The explosive growth of the popularity of ChatGPT and other solutions based on generative AI caused an unprecedented demand for computing power, which led to a deficiency of AI trackers, writes Digitimes. NVIDIA occupies the lion’s share of the and chip market, and the leading suppliers of cloud services, such as Google, Amazon and Microsoft, are actively engaged in projects to develop their own accelerators, trying to reduce their dependence on external supplies.
ASIC is becoming increasingly popular among large cloud providers, as they seek to optimize the chips for their specific requirements, Digitimes noted. ASIC provides high performance and energy efficiency in a narrow range of tasks, which makes them an alternative to universal NVIDIA accelerators. Despite the dominance of NVIDIA in the market, the high energy consumption of its chips in combination with a high cost allowed ASIC to occupy a competitive niche.
ASIC is especially good for the training and information of II Models, offering significantly higher performance indicators in the recalculation of 1 watts compared to GPU general purposes. ASIC also provide customers with greater control over their technological stack. In the ASIC development market, the main competitors are Broadcom and Marvell, which use different technologies and strategic approaches.
Source of Images: Digitimes
Marvell strengthened its position in the market, in particular, thanks to the partnership with Google in the development of server ARM chips, while expanding strategic cooperation with its main client-Amazon. The Google TPU V6E is the most advanced ASIC among chips developed by four leading cloud providers, approaching in performance to H100. However, it is still behind NVIDIA accelerators for about two years, says Digitimes. Created by Marvell and Amazon, the Trainium 2 Accelerator in performance is between the NVIDIA A100 and H100.
During the last report on the financial results, Marvell shared a prognosis of significant increase in ASIC revenue, starting in 2024 (2025 financial year), due to Trainium 2 and Google Axion. In particular, the Marvell Inferential Asic project jointly with Amazon is supposed to be launched into mass production in 2025 (2026 financial year), while Microsoft Maia, as expected, will begin to bring income from 2026 (2027 financial year).
According to Morgan Stanley, although the Marvel business of custom chips is a key driver of the growth of his unit for decisions for the data center, it also carries significant uncertainty. Short -term forecasts of Morgan Stanley for Marvell/Trainium products are positive, which is confirmed by the increased TSMC capacities for the COWOS packaging method, Amazon plans for expanding production and Marvell confidence in market demand.
However, in the long run, the competitive environment creates problems. The emergence of companies like WorldChip Electronics in the computing chips sector can force Marvell to reorient on network solutions. In addition, the potential reduction in profit from Trainium after 2026 means that Marvell will need to ensure the launch of new projects to maintain growth dynamics, analysts say.
Broadcom and Marvell are examples of different development strategies in the ASIC sector, Digitimes notes. Broadcom gives the priority of large -scale integration and design of platforms, reinforcing its approach with significant investments in R&D and complex technological integration. In turn, Marvell develops due to strategic acquisitions, for example, Cavium, Avera and Innovium, which expands its portfolio of technology.