AMD's Ryzen AI Halo processor has outperformed NVIDIA's DGX Spark system in agentic workflow benchmarks, delivering faster completion times and lower costs. The results, based on internal testing, position AMD's chip as a strong contender in the growing market for AI agents that operate autonomously.
The benchmark results
In tests focused on agentic workflows—where AI models execute multi-step tasks without human intervention—the Ryzen AI Halo completed workloads more quickly than the DGX Spark. AMD also reported that the Halo system achieved these results at a lower total cost, though specific figures were not disclosed. The company said the performance gap was consistent across several common agentic tasks, including code generation and data retrieval.
What the chips are
The Ryzen AI Halo is AMD's latest processor designed specifically for AI inference at the edge. It integrates a neural processing unit (NPU) alongside CPU and GPU cores, allowing it to run large language models locally. NVIDIA's DGX Spark, by contrast, is a compact AI supercomputer aimed at developers and researchers, built around the company's Grace Hopper architecture. Both systems target workloads that require low latency and high throughput for AI agents.
Why agentic workflows matter
Agentic workflows refer to AI systems that can plan and execute multi-step tasks autonomously, such as browsing the web, filling out forms, or writing and testing code. These workflows demand fast inference and the ability to chain multiple model calls without human oversight. The benchmark results suggest AMD's architecture may be better suited for these real-time, iterative processes than NVIDIA's current offering.
Cost and performance edge
AMD emphasized that the Ryzen AI Halo not only finished tasks faster but also did so at a lower cost per task. This could make it attractive for companies deploying large numbers of AI agents in production environments, where both speed and budget matter. The company did not provide exact pricing or power consumption figures, but the implication is that the Halo offers a better price-to-performance ratio for agentic workloads.
The results come as both AMD and NVIDIA race to capture the AI hardware market beyond training. While NVIDIA dominates training with its GPUs, AMD is betting on integrated solutions that handle inference more efficiently. The DGX Spark, released earlier this year, was NVIDIA's answer to that demand. AMD's benchmark data suggests the Halo may have an edge in the specific use case of autonomous agents.
Neither company has commented on the findings beyond AMD's announcement. Independent verification of the benchmarks has not yet been published. The competition between the two chipmakers is likely to intensify as more companies adopt agentic AI systems.
