A New Server Generation, Built for Speed and Scale

Cloudflare’s network philosophy has always centered on the idea that every server should be able to run every service. This design principle allows the company to grow capacity rapidly, shift traffic efficiently, and avoid the resource fragmentation that comes with specialized hardware. By relying on commodity components and software-defined load balancing, Cloudflare can optimize performance for its highest-tier customers while also leveraging idle capacity across its global footprint of 200 cities.

The alternative—building separate, purpose-built networks for services like the Firewall, DDoS protection, or Workers—would create wasted idle resources and reduce flexibility. A single, unified hardware target simplifies software development and ensures that security and performance improvements apply network-wide. Anycast routing sends each request to the nearest data center, maximizing both performance and the surface area available for attack mitigation. Once a request lands, the custom Unimog load balancer distributes it dynamically across servers of different generations, balancing at multiple layers: between cities, physical deployments, ports, cables, servers, and even individual CPU threads.

Scaling is straightforward: add servers, PoPs, or entire cities to the pool. Faulty components are automatically de-prioritized and repaired in batches by operations teams, which is why no dedicated Cloudflare staff are stationed at any of the 200 locations.

Gen X: A Shift to AMD

The tenth generation of Cloudflare servers, dubbed “Gen X,” is now live in major US cities and rolling out globally. Compared to the Gen 9 servers, Gen X processes up to 36% more requests at a substantially lower cost. The new hardware also delivers a ~50% reduction in L3 cache miss rate and up to a 50% decrease in NGINX p99 latency. All of this comes from a CPU with a 25% lower thermal design power (TDP) per core.

Cloudflare’s Gen X: 
Servers for an Accelerated Future Embedded Image - xEA6hm

The most notable change is the absence of Intel. For the first time, no major server component—CPU, board, memory, storage, NIC, or accelerator—comes from Intel. This marks a clear departure from previous generations, which relied heavily on Intel hardware. The pendulum of technology leadership swings, and this time, AMD is inside.

Cloudflare’s Gen X: 
Servers for an Accelerated Future Embedded Image - ncwMGt

Cloudflare selected the 2nd Gen AMD EPYC 7642 processor in a single-socket configuration for Gen X. The 48-core (96-thread) CPU runs at a base clock of 2.4 GHz with a 256 MB L3 cache. While the 225W TDP might seem high, it is lower than the combined TDP of the dual-socket Intel configurations used in Gen 9. The performance of this CPU outweighed the appeal of lower-power variants, and AMD’s 64-core option didn’t offer compelling gains for Cloudflare’s specific software stack.

Cloudflare’s Gen X: 
Servers for an Accelerated Future Embedded Image - LVR0fJ

The EPYC 7642 has proven considerably more powerful than the dual-socket pairs of high-core-count Intel Skylake and Cascade Lake processors used in the previous generation. This performance has encouraged Cloudflare to accelerate its server replacement cycle, swapping out multiple generations of Intel hardware ahead of schedule.

Cloudflare’s interest in ARM processors remains, though no current off-the-shelf ARM CPU meets its needs. The company ported its entire software stack to ARM ahead of Qualcomm’s Centriq launch, which was later shuttered, and continues to maintain that port. Looking ahead, Cloudflare is optimistic about high-core-count ARM offerings expected in 2020 and beyond, and envisions a future server fleet mixing x86 (both Intel and AMD) with ARM.

Compute is the largest investment in a Gen X server, reflecting the nature of Cloudflare’s workloads. From the Firewall to Workers, the company’s services are compute-heavy and often constrained by CPU before any other resource. The average web request is small in kilobytes, influenced by the popularity of APIs and mobile applications. This design point differs from traditional CDNs focused on large object delivery, where storage might take priority and re-architecting for serverless would be prohibitively expensive.

Gen X also includes an intentionally empty PCIe slot, leaving room for a future accelerator—be it a GPU, FPGA, SmartNIC, custom ASIC, TPU, or something else—if a card can perform specific functions more efficiently than the primary CPU.

Memory, Storage, and Form Factor

Because CPU is the usual bottleneck, Gen X gave Cloudflare the opportunity to scale RAM and SSD capacity more conservatively than compute. Memory remains at 256GB, now rated at 2933MHz. Storage stays around 3TB, but moves from SATA to a 3x1TB NVME flash configuration, delivering increased IOPS and higher endurance. This enables full disk encryption via LUKS without performance penalties. The network card remains a Mellanox 2x25G NIC.

Cloudflare’s Gen X: 
Servers for an Accelerated Future Embedded Image - d3iVgm
Cloudflare’s Gen X: 
Servers for an Accelerated Future Embedded Image - JmMoeT

The server design shifts from a multi-node chassis back to a simpler 1U form factor, which is lighter and less error-prone during data center operations. Cloudflare has also added multiple new ODM partners to diversify manufacturing and take advantage of additional global warehousing.

Cloudflare’s Gen X: 
Servers for an Accelerated Future Embedded Image - tdlfAT

Network Expansion Ahead

Cloudflare’s Gen X: 
Servers for an Accelerated Future Embedded Image - GMOnBi

The new generation of servers provides the flexibility to continue building out Cloudflare’s network closer to every user on Earth. The company credits its engineering teams and partners for the work on Gen X, and more technical details are expected in follow-up posts.