Performance of RKLLM DeepSeek-R1 on ROCK 5T

Hi, All

In the link below, the performance of the DeepSeek-R1-Distill-Qwen-1.5B_W8A8_RK3588.rkllm model is shown as 15.36 tokens/sec in the generate, but when I tried to run it, 15.36 tokens/sec doesn’t seem like a reasonable number.

https://docs.radxa.com/en/rock5/rock5t/app-development/rkllm_deepseek_r1


Performance on RK3588 reaches 15.36 tokens/s :

Stage Total Time (ms) Tokens Time per Token (ms) Tokens per Second
Prefill 122.70 29 4.23 236.35
Generate 27539.16 423 65.10 15.36

https://docs.radxa.com/en/assets/images/rkllm_ds_3-492e4c4fcea8ff85cea215ebdeec236c.webp


When I actually measured it, I got the numbers below.
12 tokens/s for low and 15 tokens/s for high.

Is the result RADXA tested the maximum value (15.36 tokens/s)? Or is it an average result?
If it’s an average result, why is it low on average in the tests I’ve done?
Or, is everyone else getting a speed of 15.36 tokens/sec?
Let me know your thoughts.

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          1929.53          /         /                        /
I rkllm:  Prefill       335.94           30        11.20                    89.30
I rkllm:  Generate      3014.55          44        68.51                    14.60
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2270.14          /         /                        /
I rkllm:  Prefill       367.06           49        7.49                     133.49
I rkllm:  Generate      22758.78         319       71.34                    14.02
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.72
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2270.14          /         /                        /
I rkllm:  Prefill       0.00             0         0.00                     0.00
I rkllm:  Generate      19324.20         268       72.11                    13.87
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.72
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          1889.13          /         /                        /
I rkllm:  Prefill       290.03           49        5.92                     168.95
I rkllm:  Generate      21828.04         319       68.43                    14.61
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.72
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          1941.29          /         /                        /
I rkllm:  Prefill       533.74           60        8.90                     112.41
I rkllm:  Generate      22971.40         332       69.19                    14.45
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.72
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          1998.16          /         /                        /
I rkllm:  Prefill       299.41           60        4.99                     200.39
I rkllm:  Generate      22896.06         332       68.96                    14.50
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.72
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2143.00          /         /                        /
I rkllm:  Prefill       271.09           41        6.61                     151.24
I rkllm:  Generate      164455.45        2047      80.34                    12.45
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.72
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2143.00          /         /                        /
I rkllm:  Prefill       192.91           33        5.85                     171.07
I rkllm:  Generate      164809.91        2047      80.51                    12.42
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.72
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          1968.99          /         /                        /
I rkllm:  Prefill       205.80           31        6.64                     150.63
I rkllm:  Generate      31269.30         452       69.18                    14.46
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2049.23          /         /                        /
I rkllm:  Prefill       226.38           33        6.86                     145.78
I rkllm:  Generate      32675.22         472       69.23                    14.45
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2049.23          /         /                        /
I rkllm:  Prefill       0.00             0         0.00                     0.00
I rkllm:  Generate      34048.97         489       69.63                    14.36
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2049.23          /         /                        /
I rkllm:  Prefill       413.06           34        12.15                    82.31
I rkllm:  Generate      36204.07         519       69.76                    14.34
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2142.52          /         /                        /
I rkllm:  Prefill       399.41           34        11.75                    85.13
I rkllm:  Generate      40692.71         581       70.04                    14.28
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2142.52          /         /                        /
I rkllm:  Prefill       144.45           8         18.06                    55.38
I rkllm:  Generate      3626.40          55        65.93                    15.17
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2142.52          /         /                        /
I rkllm:  Prefill       149.54           5         29.91                    33.44
I rkllm:  Generate      1952.60          29        67.33                    14.85
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2142.52          /         /                        /
I rkllm:  Prefill       170.69           10        17.07                    58.59
I rkllm:  Generate      6961.55          103       67.59                    14.80
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2142.52          /         /                        /
I rkllm:  Prefill       238.04           8         29.75                    33.61
I rkllm:  Generate      1894.84          26        72.88                    13.72
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2142.52          /         /                        /
I rkllm:  Prefill       191.33           6         31.89                    31.36
I rkllm:  Generate      2450.64          37        66.23                    15.10
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2142.52          /         /                        /
I rkllm:  Prefill       148.70           7         21.24                    47.07
I rkllm:  Generate      3972.95          59        67.34                    14.85
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2142.52          /         /                        /
I rkllm:  Prefill       194.88           4         48.72                    20.53
I rkllm:  Generate      2073.36          31        66.88                    14.95
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2142.52          /         /                        /
I rkllm:  Prefill       250.71           11        22.79                    43.88
I rkllm:  Generate      17271.81         253       68.27                    14.65
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------

I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Stage         Total Time (ms)  Tokens    Time per Token (ms)      Tokens per Second
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Init          2142.52          /         /                        /
I rkllm:  Prefill       236.68           6         39.45                    25.35
I rkllm:  Generate      2615.21          37        70.68                    14.15
I rkllm: --------------------------------------------------------------------------------------
I rkllm:  Memory Usage (GB)
I rkllm:  1.71
I rkllm: --------------------------------------------------------------------------------------