2

I run the exactly same code for my 2 computers.

Both of them are installed of tensorflow-gpu==2.7.0 | CUDA = 11.2 | cudnn = 8.1.

For simple Autoencoder code. (Epoch : 300)

3090 takes 1m 57.4s and 3060 takes 52.9 s.

3090 memory is 24 GB and 3060 memory is 12 GB. (I used both of them full memory and I can see it uses fully by task manager)

What should I do?

Until now I can't understand my situation.

Pls give me some advice what should I have to check.

<1> RTX 3090 computer specification

(1) CPU : Intel Xeon Silver 4310 CPU @ 2.10GHz

(2) Memory : 128 GB

<2> RTX 3060 computer specification

(1) Intel Xeon W-2245 CPU @ 3.90GHz

(2) Memory : 128 GB

3060 computer uses CPU 24~27% (4.49GHz) | memory 9%

3090 computer uses CPU 11~15% (1.38~1.56GHz) | memory 7%

Code is below

    input_data = keras.Input(shape=(20,))
    encoded = Dense(15, activation='relu')(input_data)
    encoded = Dense(5, activation='relu')(encoded)
    decoded = Dense(15, activation='relu')(encoded)
    decoded = Dense(20, activation='relu')(decoded)

    autoencoder = keras.Model(input_data, decoded)
    autoencoder.compile(optimizer='adam', loss='mse')

    autoencoder.fit(sc_x_train, sc_x_train,
                    epochs=300,
                    batch_size=16,
                    shuffle=True,
                    verbose=0)

    encoder = keras.Model(input_data, encoded)
    encoder.compile(optimizer='adam', loss='mse')
MCPMH
  • 175
  • 1
  • 1
  • 11
  • Did you check CPU and System RAM usage? The GPU might not be the cause. After all, the 3090 should definitely be faster if it is a GPU only task. – Torge Rosendahl Aug 09 '22 at 11:35
  • As your comment, I found different usage of CPU. How I can use full GPU and CPU together? I know how to control GPU but I haven't known about controlling CPU usage for keras – MCPMH Aug 09 '22 at 11:58
  • I would venture a guess that you're really running that load on primarily the CPU (and potentially in a less-than-ideal threaded design), as the 4310 has a much higher clock speed with similar IPC (in fact, nearly double in some cases). Without seeing the program you're running (and how you're running it), I don't see how others could begin diagnosing the issue. Try and provide an [MCVE](https://stackoverflow.com/help/minimal-reproducible-example) of your problem. – Rogue Aug 09 '22 at 12:17
  • Thanks for your comment. I uploaded my code. But the strange thing is that Both computers use different CPU usage percentage. – MCPMH Aug 09 '22 at 12:28
  • Have you found the answer? It seems the same case is for me. RTX3060 on works much faster on laptop "ASUS ROG 5020 Ryzen 4800H RTX3060 6GB 16GB RAM" than on desktop with "RTX3090 ryzen 7950x 64GB ddr5 5600". Do you know why? – Krzysztof Cichocki Mar 07 '23 at 20:14
  • After i changed cpu with same one, 3090 works faster. however, it's something like not double but just little faster. not dramatically faster. – MCPMH Jun 16 '23 at 11:41

0 Answers0