AMD Ryzen AI Halo Beats DGX Spark in Local Agent Benchmark
Back End News covered AMD benchmark results showing Ryzen AI Halo completing a local enterprise AI agent workflow in 311.6 seconds, ahead of NVIDIA DGX Spark at 367.1 seconds.
AMD's Ryzen AI Halo finished a local enterprise AI agent benchmark 55.5 seconds ahead of NVIDIA's DGX Spark, Back End News wrote, putting the processor side of agentic AI work at the centre of the comparison rather than text generation alone.
The company test ran on AMD's Hermes Executive Presentation Agent, known as HEPA, with a local task that required a presentation for executives, a companion memo, charts and sourced references from information held on the device.
The input set covered 302 documents and mixed stale records, repeated files, contradictions, scans and image-only material.
The AMD Ryzen AI Max+ 395 machine posted a 311.6-second total for the workflow.
NVIDIA's Grace Blackwell GB10-based DGX Spark posted 367.1 seconds, with each platform logging five accepted runs and clearing 25 automated checks.
The workflow split most of the work away from the graphics processor.
Seven of the eight stages ran on the processor, while the GPU handled text generation, leaving document intake, optical text extraction, chunking, evidence retrieval and result checking to the processor-based preparation path.
That preparation path produced the largest gap.
Ryzen AI Halo completed processor-based preparation in 152.1 seconds, compared with 229.9 seconds for DGX Spark, a 34% advantage in what AMD calls CPU orchestration.
Embedding and routing created the widest stage-level split, at 71.5 seconds on Ryzen AI Halo and 139.6 seconds on DGX Spark.
DGX Spark led the text-generation stage, finishing in 137.2 seconds against 159.5 seconds for the AMD system.
The faster generation pass did not overcome the longer preparation stages in the total workflow.
AMD also estimated a 27% lower cost per completed workflow for Ryzen AI Halo.
The calculation amortized hardware over three years of nonstop operation, left electricity out, and produced per-run figures of about $0.0132 for Ryzen AI Halo and $0.0182 for DGX Spark.
The result remains a vendor benchmark for one locally run workload, not a general performance measure across all AI models, applications or system configurations.
For enterprise AI agent buyers, the published test narrows the comparison to whether processor-heavy preparation work matters more than the generation step in their own local workflows.
