AMD's MI355X accelerator achieved 2626 tokens per second per node running Zhipu's GLM5.2 model while costing more than 50% less than Nvidia's Blackwell-based systems. The performance benchmark demonstrates AMD's competitiveness in large language model inference.
AMD's MI355X GPU successfully ran the GLM5.2 large language model at 2626 tokens per second per node, delivering cost-per-inference advantages over Nvidia's flagship Blackwell architecture.
The MI355X, part of AMD's INSTINCT MI300 family, achieved this throughput while maintaining a cost structure that undercuts Blackwell deployments by more than 50%. This represents a significant competitive positioning for AMD in the inference market, where operational costs directly impact profitability for cloud providers and enterprise customers.
GLM5.2, developed by Zhipu AI, is a high-performance language model that has gained adoption among organizations seeking alternatives to models from other vendors. The successful implementation on MI355X indicates that AMD's hardware stack—including ROCm software and optimization tools—can handle advanced model architectures effectively.
The benchmark suggests AMD is closing the gap in the inference performance competition. While Nvidia maintains dominant market share through CUDA ecosystem lock-in and extensive optimization work, AMD's improved cost-to-performance ratio could attract price-sensitive customers and organizations seeking supply chain diversification.
Inference optimization has become critical as AI deployment scales. Major cloud providers and enterprises increasingly prioritize operational efficiency, making hardware choices that deliver similar performance at lower cost strategically valuable. AMD's MI355X performance data provides concrete evidence for procurement discussions.
The result garnered significant attention on technical communities, with the Hacker News discussion attracting 50 comments and 167 points, indicating substantial interest from engineers and infrastructure decision-makers.
AMD continues expanding its AI accelerator portfolio to capture market share in inference and training workloads. The MI355X results suggest the strategy of offering competitive performance-per-dollar metrics is gaining traction in market segments where cost efficiency matters.
Twenty-six former Meta employees are suing the company for allegedly using AI tools to disproportionately target workers on protected leave during mass layoffs. The lawsuit claims Meta's ranking system failed to exclude employees on parental or medical leave.
Superhuman has released an AI-powered auto-draft feature that generates email replies requiring minimal editing. The tool represents a meaningful improvement in practical AI email assistance.
Prime Minister Anthony Albanese announced accelerated approval processes for artificial intelligence projects and datacentres to boost investor confidence and secure Australia's position in the global AI race.
Google is redesigning Images with a Pinterest-style discovery feed and integrating AI image generation into Search when real-world matches aren't available.