Roofline
Also known as frontier envelope, roofline curve
In plain English
On InferenceX a roofline is the outer envelope drawn through the best points of one hardware configuration, showing the edge of what it achieved.
Technical definition
A roofline on InferenceX is the Pareto envelope curve drawn per hardware configuration through its non-dominated points for the selected pair of axes.
Engineering details
Which corner counts as best depends on the metrics: throughput against interactivity takes the upper right, while a cost axis is better when lower, so the envelope is anchored at whichever corner the selected metric pair defines. Points with a degenerate x value are excluded from eligibility, though they still render in the show-all view.
Why it matters
Reading a cloud of raw points invites cherry-picking, since a badly tuned configuration contributes points that no operator would ever choose. The envelope shows the achievable boundary per system, which is the shape a capacity decision is actually made against.
How to read it in InferenceX
This is not the classic roofline model from HPC, which plots attainable FLOPS against arithmetic intensity to expose a compute or memory bound. The dashboard borrows only the picture of an upper bound. Roofline direction is configured per metric, so the same points can produce a different envelope on a different axis.
Source material
See the concept in real benchmarks
InferenceMAX: Open Source Inference Benchmarking
NVIDIA GB200 NVL72, AMD MI355X, Throughput Token per GPU, Latency Tok/s/user, Perf per Dollar, Cost per Million Tokens, Tokens per Provisioned Megawatt, DeepSeek R1 670B, GPTOSS 120B, Llama3 70B
InferenceX v2: NVIDIA Blackwell Vs AMD vs Hopper - Formerly InferenceMAX
GB300 NVL72, MI355X, B200, H100, Disaggregated Serving, Wide Expert Parallelism, Large Mixture of Experts, SGLang, vLLM, TRTLLM
B200 NVFP4 vs H200 FP8 on GLM-5: Up to 3.65x Better Performance per Dollar with SGLang MTP
Both SKUs run SGLang EAGLE MTP; the Blackwell generation lifts perf/$ by ~1.2x at the peak and the NVIDIA GLM-5-NVFP4 checkpoint on FlashInfer TRT-LLM sparse MLA stacks another ~2.4–3.0x on 8K/1K
GLM 5.3 on AgentX: MI355X ATOM Beats GB300 NVL72 on Part of the Curve
Where AMD’s vendor engine wins on performance per dollar, and what E2E Normalized Interactivity actually measures