Performance Metric

TFLOPS

Trillion floating-point operations per second — the standard GPU compute throughput metric.

TFLOPS (TeraFLOPS) measures how many trillion floating-point math operations a GPU can perform per second at a given precision. Ratings are precision-specific: a GPU's FP32 TFLOPS, FP16 TFLOPS, FP8 TFLOPS, and INT8 TFLOPS are all different numbers.

Marketing figures often include sparsity (treating zero-valued matrix elements as free), doubling the reported rate. Dense (non-sparse) TFLOPS is the more conservative number; AIMC uses this where possible to avoid overstating capability.

Headline 2026 datacenter ratings (dense FP16): H100 SXM ~989 TFLOPS, A100 SXM 80GB ~312 TFLOPS, B200 SXM ~2,250 TFLOPS, MI300X ~1,300 TFLOPS. For consumer parts: RTX 5090 ~104 TFLOPS, RTX 4090 ~83 TFLOPS. TFLOPS is one of three inputs to AIMC's workload fit score.

Related Terms

Concepts directly relevant to TFLOPS.

FP16

16-bit half-precision floating-point format used heavily in deep learning.

FP8

8-bit floating-point format introduced for high-throughput inference and training.

Memory Bandwidth

How fast a GPU can read and write its VRAM, measured in gigabytes per second.

This definition is part of AIMC's reference glossary — 25 concepts across 8 categories.

Browse full glossary