AI Compute Converter

What is an AI compute unit converter?#

AI hardware is sold in two parallel families of “operations per second” units, and they look interchangeable but are not. The FLOPS family (FLOPS, kFLOPS, MFLOPS, GFLOPS, TFLOPS, PFLOPS, EFLOPS, ZFLOPS) counts floating-point operations per second — the traditional measure for scientific computing, training throughput, and graphics. The OPS family (OPS, kOPS, MOPS, GOPS, TOPS, POPS) counts integer or “mixed” operations per second — the figure chip vendors quote for INT8 inference accelerators. A “100 TOPS” NPU and a “100 TFLOPS” GPU cannot be compared by the same yardstick, because a TOPS “operation” and a FLOPS “operation” measure different things.

This page converts within each family exactly (every unit is a clean power-of-1000 multiple of its family base), and it refuses cross-family conversions rather than inventing a ratio. That refusal is the whole point: any single number that “converts” TFLOPS to TOPS is a fabrication, because vendors do not even agree on whether one multiply-accumulate counts as one operation or two.

How to use it#

1. Enter the Value you want to convert.
2. Pick From using the grouped dropdown — units are listed under two optgroups, FLOPS and OPS, so the family boundary is visible right in the picker.
3. Pick To the same way.
4. Read the Result. If both units are in the same family you get a number; if you crossed families you get an explicit cross-family notice instead of a made-up figure.
5. Expand All units to see your input converted into every unit of the same family at once, rounded to six significant digits. Units of the other family are omitted because they are not convertible.

Key features#

• Two families, never conflated. FLOPS and OPS are shown as separate optgroups and never silently mixed. The math within a family is exact (value × fromFactor ÷ toFactor); across families the tool says no.
• Honest cross-family refusal. A TFLOPS-to-TOPS request returns an explicit cross-family error, not a plausible-looking but meaningless number. This is the difference between a converter and a guess.
• Full FLOPS range to ZFLOPS. From single-ops FLOPS up through k, M, G, T, P, E and Z (10²¹), covering everything from an embedded MCU to a hypothetical exascale-plus system.
• Full OPS range to POPS. OPS through k, M, G, T, P — the span vendor spec sheets actually use for INT8 inference claims.
• Six-significant-digit rounding. Results are rounded with epsilon correction to suppress binary floating-point noise, so 1 TFLOPS reads as exactly 1000 GFLOPS rather than 999.9999999.
• All-units table. One input, every same-family unit at a glance — useful when a spec sheet quotes TOPS and you want to see the same throughput in GOPS and MOPS too.

Worked example#

A GPU spec sheet claims 1 TFLOPS of FP16 throughput. To express that in GFLOPS, set Value 1, From TFLOPS, To GFLOPS:

1 TFLOPS = 1000 GFLOPS

Expand All units and the same value fills the entire FLOPS family:

FLOPS:   1,000,000,000,000
kFLOPS:  1,000,000,000
MFLOPS:  1,000,000
GFLOPS:  1,000
TFLOPS:  1
PFLOPS:  0.001
EFLOPS:  1e-6
ZFLOPS:  1e-9

Now try to compare that against an NPU rated in TOPS: set From TFLOPS, To TOPS. The result is not a number — it is the cross-family notice. There is no physically meaningful conversion, because TFLOPS measures floating-point operations (with their own precision semantics) and TOPS measures integer/mixed operations (typically INT8), and a MAC counts as one operation for some vendors and two for others. The honest answer is “they are not comparable without more information”, and that is what the tool returns.

FAQ#

Why can’t I convert TFLOPS to TOPS?#

Because they count different things. TFLOPS is floating-point operations per second; TOPS is integer (usually INT8) operations per second, and the two operation types do not have a fixed ratio — it depends on the chip’s data paths and on whether a multiply-accumulate is counted as one or two operations. Any single conversion factor would be a guess. Convert within FLOPS, or within OPS, but not between them.

What does the FLOPS family actually measure?#

Floating-point operations per second, at whatever precision the spec sheet defines (FP64, FP32, FP16, INT8-as-float…). Crucially, “1 TFLOPS FP16” and “1 TFLOPS FP32” are not the same amount of work — a chip’s TFLOPS rating depends on which precision you quote. Always read the precision alongside the number; the unit alone is not enough.

Are kFLOPS, MFLOPS, GFLOPS all powers of 1000?#

Yes. Every step within a family is ×1000: 1 kFLOPS = 1000 FLOPS, 1 MFLOPS = 1000 kFLOPS, 1 GFLOPS = 1000 MFLOPS, and so on up to ZFLOPS (10²¹). There is no binary/decimal ambiguity here the way there is with KiB versus KB — compute units are universally decimal.

Why does the All-units table omit the other family?#

Because those units are not convertible to yours. If you entered a FLOPS value, the table shows every FLOPS unit; the OPS units are omitted rather than shown with fabricated numbers. To see OPS equivalents, re-enter your value as an OPS figure (from a spec sheet that actually quotes OPS) and the table will populate the OPS column instead.