Peak FP32 rate (GPU) over time
Theoretical single-precision peak from unit count, clock and operations per cycle. An upper bound, not an achieved rate.
- Device scope
- gpu
- Method
- Theoretical peak
- Plotted in
- floating-point operations per second
Chart
19 figures from 19 machines, spanning 1999 to 2025. The lowest peak fp32 rate here is 960 MFLOP/s (NVIDIA GeForce 256, 1999) and the highest is 104.8 TFLOP/s (NVIDIA GeForce RTX 5090, 2025). That is a range of about 109 167× within one comparability group; the figures are plotted in FLOP/s. Points use the release date of the machine the figure belongs to. The table below lists absent figures with their recorded reasons.
Data
| Machine | Released | Figure | Confidence | Sources |
|---|---|---|---|---|
| NVIDIA GeForce 256 NVIDIA GeForce 256 (NVIDIA GeForce 256) | October 1999 | 960 MFLOP/s | Theoretical provisional | |
| Microsoft Xbox 360 | November 2005 | 240 GFLOP/s | Theoretical provisional | |
| NVIDIA GeForce 8800 GTX NVIDIA Tesla (NVIDIA GeForce 8800 GTX) | November 2006 | 518 GFLOP/s | Theoretical provisional | |
| Sony PlayStation 3 RSX Reality Synthesizer | November 2006 | 1.8 TFLOP/s | Vendor-rated provisional | |
| Microsoft Xbox One | November 2013 | 1.31 TFLOP/s | Theoretical provisional | |
| Sony PlayStation 4 PlayStation 4 custom processor (GPU section) | November 2013 | 1.84 TFLOP/s | Vendor-rated provisional | |
| NVIDIA GeForce GTX 1080 NVIDIA GP104 (GeForce GTX 1080) | May 2016 | 8873 GFLOP/s | Theoretical | |
| Nintendo Switch | March 2017 | 393 GFLOP/s | Theoretical provisional | |
| NVIDIA GeForce RTX 2080 Ti NVIDIA Turing (NVIDIA GeForce RTX 2080 Ti) | September 2018 | 13.45 TFLOP/s | Theoretical | |
| Apple Mac mini (M1, 2020) Apple M1 GPU | November 2020 | 2.6 TFLOP/s | Vendor-rated provisional | |
| Microsoft Xbox Series X | November 2020 | 12 TFLOP/s | Vendor-rated | |
| Sony PlayStation 5 PlayStation 5 custom processor (GPU section) | November 2020 | 10.3 TFLOP/s | Vendor-rated provisional | |
| NVIDIA H100 Tensor Core GPU (SXM) NVIDIA H100 (SXM) | 2022 | 67 TFLOP/s | Vendor-rated | |
| Valve Steam Deck | February 2022 | 1.6 TFLOP/s | Vendor-rated | |
| NVIDIA GeForce RTX 4090 NVIDIA Ada Lovelace (NVIDIA GeForce RTX 4090) | October 2022 | 82.58 TFLOP/s | Theoretical | |
| AMD Radeon RX 7900 XTX AMD RDNA 3 (AMD Radeon RX 7900 XTX) | December 2022 | 61.42 TFLOP/s | Theoretical | |
| Apple Mac mini (M4, 2024) | November 2024 | 4.3 TFLOP/s | Theoretical provisional | |
| NVIDIA GeForce RTX 5090 NVIDIA GB202 (GeForce RTX 5090) | January 2025 | 104.8 TFLOP/s | Theoretical | |
| Nintendo Switch 2 Docked mode | June 2025 | 3.072 TFLOP/s | Theoretical provisional | |
| Atari 2600 Television Interface Adapter | September 1977 | not applicable This is a tile, sprite or scanline graphics unit, not a floating-point processor; it executes no FP32 operations. | not plotted | — |
| Intel 8086 PC IBM PC display adapter | 1978 | not applicable These raster and text display adapters have no floating-point execution hardware. | not plotted | — |
| Intel 80286 PC IBM PC display adapter | 1982 | not applicable These raster and text display adapters have no floating-point execution hardware. | not plotted | — |
| Intel 80386DX PC VGA-compatible display adapter | 1985 | not applicable The represented VGA raster controller class has no programmable floating-point shading hardware. | not plotted | — |
| Nintendo Entertainment System Ricoh 2C02 PPU | October 1985 | not applicable This is a tile, sprite or scanline graphics unit, not a floating-point processor; it executes no FP32 operations. | not plotted | — |
| Sega Master System Sega 315-5124 VDP | June 1986 | not applicable This is a tile, sprite or scanline graphics unit, not a floating-point processor; it executes no FP32 operations. | not plotted | — |
| NEC PC Engine Hudson Soft HuC6270 VDC | October 1987 | not applicable This is a tile, sprite or scanline graphics unit, not a floating-point processor; it executes no FP32 operations. | not plotted | — |
| Sega Mega Drive Sega Mega Drive VDP | October 1988 | not applicable This is a tile, sprite or scanline graphics unit, not a floating-point processor; it executes no FP32 operations. | not plotted | — |
| Intel 80486DX PC VGA-compatible display adapter | 1989 | not applicable The represented VGA raster controller class has no programmable floating-point shading hardware. | not plotted | — |
| Nintendo Game Boy Game Boy LCD controller | April 1989 | not applicable This is a tile, sprite or scanline graphics unit, not a floating-point processor; it executes no FP32 operations. | not plotted | — |
| SNK Neo Geo AES SNK LSPC2-A2 | April 1990 | not applicable This is a tile, sprite or scanline graphics unit, not a floating-point processor; it executes no FP32 operations. | not plotted | — |
| Super Nintendo Entertainment System Super Nintendo PPU pair | August 1991 | not applicable This is a tile, sprite or scanline graphics unit, not a floating-point processor; it executes no FP32 operations. | not plotted | — |
| Intel Pentium PC VGA-compatible display adapter | 1993 | not applicable The represented VGA raster controller class has no programmable floating-point shading hardware. | not plotted | — |
| Sega Saturn | November 1994 | unknown Sega states 25 MIPS for each SH2 and no floating-point rate at any precision. | not plotted | — |
| Sony PlayStation | December 1994 | not applicable Sony’s developer reference states that the geometry transform engine works in fixed-point format. A machine whose 3D arithmetic is fixed point has no floating-point rate to state, in the same way the 6510 has none. | not plotted | — |
| Nintendo 64 | June 1996 | unknown The programming manual describes the Reality Coprocessor by what its microcode does and never rates it in floating-point operations per second. | not plotted | — |
| 3dfx Voodoo Graphics 3dfx Interactive Voodoo (3dfx Voodoo Graphics) | November 1996 | unknown No reviewed source establishes this figure for a Voodoo Graphics board; fill rate is not treated as FLOPS. | not plotted | — |
| Intel Pentium II PC VGA-compatible display adapter | 1997 | not applicable The represented VGA raster controller class has no programmable floating-point shading hardware. | not plotted | — |
| Game Boy Advance Game Boy Advance PPU | March 2001 | not applicable This is a tile, sprite or scanline graphics unit, not a floating-point processor; it executes no FP32 operations. | not plotted | — |
| Nintendo GameCube | September 2001 | unknown Nintendo does publish a floating-point figure for this console — 10,5 GFLOPS (máximo) — and its own wording makes it a total across the processor, the geometry engine and the hardware lighting. A precision the publisher leaves unsaid is something this catalog now reads as single precision and carries as rumored, but an aggregate across unlike units is a different quantity, and no community article separates it. There is no standalone graphics rate to record. | not plotted | — |
| Microsoft Xbox | November 2001 | unknown The registered technical source does not yet provide a standalone GPU FP32 rate with a stated precision and method. | not plotted | — |
| Intel Pentium 4 PC VGA-compatible display adapter | 2002 | not applicable The represented VGA raster controller class has no programmable floating-point shading hardware. | not plotted | — |
| ATI Radeon 9700 Pro ATI Technologies R300 (ATI Radeon 9700 Pro) | August 2002 | unknown The cited vendor material states no FP32 peak in FLOP/s. | not plotted | — |
| NVIDIA GeForce 6800 Ultra NVIDIA GeForce 6 (NVIDIA GeForce 6800 Ultra) | April 2004 | unknown The cited vendor material states no FP32 peak in FLOP/s. | not plotted | — |
| Nintendo DS Nintendo DS 2D engines | November 2004 | not applicable This is a tile, sprite or scanline graphics unit, not a floating-point processor; it executes no FP32 operations. | not plotted | — |
| Intel Core 2 Duo E6600 PC Intel integrated processor graphics | July 2006 | unknown Intel's specification record for the newest of these graphics devices states an execution-unit count of 32 and no floating-point rate at any precision. An execution-unit count is not a rate: converting one to the other needs a published figure for floating-point operations per unit per clock, which the specification does not give. | not plotted | — |
| Nintendo Wii | November 2006 | unknown The registered technical source does not yet provide a standalone GPU FP32 rate with a stated precision and method. | not plotted | — |
| Apple iPhone 3GS | June 2009 | unknown Apple publishes no standalone FP32 throughput figure for this phone GPU. | not plotted | — |
| Apple iPhone 4 | June 2010 | unknown Apple publishes no standalone FP32 throughput figure for this phone GPU. | not plotted | — |
| NVIDIA GeForce GTX 580 NVIDIA Fermi (NVIDIA GeForce GTX 580) | November 2010 | unknown The launch table gives CUDA-core and processor-clock figures, but no explicitly labelled single-precision throughput. | not plotted | — |
| Intel Core i7-2600K PC Intel integrated processor graphics | January 2011 | unknown Intel's specification record for the newest of these graphics devices states an execution-unit count of 32 and no floating-point rate at any precision. An execution-unit count is not a rate: converting one to the other needs a published figure for floating-point operations per unit per clock, which the specification does not give. | not plotted | — |
| Nintendo Wii U | November 2012 | unknown Nintendo's specification table states no floating-point rate, and names no precision anywhere. | not plotted | — |
| Apple iPhone 5s | September 2013 | unknown Apple publishes no standalone FP32 throughput figure for this phone GPU. | not plotted | — |
| NVIDIA GeForce GTX 780 Ti NVIDIA Kepler (NVIDIA GeForce GTX 780 Ti) | November 2013 | unknown The page gives CUDA-core and clock figures but no explicitly labelled single-precision throughput. | not plotted | — |
| Apple iPhone 6s | September 2015 | unknown Apple publishes no standalone FP32 throughput figure for this phone GPU. | not plotted | — |
| Apple iPhone X | November 2017 | unknown Apple publishes no standalone FP32 throughput figure for this phone GPU. | not plotted | — |
| Apple iPhone 12 | October 2020 | unknown Apple publishes no standalone FP32 throughput figure for this phone GPU. | not plotted | — |
| Apple iPhone 15 Pro | September 2023 | unknown Apple publishes no standalone FP32 throughput figure for this phone GPU. | not plotted | — |
| Intel Core i9-14900K PC Intel integrated processor graphics | October 2023 | unknown Intel's specification record for the newest of these graphics devices states an execution-unit count of 32 and no floating-point rate at any precision. An execution-unit count is not a rate: converting one to the other needs a published figure for floating-point operations per unit per clock, which the specification does not give. | not plotted | — |
| Raspberry Pi 5 | October 2023 | unknown The product brief names the VideoCore VII GPU but does not publish an FP32 peak. | not plotted | — |
| Apple iPhone 17 Pro | September 2025 | unknown No checked source states a reproducible FP32 peak for the six-core A19 Pro GPU. | not plotted | — |
| Apple iPhone 17 | September 2025 | unknown Apple publishes no standalone FP32 throughput figure for the A19 GPU. | not plotted | — |
Why this chart holds one group only
Every point on a chart is implicitly compared with every other point on it. This chart uses one comparability group: one metric, one device scope, one method . There is no chart of all clock frequencies over time because they are not one quantity.
Points use the release date of the machine a figure belongs to, including figures recorded against a part. A release date applies only to a machine that shipped. What this metric measures.