3 min read
|
Saved February 14, 2026
|
Copied!
Do you care about this?
This article analyzes cloud hardware developments over the past decade, focusing on performance improvements in CPU, memory, network, and NVMe storage. While network bandwidth has significantly increased, gains in CPU and memory have stagnated, and NVMe performance in the cloud has not kept pace with on-premise hardware. The findings suggest a shift towards specialized hardware and software integration to maximize performance.
If you do, here's more
The paper analyzes cloud hardware trends from 2015 to 2025, focusing primarily on AWS while comparing it to other cloud services and on-premise hardware. A standout finding is the dramatic increase in network bandwidth per dollar, which improved tenfold. However, gains in CPU and DRAM performance have been modest. The study reveals that NVMe storage performance has stagnated since 2016, which is unexpected given the advancements in other areas.
CPU performance has benefitted from multi-core parallelism, with AWS's largest instance reaching 448 cores. Despite this, the cost-performance improvements have been around 3x over the decade, largely due to AWS Graviton. In-memory database benchmarks show even smaller gains, between 2x and 2.5x, indicating latency issues. DRAM capacity per dollar has not improved significantly, with a peak improvement from the introduction of memory-optimized instances in 2016. The cost-normalized bandwidth gains from moving to DDR5 memory are only 2x, despite a 5x increase in absolute bandwidth.
On the networking front, substantial improvements have emerged, with speeds increasing from 10 Gbit/s to 600 Gbit/s. However, much of this is due to advancements in network-optimized instances, not generic ones. In contrast, NVMe SSD performance has not kept pace, with AWS's best instance still being the i3 from 2016. The stagnation in cloud NVMe performance, alongside rising prices, may drive a shift toward disaggregated storage solutions. The paper suggests that current architectural choices may need reevaluation, especially as network speeds increase. Lastly, the paper highlights a growing need for specialized hardware and software integration as general-purpose performance plateaus, raising questions about current software's ability to leverage the available parallelism effectively.
Questions about this article
No questions yet.