Inside Gen 13: How the Most Powerful Server Ever Was Built

As Cloudflare's request-processing architecture changed with the FL2 transition, the company designed a new server generation to match: Gen 13. Its core specs are AMD EPYC 9965 Turin 192-core, 768GB DDR5-6400 memory, 24TB PCIe 5.0 NVMe storage, and dual 100GbE networking.

Compared to Gen 12, Gen 13 achieves the following goals:

• 2x throughput improvement while maintaining latency SLAs
• Up to 50% improvement in performance-per-watt
• Up to 60% higher throughput per rack while keeping rack power the same
• 2x memory, 1.5x storage, and 4x network capacity
• Added hardware support for PCIe encryption
• Improved support for hot PCIe accelerators

The key factor in CPU selection was core count rather than L3 cache. Cloudflare compared Turin 9755, 9845, and 9965, and confirmed that FL2 workloads have low dependence on L3 cache and scale almost linearly with increased core count. As a result, the 9965 showed the highest aggregate RPS and good performance-per-watt, and was ultimately selected.

The selection rationale breaks down into four points:

• FL2 resolved the L3 cache bottleneck. Gen 12's 9684X had 12MB of L3 per core, but Gen 13's 9965 was sufficient even with 2MB per core.
• TCO was better. 192 cores delivered higher density and power efficiency.
• Operations became simpler. The same compute can be provided with fewer, higher-density servers.
• Support for DDR5-6400, PCIe Gen 5.0, and CXL 2.0 gives the platform a long lifespan.

Memory is configured as 12-channel 1DPC, filling every channel. Cloudflare chose 12x 64GB DDR5-6400 ECC RDIMM for a total of 768GB, a choice that doubles total memory capacity versus Gen 12 while maintaining the existing ratio of 4GB/core. A smaller 576GB configuration was judged insufficient, while a 1152GB configuration was judged excessive in terms of cost and power.

Cloudflare also used dual-rank 2Rx4 DIMMs to boost memory throughput. Based on 384 hardware threads, this secures 614GB/s of peak bandwidth, providing roughly 1.6GB/s per thread; since current workloads are not bound by memory bandwidth, this headroom remains as room for future growth.

Storage was switched to PCIe 5.0 E1.S NVMe, and the number of slots was increased from 2 to 3, expanding total capacity from 16TB to 24TB. At the same time, front drive bays were added, allowing up to 10x U.2 PCIe 5.0 NVMe drives to be installed, and enabling flexible operation of compute SKUs and storage SKUs within the same chassis.

The target operational lifespan is 5 years, with a drive endurance target of 1 DWPD. The Samsung PM9D3a and Micron 7600 Pro drives used both meet this requirement, and support NVMe 2.0 and the OCP NVMe Cloud SSD Specification 2.0, enabling use of ZNS, Simple Copy Command, and stronger telemetry and lockdown features.

Networking moves up from Gen 12's dual 25GbE to Gen 13's dual 100GbE, tuned to handle greater throughput while keeping the power budget unchanged. Overall, Gen 13 is a server that packs more cores, wider memory bandwidth, larger storage, and faster networking into a single platform built for the FL2 era.