Principled Technologies PT has released new report examining how Dell PowerEdge XE9780LAP server with dual Intel Xeon 6979P processors performs on agentic AI workloads. PT tested server with eight NVIDIA B300 GPUs and in CPU-only configuration using custom agentic AI benchmark that runs realistic manufacturing financial services and coding workflows.
Agentic AI systems plan tasks retrieve information invoke tools and reason through multi-step workflows with minimal human intervention. As report notes agent completing single task may require hundreds or thousands of interactions between AI models orchestration frameworks tools and data sources. That makes infrastructure sizing key decision for organizations adopting agentic AI.
Key takeaways more concurrent agents with eight NVIDIA B300 GPUs server supported 205 AI agents on early-stage manufacturing workload 62 on late-stage financial services workload and 14 on complex code generation workload while meeting SLA thresholds. High throughput server delivered up to 3,139 output tokens per second on manufacturing workload. Favorable tokenomics cost per million tokens ranged from $0.25 for coding workload to $2.05 for complex mixture-of-experts MoE financial workload based on server purchase price and energy costs.
What role do CPUs play in agentic AI In PT study Intel Xeon 6979P processors served as orchestration engine managing agent control flow invoking tools enforcing policies and moving data while also handling agents local work. NVIDIA B300 GPUs accelerated model inference and token generation.
Can organizations run agentic AI without GPUs CPU-only results show CPU-only deployment makes practical sense for earlier-stage or lower-volume agentic AI use giving organizations starting point before adding GPUs. With Intel Advanced Matrix Extensions Intel AMX enabled two Intel Xeon 6979P processors supported 57 agents on moderate financial workload within 30-second SLA. With Intel AMX disabled at FP32 precision system could not sustain even one agent within SLA.
Which workloads did PT test PT tested eight manufacturing tasks eight financial services tasks and 16 coding tasks ranging from predictive maintenance and fraud investigation to autonomous bug fixes in real code repository.
Hardware that suffices for early cautious agentic use will not be enough once organization trusts its agents to do more. More work organization lets its agents perform more hardware it will need.
Report concludes Dell PowerEdge XE9780LAP servers with Intel Xeon 6979P processors and NVIDIA B300 GPUs poised to meet agentic AI infrastructure needs for organizations of many types and sizes.
About Principled Technologies, Inc.
Principled Technologies, Inc. is the leading provider of fact-based marketing and competitive analysis backed by transparent methodologies.
