How would the performance of a workstation with 2 V100s compare to one with 2 3090s?

Good People, One aspect that remains unclear to me is the comparison of “tensor performance.” The V100 boasts 130 TFLOPs, but I couldn’t find information on the 3090’s “tensor performance” specifically; only figures for fp16, 32, and 64, all of which surpass the V100’s.

This leads me to assume that the “tensor performance” of the 3090 would also be higher. I recall seeing figures for the 3090 pertaining to sparse networks, which drew significant criticism. However, advocates argue that nearly any network can be sparsified without much difficulty.

In summary, would these two builds be comparable?

4 Likes

heeey…Don’t be fooled by the hefty price tag of data center cards like the V100. While they boast features like bigger memory and better double-precision for science tasks, for everyday deep learning, a powerful consumer card like the Titan RTX might be just as good. These data center cards are designed for large-scale operations and cloud usage, not your personal computer. So, save your money unless you’re working with massive datasets or super complex models.

3 Likes

Regarding the V100’s target audience and overall performance of data center cards, there is a widespread misperception. Even the priciest consumer cards are considerably less expensive than the V100. Is it, then, several times better? Actually, no, although it varies. You won’t see many advantages in your daily deep learning tasks, except from having a larger memory. The new 3090 or a Titan are both noticeably quicker than the V100 when it comes to raw FP32 performance. These cards are very costly due to the characteristics they provide, which include improved double-precision performance for scientific applications,Higher-interconnect bandwidth, again crucial for data centers, or more tensor cores, respectively FP16 performance, are examples of virtualization that is typically required by data center/cloud operators as their consumer cards prohibit this kind of utilization. The latter is undoubtedly a selling advantage, but bear in mind that mixed-precision training typically requires a significant engineering effort that is, in my opinion, unnecessary unless you truly need to scale to enormous models and volumes of data.

In summary, if you don’t want to manage a data center, just get a Geforce. It will cost less even if there are any significant drawbacks.

1 Like

@Aiden How recently did you perform mixed precision? I don’t know if you know, but with the most recent versions of the framework, it’s typically simply two extra lines. Not what I would call an engineering endeavor.

While some networks still require more attention, the difference is getting smaller with each new framework and CUDNN upgrade.

1 Like