A compute equivalent, not an LLM benchmark. We assume 14 billion FLOPs per token for a dense 7B model.
tok/s = GFLOPS ÷ 14 No model is loaded; memory, quantization, and inference overhead are not represented.The reference bench
Your live measurement alongside 20 illustrative device profiles. Reference values are seeded examples, not verified benchmarks.
| # | Device / environment | Class | CPU GFLOPS ↓ | Tok/s-equiv. | Source |
|---|---|---|---|---|---|
| ↳ | This browserYOULocal device · measuring automatically | Live run | Measuring… | Measuring… | In progress |
| 01 | Apple M4 MaxMacBook Pro · macOS | Desktop | 18.80 | 1.343 | Illustrative |
| 02 | AMD Ryzen 9 9950XCustom workstation · Linux | Desktop | 17.40 | 1.243 | Illustrative |
| 03 | Intel Core Ultra 9 285KCustom workstation · Windows | Desktop | 16.20 | 1.157 | Illustrative |
| 04 | Apple M4Mac mini · macOS | Desktop | 15.10 | 1.079 | Illustrative |
| 05 | Apple M3 ProMacBook Pro · macOS | Laptop | 13.80 | 0.986 | Illustrative |
| 06 | AMD Ryzen 7 7840UFramework 13 · Linux | Laptop | 12.20 | 0.871 | Illustrative |
| 07 | Intel Core i7-13700KCustom desktop · Windows | Desktop | 11.60 | 0.829 | Illustrative |
| 08 | Apple M2MacBook Air · macOS | Laptop | 10.90 | 0.779 | Illustrative |
| 09 | Apple A18 ProiPhone 16 Pro · iOS | Mobile | 10.20 | 0.729 | Illustrative |
| 10 | Intel Core Ultra 7 155HThinkPad X1 · Windows | Laptop | 9.40 | 0.671 | Illustrative |
| 11 | Apple M1MacBook Air · macOS | Laptop | 8.60 | 0.614 | Illustrative |
| 12 | Snapdragon 8 Gen 3Android reference phone | Mobile | 7.70 | 0.550 | Illustrative |
| 13 | AMD Ryzen 5 5600XCustom desktop · Windows | Desktop | 7.20 | 0.514 | Illustrative |
| 14 | Apple A15 BioniciPhone 13 · iOS | Mobile | 6.60 | 0.471 | Illustrative |
| 15 | Intel Core i5-1135G7Surface Laptop 4 · Windows | Laptop | 5.90 | 0.421 | Illustrative |
| 16 | Snapdragon 7+ Gen 2Android reference phone | Mobile | 5.20 | 0.371 | Illustrative |
| 17 | Google Tensor G3Pixel 8 · Android | Mobile | 4.50 | 0.321 | Illustrative |
| 18 | Intel Core i5-8250UThinkPad T480 · Linux | Laptop | 3.80 | 0.271 | Illustrative |
| 19 | Raspberry Pi 58 GB · Raspberry Pi OS | Desktop | 2.40 | 0.171 | Illustrative |
| 20 | Intel Celeron N4020Chromebook · ChromeOS | Laptop | 1.30 | 0.093 | Illustrative |
No reference devices match. Try another name or category.
20 reference profiles + your deviceYour row stays pinned · highest reference throughput first
01 / Measure
Real work. Right here.
WebAssembly processes a 1 MiB float32 buffer with two operations per element. After warm-up, we report the median of three sustained samples. This is a single-threaded CPU test.
02 / Translate
A common unit, with context.
Each token is modeled as two FLOPs per parameter. GPU throughput gets its own equivalent; it is never added to CPU throughput. Real model performance will differ substantially.
03 / Compare
Keep the conditions in mind.
Browser, temperature, background activity, and power mode all affect results. Keep this tab visible and rerun for comparison. Reference profiles are illustrative, stored entirely in this page.