Amazon is hoping to gain an edge with its AI chips over NVIDIA's Blackwell GPUs through power efficiency and performance upgrades. NVIDIA's Blackwell GPUs are the highest-end AI chips in the market. However, their price, partially fueled by tight supply resulting from high demand, estimated to range between $30,000 to $40,000, has been a major constraining factor. NVIDIA CEO Jensen Huang has dismissed concerns of custom AI chips eating away at his firm's market share as he is confident about the performance advantage offered by his products. However, Amazon aims to deliver superior power efficiency with its Trainium chip lineup, according to remarks AWS Senior Director for Customer and Product Engineering Gadi Hutt to CNBC.
Amazon Touts Trainium2 Chip's Cost-Performance Advantages Over NVIDIA's Blackwell
The cost of NVIDIA's Blackwell chips has been one of the most widely discussed aspects of the AI GPUs. Analysts have suggested that a single Blackwell AI GPU costs at least $30,000 and the Blackwell GB200 chip costing between $60,000 to $70,000. Consequently, these chips, which are already in limited supply, cannot be bought in large quantities by most AI companies, which limits the spread of AI training and inferencing software.
On this front, Amazon believes its Trainium2 chips offer better performance when their cost is analyzed. Amazon, like several other big technology firms, self-develops custom AI chips. These chips are primarily used in-house and in tandem with advanced chips from NVIDIA. Amazon's Trainium was first launched in 2020, and Amazon followed up with Trainium2 in December of last year.
The firm claims that Trainium2 offers four times the performance of the first-generation chips and up to 40% better price performance over AI GPUs present in the market at the time.
Amazon's Gadi Hutt, speaking to CNBC, shared details about the firm's upcoming Trainium3 chips. He believes that the chip is going to save energy "by an additional 50%" over the Trainium2 and double the performance of the predecessor. Amazon's Amazon Web Services (AWS) has outlined that its Trainium2 GPUs have trained Amazon-linked Anthropic's Claud Opus 4 AI model. These chips are also powering Anthropic's Rainier supercomputer.
The Trainium chips also offer better cost performance over NVIDIA's Blackwell GPUs, Hutt outlined, while Blackwell is still the chip with the best performance. NVIDIA's AI GPUs are the backbone of AI computing, with firms such as Oracle having acquired these chips to rent out AI computing capacity to software companies.
Amazon's AWS business also relies on renting cloud capacity. The firm is pushing its Trainium chips for AI computing to customers as well, as it seeks to play up the cost advantages offered by the in-house semiconductors. Apart from Amazon and Trainium, other firms are also involved in designing custom AI chips.
Chip designers Marvell and Broadcom have confirmed these plans, while Alphabet already offers its TPU chips for AI computing. NVIDIA's CEO Jensen Huang, on the other hand, has brushed off concerns about custom chips eating market share as he believes NVIDIA's products are market leaders in performance.
About the author: Ramish is a seasoned technology writer and editor with more than a decade of experience. He specializes in semiconductor fabrication and market analysis. With a background in finance and supply chain management - via his bachelors in Finance and a micromasters in supply chain management from MIT - Ramish combines financial rigor with deep industry insight to deliver accurate and authoritative coverage.
Follow Wccftech on Google to get more of our news coverage in your feeds.
