Intel Xeon 6 658X for Workstations Launching Now.. but how's it stacking up?

TL;DR

The Intel Xeon 6 658X is a compelling 24-core workstation CPU with exceptional memory bandwidth and efficiency, but limited per-core performance versus AMD Zen 5.

Key Points

• 1.The Xeon 6 658X is a 24-core workstation CPU priced around $1,800. It offers 48 threads, 8-channel DDR5-6400, 128 PCIe 5 lanes, 144MB cache, and 250W base / 300W boost power, positioned as Intel's answer to Thread Ripper.
• 2.The ASUS ProWS W890E-SAGE SE motherboard is the launch platform for this CPU. It features 8 memory channels, 7 PCIe x16 slots, four PCIe Gen 5 M.2 slots, dual PSU support, Thunderbolt, BMC with dual 10GbE, and a dedicated memory cooling fan to address DDR5 thermal throttling.
• 3.Memory bandwidth is the 658X's standout advantage over Thread Ripper. Theoretical peak is ~409 GB/s at DDR5-6400 JEDEC; with overclocking to DDR5-7200 or 9600, real-world bandwidth exceeds 430–500+ GB/s, outperforming even the 96-core Thread Ripper 9995WX by 25%.
• 4.Per-core performance improves only 5–7% over Sapphire Rapids Refresh, but power efficiency is dramatically better. Under sustained load, the 658X delivers 50–60% better performance-per-watt versus the older Xeon W3500 generation, representing one of Intel's largest ever generational efficiency gains.
• 5.In AI and CPU inferencing workloads, the 658X excels thanks to AMX extensions. It achieves 29+ tokens/second on Llama 3.1 8B Q8 and over 258 tokens/second prompt processing — more than twice as fast as competing Thread Ripper CPUs in those specific benchmarks.
• 6.PCIe peer-to-peer bandwidth nearly doubles versus the previous generation. With four RTX Pro 6000 GPUs, the 658X delivers 63 GB/s real-world P2P throughput compared to only ~30 GB/s on Sapphire Rapids Refresh, removing the platform as a bottleneck for multi-GPU workloads.
• 7.The 658X underperforms Thread Ripper in general-purpose multi-threaded workloads due to core count. Code compilation (LLVM: 215s), HPC tasks, and OpenFOAM all trail 32–64 core Thread Ripper parts, making this CPU best suited to memory-bandwidth-heavy and AMX-accelerated workloads rather than raw throughput.
• 8.Intel's 24-core SKU makes the platform feel underutilized rather than maxed out. Unlike most lower-core-count workstation chips that hit CPU limits before platform limits, the 658X's memory controller and IO headroom suggest that higher-core-count Xeon 6 variants scaling this architecture will be significantly more competitive.