Iso-interactivity
Also known as matched interactivity, equal token rate
In plain English
Iso-interactivity compares systems while users see words appear at the same speed. This provides an apples-to-apples view of the hardware behind the experience.
Technical definition
Iso-interactivity means comparing systems at the same per-user generation rate.
Engineering details
Benchmark runs rarely land at identical tok/s/user values because each recipe has different concurrency points. An iso-interactivity comparison interpolates each Pareto frontier at a shared target and then compares throughput, cost, or efficiency there.
Why it matters
Matching streaming speed avoids declaring a system faster merely because it serves more requests at a slower token rate. It does not hold the entire user experience constant: initial waiting time and total response duration still need separate comparisons.
How to read it in InferenceX
InferenceX articles use iso-interactivity tables for hardware, precision, and software comparisons. Values outside a measured frontier are marked unreachable and are not extrapolated beyond observed data. The frontier is always built on throughput against interactivity; cost per million tokens and joules per token are then derived from the interpolated throughput rather than splined on their own, because each is a per-chip constant divided by that throughput and splining it separately would break the identity between knots.
Source material
See the concept in real benchmarks
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 ~2.4β3.0x on 8K/1K
B200 NVFP4 vs H100 FP8 on MiniMax-M2.5: Up to 8.2x Better Performance per Dollar with vLLM
vLLM PR #36307 unlocks the trtllm-gen FP8 MoE kernel for MiniMax on B200; combined with NVFP4, perf/$ scales from 4.0x at 22 tok/s/user to 8.2x at 110 on 8K/1K
B200 NVFP4 vs H200 INT4 on Kimi K2.5/K2.6: Up to 2.95x Better Performance per Dollar
On vLLM 8K/1K the NVFP4 path on B200 is 2.71xβ2.95x cheaper per million tokens than H200 INT4 across the entire 30β90 tok/s/user serving band, and 2.45xβ2.74x cheaper than B200 INT4 on the same silicon. Both factors decompose cleanly into B200's HBM bandwidth, HBM capacity, and NVFP4 tensor cores
GB300 NVL72 vs GB200 NVL72 Inference Performance & Perf per Dollar - on DeepSeek-V4-Pro 1.6T: Up to 2.83x Throughput
DSv4-Pro FP4 8K/1K, Dynamo+vLLM, disaggregated on both racks. GB300's 50% extra HBM (288 vs 192 GB/GPU) unlocks a wider prefill+decode recipe GB200 can't fit β lifting middle-of-curve perf/$ by 2.31x despite a 20% per-GPU TCO premium.
GLM 5.3 on AgentX: NVIDIA Is Up to 5x Cheaper per Token at 150 tok/s/user
At this operating point, free AMD silicon would still not close the gap
Rubin NVL72 Agentic Inference: 67x better Performance per Dollar
Jensen Sandbagging Performance Again, 2x more Annual Profit Per GigaWatt, The More you Buy, The More you Earn, AgentX, InferenceX, Extreme Co-Design