Gen 13 Hardware Specifications
Cloudflare's new Gen 13 server fleet is powered by AMD EPYC 5th Generation Turin-based processors, a significant step up from Gen 12's 96-core Genoa-X architecture. Key improvements include:
- 192 cores with 384 threads (100% core count increase vs. Gen 12's 96C/192T)
- Improved instruction-per-cycle efficiency via Zen 5 architecture enhancements
- 32% better power efficiency per core despite the higher density
- DDR5-6400 memory support for increased bandwidth to feed all cores
The Cache Tradeoff Challenge
Turin's architecture makes a deliberate throughput-over-cache optimization. The highest-density Turin SKU allocates only 384MB of L3 cache across 192 cores (2MB per core), compared to Gen 12's 12MB per core from 3D V-Cache. Initial testing revealed the problem: L3 cache miss rates spiked dramatically, driving memory latency increases of 30-50% at high CPU utilization—unacceptable for Cloudflare's SLA requirements.
Performance counter analysis showed requests were hitting memory (350+ cycle latency) far more frequently than L3 cache hits (50-cycle latency), degrading response times even as throughput improved modestly.
FL2 Rewrite: Unlocking Architecture Potential
Rather than accept the latency penalty, Cloudflare's rewrite of its request handling layer (FL1 → FL2) in Rust eliminated the cache-heavy code patterns. This architectural shift allowed:
- Throughput to scale linearly with core count, delivering the promised 2x performance gains
- Latency to remain within acceptable bounds even at high utilization
- More efficient resource allocation across the system
Combined with targeted optimizations (NUMA-aware core affinity, AMD Platform Quality of Service for per-CCD cache allocation), Gen 13 now delivers substantial TCO improvements without compromising customer experience.
Availability and Impact
Cloudflare is deploying Gen 13 across its edge network, improving compute capacity at the edge for Workers, request handling, and other latency-sensitive services. The 2x performance jump represents a major generational leap driven equally by hardware capability and software optimization.