to about $120 billion in 2025. The number of chips sold largely depends on the ASP. Nvidia's H100 GPUs sell for between $20,000 and $40,000, and if we assume an average ASP of $30,000, we can estimate:

$120 billion ÷ $30,000 = 4 million chips.

Google's TPU Production

Google develops Tensor Processing Units (TPUs) primarily for its internal use. In 2024, total shipments for self-developed cloud AI ASIC accelerators, which includes TPUs, are expected to be around 3.45 million units, with Google holding a 74% market share, producing approximately 2.55 million TPUs. With an anticipated 20% market growth in 2025, total shipments could reach an estimated 4.14 million units, with Google likely maintaining its 74% share:

4.14 million × 0.74 = approximately 3.06 million TPUs.

As for performance, the current TPU v4 offers 275 teraFLOPs and the next generation, TPU v5e, provides 197 teraFLOPs. The upcoming TPU v6, expected by 2025, might achieve around 400 teraFLOPs.

Future of Tesla's AI Chips

Tesla's Dojo 2 chips are anticipated to enter high-volume production by late 2025, with Tesla investing around $500 million annually. Assuming an ASP of about $10,000 per chip, this investment could yield:

$500 million ÷ $10,000 = 50,000 chips.

Dojo 2 chips are projected to be ten times more powerful than their predecessor, Dojo 1, potentially reaching around 3-4 petaflops. The Dojo 3, expected in 2026, might deliver a further tenfold increase in performance, aiming for around 40 petaflops, which would put Tesla on par with Nvidia's offerings.

If Tesla's Dojo 3 chip performs successfully, especially with major customers in XAI and Tesla itself, it could surpass AMD, claiming the position of the second-best in AI chip production.

This analysis highlighted the competitive landscape of AI chip manufacturing among major firms and the potential shifts the industry might experience in the coming years.

AI, Chips, Tesla