F4 by Microsoft Azure

Precomputed compound score for Cache Intensive workloads. A weighted average (geometric mean) of benchmark scores compared to their medians: score = ∏ (x_i / m_i)^(w_i / Σw). The score of 1.0 represents a synthetic baseline server with the median performance of each component benchmark; 0.5 means roughly half the performance; and 2.0 means twice the performance of that reference profile. Component weights: 50% Redis RPS (pipeline=1, SET), 20% Redis RPS (pipeline=16, SET), 10% PassMark Memory Mark (composite), 10% Memory bandwidth (read, 16 MB ~ L3), 10% PassMark single-thread CPU. Rationale for component selection: In-memory key-value store workload, mixing direct Redis performance metrics with memory speed and latency benchmarks, and single-core CPU performance profiles.

Precomputed compound score for CI/CD Build workloads. A weighted average (geometric mean) of benchmark scores compared to their medians: score = ∏ (x_i / m_i)^(w_i / Σw). The score of 1.0 represents a synthetic baseline server with the median performance of each component benchmark; 0.5 means roughly half the performance; and 2.0 means twice the performance of that reference profile. Component weights: 50% Geekbench Clang compilation (multi-core), 10% Geekbench Clang compilation (single-core), 20% stress-ng div16 best-N cores, 5% PassMark integer math, 5% PassMark compression, 5% Brotli compression (multi-core, level 0), 5% PassMark string sorting. Rationale for component selection: Build performance is mainly driven by multi-core compilation throughput, but also bundles single-core compilation speed and general CPU performance, multi-core compression and text/scripting processing.

Precomputed compound score for Compute Heavy Applications workloads. A weighted average (geometric mean) of benchmark scores compared to their medians: score = ∏ (x_i / m_i)^(w_i / Σw). The score of 1.0 represents a synthetic baseline server with the median performance of each component benchmark; 0.5 means roughly half the performance; and 2.0 means twice the performance of that reference profile. Component weights: 15% stress-ng div16 best-N cores, 10% stress-ng div16 single core, 20% PassMark CPU Mark (composite), 10% Memory bandwidth (read, 64 MB), 15% PassMark floating point, 15% PassMark AVX/SSE/FMA (SIMD), 10% PassMark integer math, 5% PassMark physics simulation. Rationale for component selection: Number-crunching workload augmenting raw CPU performance stressing, general CPU performance benchmarks, memory bandwidth, and pure math computation speed like floating point, integer, SIMD (AVX/SSE/FMA) operations.

Precomputed compound score for Data Analysis workloads. A weighted average (geometric mean) of benchmark scores compared to their medians: score = ∏ (x_i / m_i)^(w_i / Σw). The score of 1.0 represents a synthetic baseline server with the median performance of each component benchmark; 0.5 means roughly half the performance; and 2.0 means twice the performance of that reference profile. Component weights: 70% PassMark CPU Mark (composite), 10% Gzip compression (single-core, level 5), 10% Memory bandwidth (read, 64 MB), 10% PassMark Memory Mark (composite). Rationale for component selection: Data analysis and ETL workloads are memory-bandwidth-bound and CPU-throughput-driven. The profile combines general CPU performance and memory bandwidth/latency as the primary drivers, supplemented by single-core compression speed as a proxy for serialisation-heavy ETL tasks.

Precomputed compound score for LLM Inference workloads. A weighted average (geometric mean) of benchmark scores compared to their medians: score = ∏ (x_i / m_i)^(w_i / Σw). The score of 1.0 represents a synthetic baseline server with the median performance of each component benchmark; 0.5 means roughly half the performance; and 2.0 means twice the performance of that reference profile. Component weights: 15% LLM text generation (SmolLM-135M, 128 tok), 15% LLM prompt processing (SmolLM-135M, 512 tok), 15% LLM text generation (Llama 7B, 128 tok), 15% LLM prompt processing (Llama 7B, 512 tok), 15% LLM text generation (Llama-3.3 70B, 128 tok), 15% LLM prompt processing (Llama-3.3 70B, 512 tok), 5% Memory bandwidth (read, 256 MB), 2% PassMark AVX/SSE/FMA (SIMD), 2% PassMark floating point. Rationale for component selection: VRAM and memory-bandwidth-bound LLM inference workload, using direct LLM speed benchmarks at three model sizes, and supplementing with raw memory bandwidth and SIMD performance benchmarks.