Rethinking the Rack Unit: Why Cloudflare's Next Server Gets Taller

Cloudflare is preparing its next x86 server fleet refresh, slated for 2024. The Gen 12 platform isn't just a CPU swap; it involves evaluating the entire server stack—memory, storage, security, and the physical chassis itself. A critical driver for this redesign is a simple physics problem: the latest x86 processors run significantly hotter, and the traditional 1U server footprint is hitting its cooling limits.

The platform will need to support the power demands of modern AMD and Intel silicon while also accommodating the GPUs used for Workers AI inference. The hardware team surveyed the current landscape—AMD's Genoa, Bergamo, and Genoa-X (Zen 4) alongside Intel's Sapphire Rapids—and found a common trend: rising thermal design power (TDP) across generations.

TDP represents the maximum heat a cooling system must dissipate under load, and it has climbed steadily since 2014. Cloudflare's Gen 9 (Intel Skylake) ran at 150W, Gen 10 (AMD Rome) at 240W, and Gen 11 (AMD Milan) also at 240W. Current top-tier SKUs push much higher: AMD EPYC 9004 series defaults reach 360W (configurable to 400W), and Intel Sapphire Rapids defaults hit 350W. This trend is expected to continue into the next generation of x86 products.

The 1U Fan Dilemma

Previous Gen 10 and Gen 11 servers used a 1U1N (one rack unit, one node) air-cooled design to maximize rack density. That form factor can fit a maximum of eight 40mm dual-rotor fans. Accounting for a 40°C ambient room temperature, that configuration can support CPUs with a TDP of roughly 400W—but only with fans spinning at 100% duty cycle.

Running those fans at full speed is expensive. A single fan consumes about 40W; a typical server with seven to eight fans can burn 280–320W on cooling alone. At peak load, a single node could eclipse 750W total system power. The result is twofold: rack power limits would force Cloudflare to underutilize available space by as much as 50%, and the cooling ceiling prevents using the next wave of CPUs, which are expected to reach 500W TDP.

Limiting CPU TDP via lower-power SKUs isn't a viable path either. Lab benchmarks showed services scale effectively with cores up to 128 cores (256 threads), delivering clear performance and TCO gains at and above 360W. While per-server metrics look good, rack-level constraints tell a different story. A performance sweep across configurable TDP ranges showed that rack-level TCO benefit stagnates above roughly 340W—the per-server performance gains are negated by the need to install fewer servers per rack to stay within the power budget. Even capped at 340W, 30% of rack space would remain unused.

Moving to a 2U Chassis

The solution is to increase the chassis height from 1U (1.75") to 2U (3.5"). While this reduces the number of physical servers per rack, it resolves the power and cooling bottlenecks:

  • Larger fans (up to 80mm) move more air at lower speeds.
  • A taller, larger heatsink dissipates heat more effectively.
  • Air impedance drops since most components remain at 1U height.
  • The extra space accommodates PCIe-attached accelerators/GPUs, including dual-slot cards.

Because Cloudflare is power-constrained rather than space-constrained, the typical space tradeoff of a 2U design doesn't hurt. Vendor thermal simulations show that four 60mm or four 80mm fans running at under 40W each are sufficient to cool the system—a theoretical savings of at least 150W compared to eight 40mm fans in a 1U chassis. That translates to meaningful operational expenditure savings and allows the rack to be used at full power and space capacity.

The decision seems counter-intuitive, but using more space per node actually raises rack density and improves overall TCO in this context. The Gen 12 team is validating the new chassis design and expects to move it into production soon.