More cores, more work, less power
Cloudflare's network handles roughly 11 million HTTP requests per second from more than 1 billion unique IP addresses daily, spanning 200 cities across 90 countries. Because the company's software stack is compute-intensive and CPU-bound, server hardware selection is critical. The straightforward path to more compute is adding cores; the constraint is data center power infrastructure, which forces Cloudflare to cap power per rack.
Cloudflare's key metric is Requests per Watt. More cores help Requests per Second, but only if power draw stays within the rack budget. If a rack exceeds its cap, operational costs rise because another rack must be provisioned. The company's Gen 9 servers paired dual-socket Intel Xeon Platinum 6162 processors for a 48-core setup at a TDP of 300W. For Gen X, Cloudflare moved to a single-socket, 48-core AMD EPYC 7642 with a 225W TDP. That drops TDP per core from 6.25W to 4.69W, a 25% decrease.
Cloudflare cautions that TDP is a simplified datasheet metric, not a standardized specification across manufacturers, and actual system power draw is what matters in final decisions. With higher frequency and a simplified single-socket design, the expectation was that the AMD chip would perform better than the dual Intel setup. The company ran a series of tests to verify.
Sniff tests: crypto, compression, and Go
Before committing engineering resources, Cloudflare ran a quick benchmark suite across candidate CPUs, comparing a single-socket AMD EPYC 7642 server against a dual-socket Intel Xeon Platinum 6162 server. BIOS settings were matched to production configurations, yielding average CPU frequencies of 3.03 GHz for AMD and 2.50 GHz for Intel — a rough 21% theoretical advantage for AMD with equal core counts.
In C-based OpenSSL tests, AMD won public key cryptography by 18%. For symmetric key cryptography, AMD lost on AES-128-GCM but was comparable overall. In compression tests using zlib and brotli libraries on an HTML file in memory, AMD won by an average of 29% for gzip across all quality levels, and did even better with brotli below quality 7. At brotli-9 there's a throughput cliff, likely from cache thrashing due to high memory consumption, but AMD still won by a healthy margin.
For Go-based services, Cloudflare re-ran crypto, compression, regexp, and string benchmarks. AMD performed better across all Go benchmarks except ECDSA P256 Sign, where it lost by 38% — a peculiar result given AMD's 24% advantage in the C test for the same operation. The company notes this discrepancy is worth investigating.
LuaJIT, which Cloudflare describes as the glue holding its stack together, also favored AMD. Averaging across all sniff tests equally, AMD performed about 25% better than the dual Intel setup.
Synthetic workloads simulate production
After the initial benchmarks, Cloudflare ran emulations with synthetic workloads mimicking production request patterns, varying asset size, HTTP vs. HTTPS, WAF, Workers, and other variables. With Intel normalized at 1.0, AMD achieved 1.50x better Requests per Second for simple 10KiB HTTPS requests. The average improvement across tested scenarios was 34%.
Given the power difference — 225W for the single AMD EPYC 7642 versus 300W for two Intel Xeons — AMD delivers up to 2.0x better Requests per Watt. Cloudflare shipped test servers to data centers for live production validation.
Live production results
Cloudflare deploys all server generations in homogeneous clusters with similar CPU utilization across 8-core and 24-core machines. Test Gen X AMD servers and Gen 9 Intel baseline servers showed no significant CPU usage difference, ensuring a fair comparison.
Live production data showed AMD processed about 23% more requests per core and per server on average, with the same core count and less power. Actual power draw measurements, rather than TDP figures, showed the AMD server performing 28% better on Requests per Watt. AMD's power draw ran near spec TDP at higher-than-base frequency, while Intel ran well below its TDP rating — evidence that TDP is an unreliable power estimate.
Cloudflare notes that the comparison didn't favor AMD on system design: Intel CPUs were in a multi-node system designed for high density at lower power per node, while AMD ran in a regular 1U form factor. AMD still outperformed on a power-per-node basis.
Across datasheet comparison, simulations, and live production, the 1P AMD EPYC 7642 beat the 2P Intel Xeon 6162. Some environments showed AMD up to 36% better in live production. Latency measurements for NGINX processing showed AMD about 25% faster on average, with p99 latencies 20-50% faster depending on time of day.
Gen X goes with AMD
Cloudflare's conclusion: AMD wins. The single-socket AMD EPYC 7642 configuration delivers more Requests per Second and better Requests per Watt than the dual-socket Intel setup it replaces. For a network that adds capacity and new cities regularly, lower power per server means more compute per rack and easier expansion.
The company plans to deploy its Gen X servers powered by 2nd Gen AMD EPYC processors across its network, bringing what it calls "Rome" to data centers near its users. Cloudflare says it will continue evaluating Intel, AMD, and ARM chips for future generations.
