Hacker News Reader

Comment by ml_hardware | Hacker News Reader

This is a poor comparison of performance. All of these networks are CNNs, and very old architectures at that. They are all probably memory bottlenecked which is why you see the consistent 50% improvement in FP32 perf.

It is also not clear what batch sizes are being used for any of the tests. If you switch to FP16 training, you must increase the batch size to properly utilize the Tensor Cores.

If you compare these cards at FP16 performance on large language models (think GPT-style with large model dimension), I am confident you will see Titan RTX outperform the 3090. The former has 130 TF/s of FP16.32 tensor core performance while the latter has only 70 TF/s.

Link: https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/a...

View on HN

The 3090 RTX is also $1000 cheaper than the Titan, so there's that. It would be nice if there was a good way to express value per dollar. Perhaps in GLUE accuracy and training time.

Totally agree! I think 3090 could be a lot more cost effective for researchers to dabble with NLP. But it really grinds my gears when people post these misleading benchmarks... the 3090 is handicapped at half-rate tensor core performance while the Titan RTX is not.

So if you're someone who does their work mainly in FP32, you will see improved performance with the 3090. On the other hand, if you are an FP16 speed demon who needs to train GPT-3 over the weekend, stick with your Titans :)

What do you think about TF32 in 3090? Could it replace FP32 with 5x speedup?

I've done a lot of work in ML numerics, and I think TF32 is a completely safe drop-in for FP32 for ML workloads. NVIDIA seems to think so too, which is why on A100 it won't even be an option, it will be the default mode for any FP32 matrix multiplies.

But on 3090, I don't think the speedup will be 5x, it should be closer to like 2x. The 3090 has 35.6 TF/s at TF32 and the Titan RTX has 16.3 TF/s at FP32. Once again I think there is handicapping going on for 3090.

So basically no difference to FP32. That sounds very handicapped.

For many of us, the Inception-style CNN workloads--especially at FP32--are much more realistic than large language models that may be better suited to take advantage of the tensor cores. If I'm going to be memory bottlenecked either way, I probably don't want to spend an extra $1000 on 400 tensor cores I can't take full advantage of.

If I may ask, why are the Inception style workloads still popular, rather than architectures like EfficientNet?

Also, why FP32? CNNs are some of the most robust models to train in FP16 (much easier than language models) so you could get yourself a quick XXX speedup and 2x memory savings by switching over.

(btw not intending to be accusatory or anything, I just think FP16 training deserves a lot more adoption that it currently seems to have :)

Reply on news.ycombinator.com