Leadtek has released RTX Quadro Passive graphics optimized for the data center.

Let's see the comparison between active and passive Quadro performance in AI application, 

and choose the right graphics for your workload. 

More Product Detail

Software / hardware configuration

Test items

RTX Passive Performance

<The performance test uses ImageNet with various mainstream deep learning image classification models for comparison.>

<The vertical axis of each chart in this article represents the performance of deep learning operations, based on the number of pictures that can be processed per second by various models (horizontal axis).>

With the introduction of the deep learning library cuDNN 7.6, Tensor Core has become more integrated in the deep learning framework, and the advantage of mixed precision (FP16) performance has become more obvious. Compared with cuDNN 7.3, the deep learning performance of mixed precision can be improved by nearly double. The mixed precision performance improvement with cuDNN version 7.6 is about 60% to 170%, and the performance varies according to the model architecture.

In the multi-GPU test, the performance of two GPUs is improved by 83% to 115% over a single GPU. As the RTX 8000 Passive has a larger GPU memory capacity, the performance improvement is pretty obvious.

Figure 1 Deep Learning Performance of Quadro RTX 6000 Passive with Multi-GPU

Figure 2 Deep Learning Performance of Quadro RTX 8000 Passive with Multi-GPU

Performance comparison between RTX Quadro Passive and Active

The high-end models of Quadro RTX series, RTX 6000 and RTX 8000 have launched Active and Passive versions.

The former has an active cooling fan and can be installed in the workstation; the latter does not have a cooling fan and is suitable for server installation, which can dissipate the heat outside the GPU and the system with the server fans. As can be seen from the table below, the major performance difference between the Active and Passive versions is the GPU clock. The GPU clock speed is the speed of GPU operations, and also directly affects various performance values.

The clock speed of the Passive series is about 9% to 13% lower than that of the Active. In the deep learning model training, the clock speed also directly reflects the performance results. The Active version of both models performs roughly 2% to 17% better than the Passive version.

Figure 3 Comparison of Deep Learning Performance between Quadro RTX 6000 Passive and Active

Figure 4 Comparison of Deep Learning Performance between Quadro RTX 8000 Passive and Active

The advantage of Quadro RTX 6000 Passive is that it has lower power consumption than Active, and it also has better heat dissipation when installed in the server.

As seen in Figure 5, Quadro RTX 6000 Passive maintains an average temperature of around 75oC for a long period of operation, while the Quadro RTX 6000 Active is about 10oC higher than Passive on average. Although the Quadro RTX 6000 Active has not yet reached the GPU HW thermal slowdown threshold (91oC), it may reach the GPU's maximum operating temperature (89oC), resulting in a SW thermal slowdown.

Therefore, the Active version installed in the server may cause the system to downclock the GPU due to the poor heat dissipation.

Figure 5 Thermal test of Quadro RTX 6000 Passive and Active