AWS wellarchitected documentation change
Summary
Restructured guidance on floating-point precision, added implementation recommendations for generative AI quantization, validation, and mixed-precision training to reduce resource consumption.
Security assessment
Changes exclusively address computational efficiency and accuracy trade-offs in numerical processing and generative AI workloads. No security context, vulnerabilities, or security features are introduced.
Diff
diff --git a/wellarchitected/latest/financial-services-industry-lens/fsisus15.md b/wellarchitected/latest/financial-services-industry-lens/fsisus15.md index 9202a75da..dd19fcf1d 100644 --- a//wellarchitected/latest/financial-services-industry-lens/fsisus15.md +++ b//wellarchitected/latest/financial-services-industry-lens/fsisus15.md @@ -11 +11 @@ FSISUS15-BP01 Minimize the bit count while maintaining precision -**Prescriptive guidance** +### Prescriptive guidance @@ -13 +13 @@ FSISUS15-BP01 Minimize the bit count while maintaining precision -Floating point precision is a way to represent real numbers in a finite binary format. It stores a number in a fixed-width field with the intent to reduce the memory bandwidth and storage requirements compared to double-precision arithmetic results. Although double-precision can sometimes lead to more accurate results, single-precision calculations can be faster and thus reduce overall energy consumption for particular workloads. Determine which of your workloads is suitable for use of floating-point accuracy, performance, and efficiency. Consider testing with a cluster of instances to see how well it performs at scale. +Floating point precision is a way to represent real numbers in a finite binary format. It stores a number in a fixed-width field with the intent to reduce the memory bandwidth and storage requirements compared to double-precision arithmetic results. Although double-precision can sometimes lead to more accurate results, single-precision calculations can be faster and thus @@ -15 +15,3 @@ Floating point precision is a way to represent real numbers in a finite binary f -**Implementation guidance:** +reduce overall energy consumption for particular workloads. Determine which of your workloads is suitable for use of floating-point accuracy, performance, and efficiency. Consider testing with a cluster of instances to see how well it performs at scale. + +### Implementation guidance: @@ -31,0 +34,9 @@ Floating point precision is a way to represent real numbers in a finite binary f + * Test generative AI models with reduced precision (quantization) to maintain accuracy while reducing resource consumption. + + * Validate generative AI model performance with different floating-point precisions. + + * Use mixed-precision training for generative AI models to optimize resource usage. + + + +