Ollama LLM SBC Benchmark Rankings
Compare local large language model speed and efficiency across Single Board Computers
66 boards tested · speeds from 0.2 to 23.7 (Tokens/s, DeepSeek R1 1.5B)
Top Picks
Full Results
- 1Fastest Overall: Radxa Fogwise Airbox Q900 36GBIQ-9075 · $8491.37 tok/s/W1.37 tok/s/WIQ-9075 · $849
- 2Orange Pi Zero 3W 4GBA733 · $501.31 tok/s/W1.31 tok/s/WA733 · $50
- 3Raspberry Pi 5 16GBBCM2712 · $3291.31 tok/s/W1.31 tok/s/WBCM2712 · $329
- 4Radxa ROCK 4D 8GBRK3576 · $1061.3 tok/s/W1.3 tok/s/WRK3576 · $106
- 5Raspberry Pi 5 8GBBCM2712 · $1891.24 tok/s/W1.24 tok/s/WBCM2712 · $189
- 6Dshan Pi A1 4GBRK3576 · $691.19 tok/s/W1.19 tok/s/WRK3576 · $69
- 7Radxa Dragon Q6A 6GBQCS6490 · $1191.18 tok/s/W1.18 tok/s/WQCS6490 · $119
- 8Raspberry Pi 500 8GBBCM2712 · $1941.13 tok/s/W1.13 tok/s/WBCM2712 · $194
- 9Radxa CM5 Lite 8GBRK3582 · $1191.09 tok/s/W1.09 tok/s/WRK3582 · $119
- 10Raspberry Pi 5 4GBBCM2712 · $1181.09 tok/s/W1.09 tok/s/WBCM2712 · $118
- 11Raspberry Pi 5 2GBBCM2712 · $821.07 tok/s/W1.07 tok/s/WBCM2712 · $82
- 12Libre Computer Alta 4GBA311D · $1251.05 tok/s/W1.05 tok/s/WA311D · $125
- 13Raspberry Pi Compute Module 5 4GBBCM2712 · $1001.04 tok/s/W1.04 tok/s/WBCM2712 · $100
- 14ArmSoM Sige5 8GBRK3576 · $2391.01 tok/s/W1.01 tok/s/WRK3576 · $239
- 15Forlinx FET3576-C 2GBRK3576 · $2100.96 tok/s/W0.96 tok/s/WRK3576 · $210
- 16Best Value: Radxa ROCK 2F 2GBRK3528A · $150.93 tok/s/W0.93 tok/s/WRK3528A · $15
- 17Radxa Dragon Q8B 32GBSnapdragon 8cx Gen 3 · $6770.91 tok/s/W0.91 tok/s/WSnapdragon 8cx Gen 3 · $677
- 18Banana Pi M4 Zero 4GBH618 · $1180.9 tok/s/W0.9 tok/s/WH618 · $118
- 19Orange Pi Zero 3 4GBH618 · $400.9 tok/s/W0.9 tok/s/WH618 · $40
- 20Orange Pi 5B 4GBRK3588S · $750.88 tok/s/W0.88 tok/s/WRK3588S · $75
- 21Orange Pi Zero 4 4GBA733 · $680.88 tok/s/W0.88 tok/s/WA733 · $68
- 22Raspberry Pi 500+ 16GBBCM2712 · $4430.88 tok/s/W0.88 tok/s/WBCM2712 · $443
- 23Orange Pi 5 Plus 8GBRK3588 · $1350.87 tok/s/W0.87 tok/s/WRK3588 · $135
- 24Radxa ROCK 2A 4GBRK3528A · $300.85 tok/s/W0.85 tok/s/WRK3528A · $30
- 25Banana Pi M5 4GBS905X3 · $1460.84 tok/s/W0.84 tok/s/WS905X3 · $146
- 26Radxa NX5 8GBRK3588S · $1190.82 tok/s/W0.82 tok/s/WRK3588S · $119
- 27Orange Pi 5 Pro 16GBRK3588S · $1200.78 tok/s/W0.78 tok/s/WRK3588S · $120
- 28ArmSoM AIM7 8GBRK3588 · $2970.77 tok/s/W0.77 tok/s/WRK3588 · $297
- 29Banana Pi R4 8GBMT7988A · $1460.73 tok/s/W0.73 tok/s/WMT7988A · $146
- 30Libre Computer Solitude 4GBS905D3 · $1000.73 tok/s/W0.73 tok/s/WS905D3 · $100
- 31Orange Pi 4A 4GBT527 · $500.73 tok/s/W0.73 tok/s/WT527 · $50
- 32Orange Pi Zero 2W 4GBH618 · $400.72 tok/s/W0.72 tok/s/WH618 · $40
- 33Raspberry Pi Compute Module 4 (CM4) 2GBBCM2711 · $620.72 tok/s/W0.72 tok/s/WBCM2711 · $62
- 34DEBIX Model A 4GBiMX 8M Plus · $1350.71 tok/s/W0.71 tok/s/WiMX 8M Plus · $135
- 35Radxa Zero 3W 8GBRK3566 · $920.71 tok/s/W0.71 tok/s/WRK3566 · $92
- 36Cool Pi 4B 4GBRK3588S · $1400.7 tok/s/W0.7 tok/s/WRK3588S · $140
- 37Radxa ROCK 5T 12GBRK3588 · $1250.69 tok/s/W0.69 tok/s/WRK3588 · $125
- 38Fastest Under $100: Radxa X4 8GBN100 · $1000.69 tok/s/W0.69 tok/s/WN100 · $100
- 39ArmSoM W3 8GBRK3588 · $3160.68 tok/s/W0.68 tok/s/WRK3588 · $316
- 40Radxa Orion O6N 32GBP1 · $5490.68 tok/s/W0.68 tok/s/WP1 · $549
- 41Libre Computer Le Potato 2GBS905X · $600.67 tok/s/W0.67 tok/s/WS905X · $60
- 42Raspberry Pi 4 2GBBCM2711 · $720.65 tok/s/W0.65 tok/s/WBCM2711 · $72
- 43LattePanda Mu 8GBN100 · $3980.62 tok/s/W0.62 tok/s/WN100 · $398
- 44Radxa ROCK 5 ITX 8GBRK3588 · $1200.62 tok/s/W0.62 tok/s/WRK3588 · $120
- 45Orange Pi CM4 2GBRK3566 · $400.58 tok/s/W0.58 tok/s/WRK3566 · $40
- 46Libre Computer Renegade 2GBRK3328 · $450.56 tok/s/W0.56 tok/s/WRK3328 · $45
- 47Orange Pi 6 Plus 32GBP1 · $6690.56 tok/s/W0.56 tok/s/WP1 · $669
- 48Raspberry Pi 400 4GBBCM2711 · $650.56 tok/s/W0.56 tok/s/WBCM2711 · $65
- 49DEBIX Model C 2GBiMX 93 · $650.54 tok/s/W0.54 tok/s/WiMX 93 · $65
- 50Orange Pi 3B 4GBRK3566 · $940.53 tok/s/W0.53 tok/s/WRK3566 · $94
- 51Radxa 4 SE 4GBRK3399-T · $800.53 tok/s/W0.53 tok/s/WRK3399-T · $80
- 52Radxa ROCK 4C+ 4GBRK3399-T · $800.5 tok/s/W0.5 tok/s/WRK3399-T · $80
- 53Radxa ROCK 3C 2GBRK3566 · $400.44 tok/s/W0.44 tok/s/WRK3566 · $40
- 54Radxa ROCK 4B+ 4GBRK3399 OP1 · $1700.35 tok/s/W0.35 tok/s/WRK3399 OP1 · $170
- 55ADLINK I-Pi SMARC 1200 4GBMT8395 · $3940.33 tok/s/W0.33 tok/s/WMT8395 · $394
- 56Banana Pi CM4 4GBA311D · $1040.18 tok/s/W0.18 tok/s/WA311D · $104
- 57Radxa Cubie A7A 6GBA733 · $800.09 tok/s/W0.09 tok/s/WA733 · $80
- 58Radxa ROCK 5B+ 16GBRK3588 · $1350.09 tok/s/W0.09 tok/s/WRK3588 · $135
- 59Libre Computer Renegade Elite 4GBRK3399 · $1000.04 tok/s/W0.04 tok/s/WRK3399 · $100
- 60Banana Pi M2S (S922X) 4GBS922X · $670.03 tok/s/W0.03 tok/s/WS922X · $67
- 61ArmSoM Sige7 8GBRK3588 · $3120.02 tok/s/W0.02 tok/s/WRK3588 · $312
- 62Radxa ROCK 5A 4GBRK3588S · $1300.02 tok/s/W0.02 tok/s/WRK3588S · $130
- 63Radxa ROCK 5B 8GBRK3588 · $1600.02 tok/s/W0.02 tok/s/WRK3588 · $160
- 64Radxa ROCK 5C 4GBRK3588S2 · $1660.02 tok/s/W0.02 tok/s/WRK3588S2 · $166
Frequently Asked Questions
What is the fastest single board computer in Ollama LLM (DeepSeek R1 1.5B)?
The Radxa Fogwise Airbox Q900 36GB currently tops our Ollama LLM (DeepSeek R1 1.5B) ranking with 23.7 (Tokens/s), measured on real hardware with our standardised test setup.
What is a good Ollama LLM (DeepSeek R1 1.5B) score for an SBC?
Across the 66 boards we have tested, Ollama LLM (DeepSeek R1 1.5B) results range from 0.2 to 23.7 (Tokens/s), with a median of 5.61. Anything above the median puts a board in the faster half of the market.
About Ollama LLM SBC Benchmark Rankings testing
This ranking shows how fast each board runs a large language model locally with Ollama. We use the DeepSeek R1 distilled models and run each one three times through the ollama-benchmark script, reporting the average generation speed (eval rate) in tokens per second: how quickly the answer appears, not counting prompt processing. Ollama runs on the CPU on these boards, so RAM size decides which models we can run: every board with around 2GB or more gets the 1.5B model, 8GB boards add the 8B model, and 16GB boards add 14B. Each model has its own ranking; switch between them above. Tokens per watt divides that speed by the board's peak power draw during the run, measured with a USB power meter, and tokens per dollar divides it by the board's current price. Higher is better for all three.