The updated report is now 153 pages, and has quite a few new results. In the February report, I found the results on Kalamang translation for the Machine Translation from One Book benchmark quite exciting. In this updated report, we’ve extended this line of evaluation to test
@jeffdean
-
Google Releases Gemini 1.5 Pro and Flash Models
By
–
Gemini 1.5 Model Family: Technical Report updates now published In the report we present the latest models of the Gemini family – Gemini 1.5 Pro and Gemini 1.5 Flash, two highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information
-
SmartChoices: Learning Integration in YouTube Video Caching
By
–
And the SmartChoices system, which makes it easy to integrate learned choices in the middle of "classic hand-written code", and as one example of this system, how it was used to use learning to select which items to replace in the YouTube video caching system (result in a huge
-
Machine Learning for Memory Lifetime Prediction and Fragmentation
By
–
And learning for predicting object lifetimes, so as to reduce memory fragmentation in memory allocation systems, from this work: https://
cacm.acm.org/research/combi
ning-machine-learning-and-lifetime-based-resource-management-for-memory-allocation-and-beyond/
… -
Machine Learning for Better Compiler Decision Making
By
–
I also talked about a number of ML for Computer Systems topics. One was the use of learning for better compiler decisions.
-
Google Cloud Dashboard Enables Low-Carbon Region Selection
By
–
Google Cloud offers a dashboard that provides data on the %age of CFE used for different Cloud regions to help users select appropriate regions, as well as "meta-regions" to enable users to select low-carbon regions automatically (using, for example, "in:us-low-carbon-locations")
-
NAS Emissions Estimation Errors: 19X and 5X Analysis
By
–
Please look at https://
arxiv.org/pdf/2104.10350. In particular, Appendix C shows details on the 19X error, and Appendix D shows details on the further 5X error in estimating the emissions of the NAS, and the right hand column of Table 1 shows the measured data for using the Evolved -
RL Achieves Wirelength Reduction vs Human Expert Placement
By
–
Percentage wirelength reduction of the RL solution vs. the same block placed by a human expert.
-
Making Inference Performance Faster: Key Discussion Points
By
–
I also discussed a number of issues around how to make inference performance faster and why this was important.
-
Silent Data Corruption Risks in Large-Scale ML Training
By
–
I discussed the challenges of silent data corruption in ML training jobs, and how one faulty piece of hardware can infiltrate and affect the results of a large scale training jobs on thousands of chips.