TechPrufer
Get started
Dropbox
Dropbox logoDropbox · Feature trail

Server Hardware Architecture

6 posts · Jan 2019 – Jul 2025

What this trail covers

Dropbox's ongoing efforts to improve infrastructure efficiency, focusing on planning, continuous optimization of the active fleet (including Deep Sleep and workload balancing), and extending hardware lifecycle. This capability evolves to address growing demand, particularly from AI workloads, by maximizing resource utilization and minimizing waste. This post details the design and deployment of their seventh generation server hardware, including specialized tiers for compute (Crush), database (Dexter), and storage (Sonic), as well as new GPU tiers (Gumby and Godzilla). Key improvements include higher core count CPUs (84-core AMD EPYC 9634), increased memory capacity (512GB), DDR5 memory, 100Gb networking, NVMe gen5, and a new storage chassis designed to minimize vibration and heat. The database platform saw a 30% increase in IPC and higher base frequency, leading to reduced replication lag. The storage platform aims for 200Gbps throughput and is 400G-ready.

All Dropbox trails
  1. Seventh-generation server hardware at Dropbox: our most efficient and capable architecture yetThis post details the design and deployment of Dropbox's seventh-generation server hardware platform, codenamed Crush (compute), Dexter (database), and Sonic (storage), along with new GPU tiers Gumby and Godzilla. Key technical advancements include: a shift to 84-core AMD EPYC 9634 CPUs for compute, offering 75% more cores and 2x memory capacity compared to the previous generation; DDR5 memory and 100Gb networking for improved bandwidth; NVMe gen5 for faster local disk access; a database platform with 30% higher IPC and increased base frequency, reducing replication lag; and a storage platform designed for 200Gbps throughput and 400G networking readiness. The design also focuses on thermal and power management to accommodate increased power demands.
  2. How the data center site selection process works at DropboxThis post details Dropbox's data center site selection process, which involves defining resource requirements (power, space, time) based on capacity engineering assessments. It outlines the RFP process, focusing on facility-level requirements like supporting infrastructure design (Tier III), cabinet weight, and network design (ingress/egress, carrier availability, redundancy). A significant emphasis is placed on sustainability and Power Usage Effectiveness (PUE) as a cost multiplier and driver for efficiency. The process includes technical questionnaires to assess detailed aspects of the facility and in-person site walks to validate information and uncover potential issues. Facilities are then scored on key parameters like space, power, and network design, with a focus on reliability and efficiency.
  3. After four years of SMR storage, here's what we love—and what comes nextThis post details Dropbox's continued leadership and learnings in adopting Shingled Magnetic Recording (SMR) technology for its storage drives, now comprising 90% of its HDD fleet. It highlights improvements in storage density through advanced recording techniques (TDMR), increased platter count, and optimized enclosure design. The post also quantifies the power efficiency gains (5-6x decrease in TB/watt) and cost benefits (10-20% more data per drive at similar cost) achieved through SMR adoption. Furthermore, it discusses software ecosystem maturation, including better support for zoned devices via libraries like libzbd and advancements in testing tools like SMRtest. Finally, it emphasizes deeper collaboration with HDD vendors, integrating them into large-scale testing for improved bug reproduction and velocity.
  4. A look inside our sixth generation of server hardwareThis post details the design and implementation of Dropbox's sixth generation server hardware, codenamed Scooby (storage), Diego (database), and Cartman (compute), along with Coco for the edge network. It highlights specific hardware configurations for each tier, such as drive density for storage, CPU cores and NVMe SSDs for databases, and core count and RAM for compute. It also discusses the rationale behind general-purpose server designs over service-specific 'snowflakes' and outlines future exploration of ML accelerators and HAMR drives.
  5. RunBMC: OCP hardware spec solves data center BMC pain pointsDropbox contributed the v1.0 of the RunBMC hardware specification to the Open Compute Project (OCP). This specification defines the interface between BMC subsystems and the compute platforms they manage, aiming to solve pain points related to BMC hardware in data centers. The spec uses a 260-pin DDR SO-DIMM connector and standardizes various I/O connectivity like RGMII, RMII, LPC, eSPI, and PCIe, promoting security, supply chain benefits, and faster industry innovation.
  7. Automating Datacenter Operations at DropboxThis post introduces Pirlo, a system designed to automate the provisioning and validation of network switches and servers. Pirlo consists of a ClusterOps queue, a TOR Starter component for switch provisioning, and a Server Validation component for server provisioning and repair. The TOR Starter automates switch validation and configuration through a plugin system, capturing detailed logs and providing a UI for monitoring. Server Validation uses a PXE-booted OS image called Hotdog for minimal Ubuntu with vendor tooling and Dropbox-specific scripts. Both components leverage the in-house ClusterOps queue for job management and integrate with the asset database for inventory tracking.