Emerging Computing Trends in the Datacenter
summary
Moore-era voltage scaling breaks down around 2003, forcing a shift from single-thread speed to energy-efficient multicore designs and then to heterogeneous accelerators; CPUs remain flexible but deliver limited inference throughput while GPUs dominate training, FPGAs sit in the middle, and ASICs buy an order-of-magnitude efficiency at a development cost of roughly $50–200M. Dennard scaling ends ~2003 Centriq 2400: 48 cores, 10 nm, 2.6 GHz, 120 W average SPEC power ~65 W, idle 8 W ASIC cost ~$50–200M.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:33Outline
- 2:08Dennard Scaling Scaling Factor
- 4:48The last 5 Generations of 135W Xeon Processors
- 10:05Mobile Technology Disrupting the Cloud Datacenter
- 13:49Computational + server growth fuel datacenter energy efficiency considerations
- 18:01Qualcomm Centriq 2400: Built for The Cloud World's first 10nm server processor
- 23:02Qualcomm Centriq 2460 Lowers Average and Idle Power to Improve Cloud Server Density in Datacenters
- 28:16Lessons from Mobile Computing
- 30:03The Age of Application Specific Accelerators
- 32:28The Emergence of Deep Neural Networks
- 35:38What is "Edge"?
- 36:45Silicon alternatives for DNNS
related talks
33mReimagining the Modern Data Center
24mSilicon Development
25mSustainability at the Core with Cloud Native Processors
24mClosing Keynote
16mChange Your Chip Architecture and Save the World
36mThe journey of silicon innovation at
29mUnlocking the Future of AI: Trends, Infrastructure, and Tokenomics
22mFireside Chat