Dell Precision T7920 Dual Intel Xeon Workstation Review

8

Dell Precision T7920 Workstation Deep Learning Benchmarks

Deep learning training and inferencing workloads utilize GPUs heavily. Here, we can lean on our large data set to give some comparisons showing just how fast the Quadro RTX 8000 is.

ResNet-50 Inferencing in TensorTR using Tensor Cores.

ImageNet is an image classification database launched in 2007 designed for use in visual object recognition research. Organized by the WordNet hierarchy, hundreds of image examples represent each node (or category of specific nouns).

NVIDIA Quadro RTX 8000 ResNet50 Inferencing INT8 Precision
NVIDIA Quadro RTX 8000 ResNet50 Inferencing INT8 Precision

Using precision of INT8 is by far the fastest inferencing method if at all possible converting code to INT8 will yield faster runs. Installed memory has one of the largest impacts on these benchmarks which the Inferencing on NVIDIA RTX graphics cards does not tax the GPU’s to a great deal, however additional memory allows for larger batch sizes, the NVIDIA Quadro RTX 8000 could easily do batch sizes of 500+.

Let’s look at FP16 and FP32 results.

NVIDIA Quadro RTX 8000 ResNet50 Inferencing FP16 Precision
NVIDIA Quadro RTX 8000 ResNet50 Inferencing FP16 Precision
NVIDIA Quadro RTX 8000 ResNet50 Inferencing FP32 Precision
NVIDIA Quadro RTX 8000 ResNet50 Inferencing FP32 Precision

For inferencing workloads using NVIDIA CUDA, the Dell Precision T7920 performs extremely well.

ResNet-50 Training using Tensor Cores.

We also wanted to train the venerable ResNet-50 using Tensorflow. During training the neural network is learning features of images, (e.g. objects, animals, etc.) and determining what features are important. Periodically (every 1000 iterations), the neural network will test itself against the test set to determine training loss, which affects the accuracy of training the network. Accuracy can be increased through repetition (or running a higher number of epochs.)

NVIDIA Quadro RTX 8000 ResNet50 Training FP16 Precision
NVIDIA Quadro RTX 8000 ResNet50 Training FP16 Precision
NVIDIA Quadro RTX 8000 ResNet50 Training FP32 Precision
NVIDIA Quadro RTX 8000 ResNet50 Training FP32 Precision

Here, again we can see excellent performance from the Quadro RTX 8000 in the Dell Precision T7920.

Training using OpenSeq2Seq (GNMT)

While Resnet-50 is a Convolutional Neural Network (CNN) that is typically used for image classification, Recurrent Neural Networks (RNN) such as Google Neural Machine Translation (GNMT) are used for applications such as real-time language translations.

We should note that other consumer GPUs we have tried on this benchmark could not complete training due to the lack of installed memory.

NVIDIA Quadro RTX 8000 OpenSeq2Seq Training FP16 Mixed Precision
NVIDIA Quadro RTX 8000 OpenSeq2Seq Training FP16 Mixed Precision
NVIDIA Quadro RTX 8000 OpenSeq2Seq Training FP32 Mixed Precision
NVIDIA Quadro RTX 8000 OpenSeq2Seq Training FP32 Mixed Precision

As the NVIDIA Quadro RTX 8000 has 48GB of installed memory, double that of the Titan RTX. The Quadro RTX 8000 is easily equal to the Titan RTX but offers larger batch sizes on a single GPU. Offering the ability to house multiple NVIDIA Quadro RTX 8000 GPUs in the Dell Precision T7920 allows for desktop data science to be done at levels formerly reserved for the data center.

Next, we are going to look at AIDA64 CPU Benchmarks.

8 COMMENTS

  1. I’ve got one of these in the office – great machine, and I love the layout of the motherboard (lenovo is also using a layout like that now too), but theres some things to look out for depending on what you’re using it for and how you’re using it:
    1- if you have nosy coworkers who like looking at the hardware in the chassis, just show them the machine before its turned on, since opening the chassis during operation will shut the machine down while triggering a chassis intrusion alarm. At least on my early model, it also bugged out the dell support assist feature so it permanently thinks a fan is broken, when that fan was never installed on that model.
    2- in order to keep the overall width of the unit within reason for product dimensions and cooling reasons, the clearance for cards is roughly equivalent to a 3U server. In places, that is actually slightly less, because the side panel latch has an internal bump running the height of the server. Tall cards like some consumer GPUs will not fit in this case, nor will short-length consumer GPUs that have outward/side facing (as opposed to front facing) pcie power plugs, as the bump will be in the way of the cable. Long standard height consumer GPUs with side facing pcie power plugs will fit, but you might have to squeeze the cabling a bit
    3- There were 3 included pcie 8/6pin cables. If you are using consumer GPUs, you may need to buy additional cables from dell or use splitters. At the time I bought my unit, dell did not have the cables for purchase individually, but I managed to cludge an equivalent cable together out of adapters, since the PSU breakout board’s connectors for the pcie power plugs have the same pinout as (cont in next)

  2. 3 (cont)- the plug would on the pcie card as well.
    4- If you want NVME support, you’ll need an adapter. You can bifurcate the lanes, but not explicitly – there is a setting in the bios (i forget which, but it has to do with pcie ssds), that you can toggle. Out of the box it was set to a setting where it would not detect dell’s own quad nvme adapter card, and dell’s enterprise support doesn’t have enough experience with those cards (or even the product specifications on hand) to do more than manually walk through trial and error experimentation of figuring out how to use the card in this machine. Turns out its that bios setting.
    5- If you need audio for the work you intend to do on this machine and are expecting to get by with the motherboard audio since its usually ‘good enough’ on highend prosumer motherboards, you will be disappointed. Get a soundcard for this – I didn’t need amazing audio or anything, just audio good enough to listen to people speaking during meetings and checking the contents of files I was processing or working with, but the built in audio was really bad.

    It should be kept in mind that my forewarnings are based on a very early unit (I got it within a month of launch, since I needed something more powerful right then in order to complete a project)
    Despite these issues I encountered, I still very much like this machine, and they wont be problems for everyone.

  3. Thanks for the info Syr. I am curious about your point #3 — I have avoided Dell workstations because they only offer 8/6 pin PCIe power connectors, while most consumer GPUs (e.g. 2080 Ti) require 8/8 connectors. It would be great to learn more about how you figured this out.

  4. Hi Michael – to clarify what I had meant by 8/6 was that the cables included with the system supported the full 6+2 connector, but only 1 per cable. The system came with 3 such cables pre-installed, but had 4 available headers, thus I was able to determine the pinout by simply matching the cables on the card side to the PSU breakout board side. It helped that Dell used standard cable color codes (yellow for +12 and black for 0).
    I’m using 2x 1080ti cards in the system with 2x 8-pin plugs on them for a total of 4 plugs required, so I had to cobble together a cable out of adapters to power the second plug on the second card from the 4th empty plug on the PSU breakout board.

  5. Syr–thanks for the reply! So it sounds like there are a maximum of 4 x 8 PCIe plugs available. That’s enough for 2 consumer GPUs. A 3rd GPU may not be possible.

LEAVE A REPLY

Please enter your comment!
Please enter your name here