Match the workload
Compare the model, quantization, context length, batch size, and serving runtime before comparing throughput figures.
Evidence, not assertions
A curated index of third-party benchmarks, engineering analysis, and operational references for AMD MI300X infrastructure and inference workloads.
Start nowBenchmark context is part of the result
Useful benchmarks document what was tested and how. These references are intended to help you frame the right questions before running your own.
Read the full system
Performance is a system result. GPU, CPU, memory, network, storage, drivers, and runtime configuration all contribute.
Validate in context
Use third-party data as a starting point, then validate the configuration and operating conditions that matter to your deployment.
Research, benchmarks, and field reports
External sources from engineering teams, cloud providers, researchers, and the AMD community.
MI300X community and current discussion
Open sourceHugging Face
Magnum 72B, trained on AMD MI300X
Open sourceAMD MI300X vs. Nvidia H100 FFT benchmark
Open sourceSingle MI300X vLLM benchmark
Open sourceMI300X Geekbench OpenCL performance
Open sourceChips and Cheese
Testing AMD’s Bergamo
Open sourceOracle Cloud
Early LLM serving experience with MI300X
Open sourceChips and Cheese
Testing AMD’s Giant MI300X
Open sourceNscale
MI300X GEMM tuning and throughput
Open sourceRunPod
MI300X vs. Nvidia H100 SXM for Mixtral 8x7B
Open sourceAMD Community
MLPerf results on MI300X
Open sourceGitHub
Updated MI300X GEMM tuning with Gradlib
Open sourcedstack
Llama 3.1 405B on 8x MI300X
Open sourceFireworks.ai
FireAttention V3 and AMD inference
Open sourceAMD Community
Tuning vLLM for efficient inferencing
Open sourceAMD Community
vLLM serving best practices on MI300X
Open sourceLinkedIn Engineering
Liger Kernels and the CUDA moat
Open sourceMedium
Anton Lokhmotov’s MLPerf CPU benchmarks
Open sourceHot Aisle 8x MI300X speed test
Open sourceSemiAnalysis
Faster tokens for fewer dollars
Open sourceRun the test that matters to you
Bring your model, runtime, and request profile to an isolated AMD GPU VM, then validate performance against your own operating requirements.
Launch compute