Poll: ** The AMD VEGA Thread **

[LoadsaMoney]

As said earlier in the thread, AMDs cards are now 'All In Ones', hence the HBM use.

 

[Rroff]

Below is an old GPUZ screenshot of one of my Titan Xp cards with a small overclock, checkout the memory bandwidth using cheap GDDR5X.

And this is still on 22-20nm for the memory - can't wait to see what is possible on 1x nm nodes.

 

[nashathedog]

They are limited by money, they couldn't afford to do multiple chips for every use.

That should change going forward, AMD's current money troubles can be traced back to Bulldozer, Now that's been replaced with Ryzen AMD's money issues will improve so thing's should get better for RTG too.

This is the really sad bit.

They tried to be innovative using HBM whereas for gaming GDDR5X would have been better and cheaper. Unfortunately this would not have been so good for their pro cards for the reasons N19h7m4r3 stated above.

If it's not too late in Navi's design progress there's no reason why they couldn't go with GDDR6 if money allows and HBM sourcing proves unreliable.

They don't do gaming cards anymore, as said earlier in the thread, AMDs cards are now 'All In Ones'.

Considering that the Polaris cards hold their own against the Pascal equivalents and they're still current they do do gaming cards but yeah Vega's gaming ability took second place to the pro market's requirements.

 

[Rroff]

If it's not too late in Navi's design progress there's no reason why they couldn't go with GDDR6 if money allows and HBM sourcing proves unreliable.

Future trajectory (and probably something they'd want to be introducing at least with Navi) is towards HBM type technologies - an MCM type implementation would require small, highly power/thermal efficient memory modules situated very close to the GPU modules. Current architectures though absolutely don't need or take significant enough advantage of it even if designed around it to justify using over GDDR technologies.

 

[Mauller]

You are still trying to tell me that HBM needs DX12 on AMD cards to get the best out of them, there are plenty of NVidia cards that are faster that don't use DX12 or HBM to get the performance.

You and DM really need to admit that you got it wrong with HBM and that there are better memory solutions for gaming cards.

Below is an old GPUZ screenshot of one of my Titan Xp cards with a small overclock, checkout the memory bandwidth using cheap GDDR5X.

With GDDR6 coming soon AMD really need to drop HBM for gaming cards.

You are the one saying HBM is the problem and you are like a broken record about it. I am saying its software and architecture of the GPU die itself, since the FPS at low resolution would still have issues in DX12/Vulkan/Mantle if it was HBM causing the problem.

Which is where driver improvements helped in DX11 over time with 1080p performance etc, but it was shown far earlier on that HBM was not the problem.

 

[LoadsaMoney]

That should change going forward, AMD's current money troubles can be traced back to Bulldozer, Now that's been replaced with Ryzen AMD's money issues will improve so thing's should get better for RTG too.



If it's not too late in Navi's design progress there's no reason why they couldn't go with GDDR6 if money allows and HBM sourcing proves unreliable.



Considering that the Polaris cards hold their own against the Pascal equivalents and they're still current they do do gaming cards but yeah Vega's gaming ability took second place to the pro market's requirements.

Polaris cards are for a different market, they're for the VR one, as they said, thats the market they want to grow with them, everyone buying a Polaris card, should be buying a headset to go with them, or have a one already, otherwise Polaris has been a failure for them.

 

[TNA]

Don't think AMD will drop HBM anytime soon on their high end cards. I also do not think HBM is the problem either. The only problem I can see is the manufacturers of HBM are lagging behind and not delivering things to schedule and AMD are suffering for this reason.

It seems AMD were probably expecting HBM2 that would deliver 512GB/s bandwidth at a lower voltage than even what they are getting currently at the lower bandwidth they have.

 

[LoadsaMoney]

Don't think AMD will drop HBM anytime soon on their high end cards. I also do not think HBM is the problem either. The only problem I can see is the manufacturers of HBM are lagging behind and not delivering things to schedule and AMD are suffering for this reason.

AMD were probably expecting HBM2 that would deliver 512GB/s bandwidth at a lower voltage than even what they are getting currently at the lower bandwidth they have.

They won't, unless they go back to doing gaming cards as well, but i can't see that, as they've not been able to compete with Nvidia for eons now, they are just so far behind them now, then again, they've been so far behind Intel for years, and now they're back, so its possible i guess, in time.

 

[TNA]

They won't, unless they go back to doing gaming cards as well, but i can't see that, as they've not been able to compete with Nvidia for eons now, they are just so far behind them now, then again, they've been so far behind intel for years, and now they're back, so its possible i guess, in time.

Lol. This post gets top marks Loadsa

But yeah, It seems when designing they build a card for compute and then carve a gaming card out of that. They don't have the budget to do a separate one for gaming like Nvidia.

 

[LoadsaMoney]

Lol. This post gets top marks Loadsa

But yeah, It seems when designing they build a card for compute and then carve a gaming card out of that. They don't have the budget to do a separate one for gaming like Nvidia.

Hopefully their next one, Navi, is the one to start things turning around for them, as some are saying.

 

[TNA]

Hopefully their next one (which i forget the name of now), is the one to start things turning around for them, as some are saying.

Navi. Yep, I am hoping this too. It is the first one Raja has worked on from the start. Sounds like it will be like a multi core CPU but on a GPU.

I remember years ago people saying that could not be done as the GPU is already multi core. But my understanding is that is what Navi will be. Which is great if they can make it work properly.

 

[Rroff]

Hopefully their next one, Navi, is the one to start things turning around for them, as some are saying.

I'm not quite at the point of writing off Vega yet but I think many of the advances in the architecture just aren't going to benefit it until Navi is relevant anyhow by which point Vega will look about as relevant as Fermi does today.

 

[Kaapstad]

You are the one saying HBM is the problem and you are like a broken record about it. I am saying its software and architecture of the GPU die itself, since the FPS at low resolution would still have issues in DX12/Vulkan/Mantle if it was HBM causing the problem.

Which is where driver improvements helped in DX11 over time with 1080p performance etc, but it was shown far earlier on that HBM was not the problem.

Lets see how Vega performs at 1080p and 2160p it could make interesting reading.

 

[LoadsaMoney]

Navi. Yep, I am hoping this too. It is the first one Raja has worked on from the start. Sounds like it will be like a multi core CPU but on a GPU.

I remember years ago people saying that could not be done as the GPU is already multi core. But my understanding is that is what Navi will be. Which is great if they can make it work properly.

Well they've done it with their CPUs, so its looking good

I'm not quite at the point of writing off Vega yet but I think many of the advances in the architecture just aren't going to benefit it until Navi is relevant anyhow by which point Vega will look about as relevant as Fermi does today.

Vega' just way too late, which ever way you look at it.

 

[nashathedog]

Future trajectory (and probably something they'd want to be introducing at least with Navi) is towards HBM type technologies - an MCM type implementation would require small, highly power/thermal efficient memory modules situated very close to the GPU modules. Current architectures though absolutely don't need or take significant enough advantage of it even if designed around it to justify using over GDDR technologies.

Well that was a waste of a perfectly good explanation.

I'm kidding, I got it....

 

[TNA]

Vega' just way too late, which ever way you look at it.

+1

 

[Davedree]

I wouldn't say Vega too late, more Amd had polaris on a shelf waiting for a suitable process, thus they cancelled the 44cu polaris Rx 490 and brought vega forward.


But this vega isn't looking all that great. should have either developed a 6 geometry engine at 1200-1300 mhz, or at least made a 4608/5120 vega.

 

[Sargatanas2511]

I wouldn't say Vega too late, more Amd had polaris on a shelf waiting for a suitable process, thus they cancelled the 44cu polaris Rx 490 and brought vega forward.


But this vega isn't looking all that great. should have either developed a 6 geometry engine at 1200-1300 mhz, or at least made a 4608/5120 vega.

We can say this all day long, but ultimately I am pretty sure AMD know better than us about what they're doing with their own architecture. If they thought it was better to do it another way I am sure they would have at least heard it out.

 

[Gerard]

In reference to the dx12 posts a few pages back, tried bf1 tonight on my card and was a bit choppy, fps were in the 80's and 90's but it just didn't feel like it. Just lacked fluidity. At the start of the round that also hitched a few times which it never does in dx11, seems NV need more work on their dx12 implementation.

 

[Davedree]

We can say this all day long, but ultimately I am pretty sure AMD know better than us about what they're doing with their own architecture. If they thought it was better to do it another way I am sure they would have at least heard it out.

Don't know i was spot on with my concerns over fury before it came out.

But yeah agree i know it's all if's and but's right now but if the rx vega drivers are fully compatible with frontier vega, then an analysis of clock scaling and power consumption vs voltage plot will tell us if Amd have reached their intended clockspeeds, or if they are pushing too hard. Or if frontier is just a batch of prototype dies, and rx vega is an improved v2 of the silicon. Much like the very early high leakage p10's vs the later p10's and current p10's.

If there's no difference in silicon and the drivers really sort out the problems we've seen then brilliant. But as i said way back in november, Gp104 has done it's damage and with an outlook of Amd needing to really grow, then Rx Vega doesn't appear to be a great execution towards revenue.