Internet-grooming company Cloudflare has revealed that it was unable to put Intel inside its new home-brew servers, because they just used too much energy.
A Tuesday post by platform operations engineer Chris Howells reveals that Cloudflare has been working on designs for an eleventh-generation server since mid-2020.
"We evaluated Intel's latest generation of 'Ice Lake' Xeon processors," Howells wrote. "Although Intel's chips were able to compete with AMD in terms of raw performance, the power consumption was several hundred watts higher per server – that's enormous."
Fatally enormous – Cloudflare's evaluation saw it adopt AMD's 64-core Epyc 7713 for the servers it deploys to over 200 edge locations around the world.
Power savings also influenced a decision to go from three disks to two in the new design. A pair of 1.92TB Samsung drives replaced the three of the Korean giant's 960GB units found in previous designs. The net gain was a terabyte of capacity, and six fewer watts of power consumption.
Howellls's post also reveals that testing produced data showing that equipping its servers with 512GB of RAM did not produce enough of a performance boost to justify the expense. The company has therefore settled on 384GB of memory, but did jump from DDR4-2933 to DDR4-3200 as the slight cost increase delivered a justifiable performance boost.
- It's time to decentralize the internet, again: What was distributed is now centralized by Google, Facebook, etc
- Don't rush to adopt QUIC – it's a slog to make it faster than TCP
- Cloudflare slams AWS egress fees to convince web giant to join its discount data club
Cloudflare stuck with Mellanox ConnectX-4 dual-port 25G Ethernet adapters.
"We investigated higher-speed Ethernet, but we do not currently see this as beneficial," Howells wrote. That's not a brickbat for fast Ethernet, but a decision made possible by Cloudflare's highly distributed architecture that removes the need for higher speeds and the higher cost of faster kit.
One change in the new servers is the use of OpenBMC to deploy firmware.
"With access to the source code, we have been able to configure BMC features such as the fan PID controller, having BIOS POST codes recorded and accessible, and managing networking ports and devices," Howells wrote.
"Whilst our current BMC is an industry standard, we feel that OpenBMC better suits our needs and gives us advantages such as allowing us to deal with upstream security issues without a dependency on our vendors.
"Some opportunities with security include integration of desired authentication modules, usage of specific software packages, staying up to date with the latest Linux kernel, and controlling a variety of attack vectors."
Howells rated Cloudflare's new servers as a refinement, rather than the "enormous" change between its ninth- and tenth-generation boxes that saw it go from dual-socket Intel servers to single-socket AMDs.
But the refinement delivered a huge payoff: the ability to handle "approximately 29 per cent more requests than generation ten without an increase in power consumption".
Howells's post states that Ampere's Arm servers also made its shortlist for eleventh-gen servers, and promised to reveal details of those designs in a future post.
He also wrote that he and Cloudflare value ongoing competition between AMD and Intel, "and we look forward to seeing how Intel's next generation shapes up".
Intel claims that new generation – named Sapphire Rapids – delivers a generational jump in performance and efficiency.
Cloudflare's decision to go with AMD shows it needs to. ®