Unlocking AI Inference Performance with Intel® Xeon® 6 Processors with Priority Core Turbo

Author: Divakar Mariyanna, Cloud Systems and Solutions Engineer, Intel

As Large Language Models continue to scale, the CPU host node in CPU-GPU AI-accelerated systems has become an important performance component, orchestrating request scheduling, tokenization, KV cache management, and result assembly. Choosing the right AI host CPU is a platform decision that demands transparent, reproducible benchmarks.

A recent report from Signal65 compared AMD EPYC 9575F and Intel Xeon 6960P as AI headnodes using NVIDIA B200 GPUs, concluding that AMD outperforms Intel. Intel reviewed the methodology and ran benchmarks on Intel® Xeon® 6 processors with Priority Core Turbo (PCT) technology, Intel's purpose-built feature for GPU-hosting workloads. This post shares Intel's findings and presents fully disclosed results for community validation.

What is Priority Core Turbo (PCT) Technology?

As AI workloads grow more complex and data-intensive, the CPU host node must sustain peak performance on a small number of latency-critical cores, the cores that service each GPU. Intel Priority Core Turbo (PCT) Technology, available on select Intel® Xeon® 6 processors with P-cores and built on Intel® SST-TF, targets exactly that: it boosts up to 8 designated High Priority (HP) cores up to 4.6 GHz, well above standard all-core turbo limits. In a CPU-GPU AI-accelerated system where one or two CPU cores typically service each GPU, PCT keeps those latency-critical cores running at peak frequency, directly reducing time-to-first-token (TTFT), improving system responsiveness, and sustaining peak GPU utilization across modern AI infrastructure.

Methodology Review of the Signal65 Report

1. Missing parameters limit reproducibility. Signal65's vLLM Bench results for DeepSeek-R1-FP4, Llama-4-Scout-17B-16E-FP4, and Qwen2.5-VL-72B-Instruct report single throughput numbers without disclosing input/output sequence length, request rate, or concurrency, parameters that can shift throughput by 10X or more. Similarly, the GenAI-Perf results omit request rate and concurrency.
2. Output length appears to have no effect on throughput. Signal65 reports GenAI-Perf results across four input/output configurations. Assuming each model was tested at a fixed request rate and concurrency, the published data shows request throughput essentially unchanged when output length varies — for example, GPT-OSS-120B on AMD: 1.720 req/s at 1024/1024 vs. 1.721 req/s at 1024/128. This pattern holds across all four models and both platforms, which is unexpected: an 8X reduction in output length should produce a meaningful throughput gain.
3. The chosen Intel SKU does not include PCT. Signal65 tested the Intel Xeon 6960P (72 cores, 3.9 GHz max turbo), a standard SKU without PCT. Signal65's own analysis notes: "The AI inference tests included in this evaluation involve lightly threaded jobs, which do not benefit from a higher core count. Instead, higher clock frequency and stronger per-core performance become critical." This directly supports the case for PCT-enabled SKUs like the Xeon 6776P, which boosts GPU-serving cores to higher frequencies.

vLLM Benchmark Results: Intel® Xeon® 6 processors with PCT(1)

We tested the same models reported by Signal65 using vLLM on our Intel Xeon 6 PCT platform, varying input length, output length, request rate, and concurrency. Since Signal65 did not disclose vLLM test parameters, we present Signal65’s published numbers alongside Intel’s results at best configurations.

1. DeepSeek-R1 (NVFP4 Quantization)

* in/out: input/output sequence length; RR: request rate; conc: concurrency. Best observed comparison under specific config

Key Takeaway: 

On DeepSeek-R1, Intel® Xeon® 6 processors with Priority Core Turbo delivers 10.7× higher output throughput and 1.5× faster TTFT than Signal65's published AMD result.

2. Qwen2.5-VL-72B-Instruct

* in/out: input/output sequence length; RR: request rate; conc: concurrency. Best observed comparison under specific config

Key Takeaway: 

On Qwen2.5-VL-72B-Instruct, Intel® Xeon® 6 processors with Priority Core Turbo delivers up to 1.58× higher output throughput and up to 8.2× faster TTFT than Signal65's AMD result.

3. Llama-4-Scout-17B-16E (FP4 Quantization)

* in/out: input/output sequence length; RR: request rate; conc: concurrency. Best observed comparison under specific config

Key Takeaway: 

On Llama-4-Scout-17B-16E (FP4 Quantization), Intel® Xeon® 6 processors with Priority Core Turbo are throughput-competitive with AMD and deliver up to ~3× faster TTFT than Signal65's AMD result.

Figure 1. Across different vLLM performance: Intel PCT delivers up to 10.7× higher output throughput and up to 8.2× faster TTFT than Signal65’s AMD results.

GenAI-Perf Benchmark Results(1)

1.  Intel PCT Shows Expected Output-Length Scaling

Intel's PCT measurements show the expected behavior: reducing output length from 1024 to 128 tokens results in a proportional increase in request throughput.

Key Takeaway:

Output-Length scaling factors of 4–8× align with expectations.

While reducing the output length improves throughput by reducing memory-bandwidth constraints, the actual speedup is governed by the fixed-latency anchor of the prefill phase and by framework scheduling overhead in multi-instance or lightweight models.

2.  Benchmark Comparison: Intel® Xeon® 6 with PCT vs. Signal65's Reported AMD Results

The following tables compare Intel PCT against Signal65's AMD numbers at both 1024/1024 and 128/128 configurations for request throughput, token throughput, and TTFT.

a. Request Throughput (req/s) (higher is better)

**Concurrency and Request Rate are applicable for Intel only as no such information was shared in Signal65 paper.

Key Takeaway:

Intel PCT leads request throughput in all eight comparisons, with significant gains appearing at 128/128 for Llama-3.3-70B, Qwen2.5-Coder, and Llama-3.1-8B.

b. token throughput (tok/s) (higher is better)

**Concurrency and Request Rate are applicable for Intel only as no such information was shared in Signal65 paper.

Key Takeaway:

Intel PCT leads token throughput in seven of eight comparisons and is especially strong on Llama-3.3-70B, Qwen2.5-Coder, and Llama-3.1-8B across both configurations.

c. Time to First Token - TTFT (ms) (lower is better) 

**Concurrency and Request Rate are applicable for Intel only as no such information was shared in Signal65 paper.

Key Takeaway:

Intel PCT delivers faster TTFT in six of eight comparisons, with the biggest speedups at 1024/1024 for Llama-3.3-70B, Qwen2.5-Coder, and Llama-3.1-8B.

The large absolute differences likely reflect a combination of factors: different software stacks and the output-length anomaly in Signal65's measurements (where 128-output numbers appear identical to 1024-output). Intel's PCT results demonstrate that using the right SKU for GPU-hosting workloads delivers performance that meets or exceeds the AMD numbers Signal65 published.

Summary

Intel's data confirms Signal65's core observation — higher per-core frequency improves overall CPU-GPU AI accelerated system performance. However, the conclusion differs: Optimizations on Intel® Xeon® 6 with PCT deliver that frequency advantage on the cores that matter, without requiring all cores to run at maximum frequency. Across vLLM and GenAI-Perf benchmarks with matching GPU configurations (8× B200), Intel PCT matches or exceeds AMD's published results on throughput and delivers consistently faster TTFT — the most user-perceptible latency metric.

Learn More

Contact your Intel representative to discuss the value of Intel Xeon 6 processors with PCT for advanced AI workloads. And learn more about Intel Xeon 6 processors SKUs with PCT on Intel’s website:

Intel Xeon 6776P processor

Unleash the Power of Intel Xeon 6 Processors as AI Host CPUs with Priority Core Turbo

Priority Core Turbo Technology (PCT Technology) Technical Article

Endnotes

(1)Test System and Software Configuration: 1-node, Supermicro X14DBG-LC+, 2× Intel® Xeon® 6776P (64 cores, 350W TDP, PCT enabled, HT On, Turbo On), 8× NVIDIA B200, 1024 GB DDR5 6400 MT/s (16×64 GB), BIOS 1.5, microcode 0x1000405, 2× Intel X710 10GBASE-T, Ubuntu 24.04.4 LTS, kernel 6.8.0-111-generic. vLLM-v0.18.1, AI PERF-0.6.0.post1 Tested by Intel as of May 19, 2026.

Notices and Disclaimers

Performance varies by use, configuration, and other factors. Learn more on the Performance Index site.
Performance results are based on testing as of the dates shown in configurations and may not reflect all publicly available ​updates. See backup for configuration details. No product or component can be absolutely secure.
Your costs and results may vary.
Intel technologies may require enabled hardware, software, or service activation.
© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries. Other names and brands may be claimed as the property of others.