This week, insights into the evolving landscape of datacenter technology have emerged, revealing a shift in the dominance of GPUs. While these powerful processors have long ruled the computing realm, their monopoly is being challenged by innovative alternatives. The latest episode of <span class="tertiary colormobiletertiary font-weight-bold m-font-weight-bold”>The Kettle delves into these developments, featuring Systems Editor Tobias Mann and Microsoft Ecosystem Reporter Richard Speed discussing the implications of these changes.
Emerging Alternatives to Traditional GPUs
As the AI sector continues to expand, many companies are exploring options beyond traditional GPUs. Mann highlights Cerebras, a company that has been tackling AI infrastructure challenges for years. Founded by Andrew Feldman, Cerebras distinguishes itself with chips that are not just GPU-sized but rather the size of dinner plates. This design allows them to pack an impressive 44 gigabytes of ultra-fast SRAM, circumventing the supply constraints of high-bandwidth memory.
In 2024, Cerebras launched an inference server capable of processing tokens faster than any GPU configuration. Their approach leverages structured sparsity to optimize compute power, and they have recently partnered with AWS and AMD to enhance their capabilities. Notably, Feldman has openly criticized Nvidia, referring to it as an “AI arms dealer” for its aggressive GPU sales strategies.
Innovations in AI Hardware
The latest generation of Cerebras chips, known as the CS-4 systems, showcases a modular architecture that allows for easier maintenance and improved power delivery. This advancement enables double the performance and bandwidth compared to previous models. As AI companies seek efficient solutions for generating premium tokens, Cerebras positions itself as a viable alternative to Nvidia, particularly for clients like OpenAI.
Meanwhile, Google is also pivoting towards custom silicon, moving away from traditional GPUs. The tech giant has been developing its Tensor Processing Units (TPUs) since 2015, primarily for internal workloads. Recently, Marvell has emerged as a competitor to Broadcom in providing IP for custom chips, indicating a growing trend among cloud operators to develop tailored solutions for their specific needs.
Waymo’s Custom Chip Development
In an unexpected move, Waymo has announced the creation of a custom machine learning accelerator for its autonomous vehicles. This decision stems from the need to process vast amounts of sensor data with minimal latency. While off-the-shelf options exist, Waymo’s shift to a custom solution aims to enhance compute density and streamline development, addressing the limitations of their previous use of FPGAs.
Windows 11: A Long-Awaited Return to User Preferences
On the software front, Microsoft is finally addressing user feedback regarding Windows 11. After nearly five years, the company is reintroducing the ability to move the taskbar—a feature that many users have missed since its removal. Richard Speed notes that this change reflects Microsoft’s renewed focus on user experience, as the company aims to win back its audience by listening to their requests.
Speed emphasizes that while the return of the taskbar customization is a positive step, the delay raises concerns about Microsoft’s responsiveness to user needs. The company is also experimenting with context menus that allow for greater customization, echoing sentiments from users who appreciated the functionality of previous Windows versions.
As the conversation shifts to the broader implications of these developments, Mann points out that the competitive landscape has changed significantly since Windows 11’s launch. With alternatives like macOS and Linux gaining traction, Microsoft is compelled to adapt and enhance its offerings to retain its user base.
In summary, the landscape of AI hardware is diversifying, with companies like Cerebras and Waymo leading the charge in innovation. Concurrently, Microsoft is taking steps to realign Windows 11 with user expectations, signaling a potential shift in how tech giants respond to consumer demands in an increasingly competitive market.
