Aishwarya Mahesh on X: "@opticuddle Rn the fluctuations stem from training workloads. DCs can’t easily smooth training workloads since these jobs require strict synchronization across thousands of GPUs. Slowing them down to manage power would mean costly idle GPUs and longer training times. As a result, most" / X
To view keyboard shortcuts, press question mark View keyboard shortcuts Messages Home Explore Notifications Messages Grok Bookmarks Communities Premium Verified Orgs Profile More Post kristie! @kristiehuang Post Reply See new posts Conversation Aishwarya Mahesh @ash_locked · Jul 7 Just dropped my third Power Bytes note introducing a new metric I developed: Data Center Reserve Margin (DCRM). It’s designed to quantify how much additional AI/data center load a region can realistically absorb before grid reliability breaks down. (1/4) Full analysis + case Show more 2 8 26 3.9K ryan mei @opticuddle · Jul 7 Are the gigawatt-scale fluctuations in consumption driven by demand for inference or training services? Why don’t datacenters optimize training workloads to smooth energy consumption? 1 1 153 Aishwarya Mahesh @ash_locked Rn the fluctuations stem from training workloads. DCs can’t easily smooth training workloads since these jobs require strict synchronization across thousands of GPUs. S
Aishwarya Mahesh @ash_locked Replying to @opticuddle Rn the fluctuations stem from training workloads. DCs can’t easily smooth training workloads since these jobs require strict synchronization across thousands of GPUs. Slowing them down to manage power would mean costly idle GPUs and longer training times. As a result, most operators are looking to use grid-side solutions like BESS and UPS to buffer these fluctuations without compromising performance. 1:30 PM · Jul 8, 2025 86 Views 3 0 3
Explore this link on the map →related reading
- The Most Expensive GPU is The One You Didn't Hedge - ADM Investor Servicesadmis.com
- The Most Expensive GPU is The One You Didn't Hedge - ADM Investor Servicesadmis.com
- Aishwarya Mahesh on X: "@tylerhnorris +1 ... hard to imagine the AI race hinging on whether training runs continue during a winter storm. My understanding though is the bigger tension with training isn’t just reliability but the economic cox.com
- Making Deep Learning go Brrrr From First Principleshorace.io
- Multi-Datacenter Training: OpenAI's Ambitious Plan To Beat Google's Infrastructuresemianalysis.com
- Can AI scaling continue through 2030? | Epoch AIepoch.ai
- The Best GPUs for Deep Learning in 2023 — An In-depth Analysistimdettmers.com
- How to Think About GPUs | How To Scale Your Modeljax-ml.github.io
- From bare metal to a 70B model: infrastructure set-up and scripts - Imbueimbue.com
- Strangely, Matrix Multiplications on GPUs Run Faster When Given "Predictable" Data! [short]thonking.ai
- Reiner Pope – The math behind how LLMs are trained and serveddwarkesh.com
- 2410.21680arxiv.org