Ampere vs Turing Tensor Cores

[ixwild]

I've got a 3080 on order and really looking forward to using the tensor cores in some python stuff I'm doing, I noticed though that the 3080 has the same number of tensor cores as the 2060 super and far lower than the 2080s and such. Would it be appropriate to assume the 3080 will be far slower in these tasks compared to the same card from the Turing range?

It's hard to find benchmarks on this stuff at the moment so any advice is appreciated!

 

[Rroff]

Whitepaper with a lot of information here: https://www.nvidia.com/content/dam/...pere-GA102-GPU-Architecture-Whitepaper-V1.pdf

Pertinent bit:

Unlike the TU102 SM which includes two second-generation Tensor Cores per partition and eight Tensor Cores total, the new GA10x SM includes one third-generation Tensor Core per partition and four Tensor Cores total, with each GA10x Tensor Core being twice as powerful as a Turing Tensor Core.

Off the top of my head Tensor core performance per GPU is something like a minimum of 40% faster on a 3080 compared to a 2080S.

EDIT: As below they are also more advanced and allow acceleration considerably above that depending on what you are doing and how you approach it.

 

[alec]

https://developer.nvidia.com/blog/nvidia-ampere-architecture-in-depth/

New TensorFloat-32 (TF32) Tensor Core operations in A100 provide an easy path to accelerate FP32 input/output data in DL frameworks and HPC, running 10x faster than V100 FP32 FMA operations or 20x faster with sparsity. For FP16/FP32 mixed-precision DL, the A100 Tensor Core delivers 2.5x the performance of V100, increasing to 5x with sparsity.

 

[PrincessFrosty]

Tensor performance is way up on the 30 series as you'd expect, about 2.7X performance according to the Nvidia release. They've doubled the size of the Tensor cores and added in some efficiency gains.

89 Tensor-TFLOPS of last gen to 238 Tensor-TFLOPS of Ampere

https://youtu.be/E98hC9e__Xs?t=1023

 

[ixwild]

Whitepaper with a lot of information here: https://www.nvidia.com/content/dam/...pere-GA102-GPU-Architecture-Whitepaper-V1.pdf

Pertinent bit:



Off the top of my head Tensor core performance per GPU is something like a minimum of 40% faster on a 3080 compared to a 2080S.

EDIT: As below they are also more advanced and allow acceleration considerably above that depending on what you are doing and how you approach it.

https://developer.nvidia.com/blog/nvidia-ampere-architecture-in-depth/

Tensor performance is way up on the 30 series as you'd expect, about 2.7X performance according to the Nvidia release. They've doubled the size of the Tensor cores and added in some efficiency gains.

89 Tensor-TFLOPS of last gen to 238 Tensor-TFLOPS of Ampere

https://youtu.be/E98hC9e__Xs?t=1023

Thank you all very much! Even more excited to try it out now

 

[clv101]

... really looking forward to using the tensor cores in some python stuff I'm doing,

What kind of stuff are you using them for?

 

[ixwild]

What kind of stuff are you using them for?

Nothing crazy, I've been generally using other peoples stuff that they've made and giving it my own data so far like Jukebox AI (had to use my dads 2070), but I've got my dissertation for uni coming up and part of my project is using TensorFlow for image classification to spot fake logos on phishing emails. I haven't worked out the fine details yet but what I'm hoping to do is feed it a data set of logos (which I've found created for this purpose) then compare its bank of images with the one that come in the email on a variety of things like first can it find a match then if it can find a match how close is it to the image it has and so on.

 

[FxUK]

But, don't phishing emails generally use real source images and logo's, so as to come across as official.. I know, I know, not the point

You are number 1,059,882 in the queue for a GPU, pay us an extra £100
and we will reposition you closer to the front of the queue, also click this link
to give us all your bank information, because we can't find the post-it note

Ofishal emale from ocverlockers uk.

 

[ixwild]

But, don't phishing emails generally use real source images and logo's, so as to come across as official.. I know, I know, not the point

You are number 1,059,882 in the queue for a GPU, pay us an extra £100
and we will reposition you closer to the front of the queue, also click this link
to give us all your bank information, because we can't find the post-it note

Ofishal emale from ocverlockers uk.

Generally a lot do but you'd be surprised how many are slightly off, the whole project is to first filter out the emails then once they've been filtered determine which ones would've been most likely to succeed based on a load of factors I'm yet to determine through research but likely stuff just like non-dodgy email addresses, spelling and grammar. Then use the most succesful ones as templates for training staff. I'm obviously super early on in the project and so this is all very likely to change haha.

Wow only £100! What a bargain

 

[FxUK]

Interesting stuff nonetheless