Home / Technology / Old chips power massive AI models, Nvidia/AMD challenged
Old chips power massive AI models, Nvidia/AMD challenged
11 May
Summary
- Skymizer's card runs 700B parameter models on single device.
- It uses older 28nm chips with standard memory, consuming 240W.
- Performance claims challenge high-power Nvidia and AMD accelerators.
A Taiwanese firm, Skymizer, has introduced an AI accelerator card that could disrupt the market dominated by Nvidia and AMD. The HTX301 card utilizes older 28nm chip technology and standard LPDDR memory, yet it boasts the capability to run language models with up to 700 billion parameters. This impressive feat is achieved with a remarkably low power consumption of just 240 watts.
The HTX301 card integrates six HTX301 chips, offering up to 384 GB of memory capacity. Skymizer claims its HyperThought platform, featuring next-generation LPU IP, delivers efficient weight and KV cache compression. This design reportedly outperforms open-source solutions like llama.cpp and consumes less than half the power of leading PCIe AI accelerators from competitors.
Skymizer positions the HTX301 as a solution for enterprises seeking on-premises AI infrastructure without the need for expensive hyperscale GPU clusters. The card fits into standard air-cooled servers, addressing privacy concerns and unpredictable costs associated with cloud-based AI. The company suggests the era of requiring massive GPU infrastructure for ultra-large LLMs may be ending.
The HTX301 is slated for a preview at Computex, where independent verification of its performance claims will be possible. If successful, the technology could dramatically reduce the entry cost for on-premises AI deployments, offering a viable alternative to the power-hungry offerings from AMD and Nvidia.
