Apples-to-apples comparison
Also known as apples to apples, apples to bananas, like-for-like comparison
In plain English
An apples-to-apples comparison holds precision, serving mode, and workload constant so a difference in results can be attributed to the hardware.
Technical definition
An apples-to-apples comparison in InferenceX matches numerical precision, aggregated or disaggregated serving mode, speculative decoding settings, and workload shape across systems so that only the hardware and its software stack differ.
Engineering details
FP8 on one side and FP4 on the other is not apples to apples, because FP4 carries a quality loss relative to FP8 and reads half the bytes per weight. Aggregated on one side and disaggregated on the other mixes a deployment-mode advantage into the result. Single-token prediction against MTP does the same for speculative decoding. InferenceX calls the mismatched case apples to bananas and still publishes it, labeled, because buyers face those choices, but headline claims come from matched settings. Ironwood has no native FP4, so its fair comparison today is FP8 versus FP8 Blackwell; TPUv8i will be compared FP4 versus FP4.
Why it matters
A large share of vendor performance claims rest on mismatched settings. Insisting on matched precision and serving mode is what makes a performance-per-dollar ratio a statement about the chip rather than about the recipe chosen for the marketing slide.
How to read it in InferenceX
The InferenceX Official Preview headline, up to 50% better performance per dollar for TPUv7, is an apples-to-apples FP8 aggregated single-token comparison against B200 and B300. The GB300 NVL72 disaggregated versus TPUv7 aggregated chart is explicitly labeled apples to bananas and shows about a 30% GB300 advantage mid-curve.
Source material
See the concept in real benchmarks
TPU Inference Externalization Full Steam Ahead
InferenceX, Up to 50% Better Performance per Dollar, Rapid Externalization of TPU stack, Growing Customer Base, Ironwood, TPUv8i, Reducing CUDA Moat
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
Vera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture Analysis
Rubin LUT Based Tensor Core, Feynman, Rack Scale, Perf Per MegaWatt, Perf Per Dollar, Software Improvements, Public Rubin Software, PyTorch, vLLM, OpenAI Triton
OpenAI Jalapeño: Better Than Nvidia Blackwell
OpenAI’s self-designed ASIC compared with Rubin, Jalapeño’s TCO, throughput per MW, and spicy deets