• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

AMD 7nm GPU News and Rumours 2018/2019

Gibbo said:
Our price won’t be far off.

Let’s just say come 2pm this Thursday many of your jaws will drop and you’ll be opening your wallets as we’re doing something insane and the new games bundle is AMAZING!
Click to expand...
I imagine we're talking about people who haven't already got 290/390/480/580 level cards?

Which must be a fairly small market these days, surely...
 
Gibbo said:
Our price won’t be far off.

Let’s just say come 2pm this Thursday many of your jaws will drop and you’ll be opening your wallets as we’re doing something insane and the new games bundle is AMAZING!
Click to expand...

DayZ just went Beta last week, but i kinda hope it isnt in the new bundle.. :D

Btw @Gibbo can i not sell my copy of Assassin's Creed Odyssey as it states i need to redeem on UPlay? Was there Steam codes for it when the offer started?
 
4K8KW10 said:
If happens, a nice upgrade over 580/590.

About GCN - I don't like it - it may be powerful for compute but AMD did best once VLIW5 ruled.
Click to expand...

GCN isn't a problem as such (though does have some limitations in terms of reliance on fixed function systems that don't scale well and how much you can increase the number of certain systems) they keep persisting in trying to build an implementation with it for a future that never arrives.
 
Rroff said:
GCN isn't a problem as such (though does have some limitations in terms of reliance on fixed function systems that don't scale well and how much you can increase the number of certain systems) they keep persisting in trying to build an implementation with it for a future that never arrives.
Click to expand...

I think with GCN they build not working implementations for the future that is already here. Read 4K gaming performance, 4K everything. Not to mention 8K.
4K and 8K will never happen with GCN.
 
Isn't gcn limited to 64 rops? Seen a few posts a while back talking about this, apparently one of the reasons vega felt more like a reworked fury x.
 
So in the 'Tensorflow Resnet 50 benchmark' AMD didn't allow the competition to use their Tensor cores just because AMD don't have any and it would show their new 7nm flagship card to be rubbish in that particular benchmark.

Well that certainly seems a fair and transparent thing to do.;)
 
.
bru said:
So in the 'Tensorflow Resnet 50 benchmark' AMD didn't allow the competition to use their Tensor cores just because AMD don't have any and it would show their new 7nm flagship card to be rubbish in that particular benchmark.

Well that certainly seems a fair and transparent thing to do.;)
Click to expand...

As I understand it, tensor cores are not used due to poor accuracy.
 
bru said:
So in the 'Tensorflow Resnet 50 benchmark' AMD didn't allow the competition to use their Tensor cores just because AMD don't have any and it would show their new 7nm flagship card to be rubbish in that particular benchmark.
Click to expand...
According to the article it was because enabling the tensor cores makes the results invalid due to the drop in accuracy. I.E sacrificing accuracy for speed isn't an option for the banking sector.
 
bru said:
So in the 'Tensorflow Resnet 50 benchmark' AMD didn't allow the competition to use their Tensor cores just because AMD don't have any and it would show their new 7nm flagship card to be rubbish in that particular benchmark.

Well that certainly seems a fair and transparent thing to do.;)
Click to expand...

ubersonic said:
According to the article it was because enabling the tensor cores makes the results invalid due to the drop in accuracy. I.E sacrificing accuracy for speed isn't an option for the banking sector.
Click to expand...

brings popcorn
 
bru said:
So in the 'Tensorflow Resnet 50 benchmark' AMD didn't allow the competition to use their Tensor cores just because AMD don't have any and it would show their new 7nm flagship card to be rubbish in that particular benchmark.

Well that certainly seems a fair and transparent thing to do.;)
Click to expand...

Actually is known for years now with Volta, when Tensor cores are activated the results accuracy goes down the drain.
It was especially common and potential dangerous issue, found on some protein research when the results were varying every time the calculation was run.
Same applies to Banking sector, AI etc.
On the other hand Tensor cores shine on image recognition processing.

As for Volta (TV100) performance, even on Tensor Flow matrix calculations (were Tensor cores should show their teeth), is only just 10% faster over the Vega FE and <20% over the humble Vega 64. (with ROCm 1.3, now we are on 1.9.1 and 2.0 coming out end of year). And last time we discussed about this provided you the benchmarks. :)
 
Btw apparently Radeon Group VP said that there won't be any DXR support until the tech is supported by whole product range.
From low end to high end.

Imho he is right to some extent. Given that we need a RTX2080Ti for 1080p @ 60fps using DXR......

Personally I am bit annoyed because Vega had DX12.1 compatible hardware, which is needed for DXR.
 
Wrinkly said:
.


As I understand it, tensor cores are not used due to poor accuracy.
Click to expand...

Which is utterly BS excuse. Teh reason AMD didn;t is because their fancy Vega 20 would get annihilated in their own benchmarks.

Remember, with the original Vega AMD were hyping up how fast their fp16 compute was. Now all of a sudden it is not accurate enough, despite the thousands of deep learning researching saying otherwise. In fact, the whole industry is moving to int8
 
Panos said:
Actually is known for years now with Volta, when Tensor cores are activated the results accuracy goes down the drain.
It was especially common and potential dangerous issue, found on some protein research when the results were varying every time the calculation was run.
Same applies to Banking sector, AI etc.
On the other hand Tensor cores shine on image recognition processing.

As for Volta (TV100) performance, even on Tensor Flow matrix calculations (were Tensor cores should show their teeth), is only just 10% faster over the Vega FE and <20% over the humble Vega 64. (with ROCm 1.3, now we are on 1.9.1 and 2.0 coming out end of year). And last time we discussed about this provided you the benchmarks. :)
Click to expand...


Apart form all the other nonsense in this post, I have highlighted the critical part. Maybe you should go back and have another think before posting. Hint, Image ResNet is a benchmark of image recognition processing, where by your own words Tensor cores shine, so why would AMD not test fairly with Tensor cores enabled on an image processing task?
 
You must log in or register to reply here.
Back
Top Bottom