Vega refresh in 2018, current Vega is broken!

AmateurExpert · 20 Dec 2017 at 16:14

4K8KW10 said:
Is Samsung the single manufacturer of HBM?

The only other major manufacturer working on HBM is Hynix and they were many months behind Samsung in HBM2 production; it's possible that Hynix might beat them to volume production, but it's unlikely.

4K8KW10 said:
That article is extremely old and the slide there is extremely not serious. Just look at what kind of meaningless words there were used.

Of course all slideware information is subject to change, but there are also extremely few instances of non-trivial products being made available years ahead of planned availability.

nashathedog · 20 Dec 2017 at 16:32

Kirby Wurm said:
My biggest question for Navi is HBM2, GDDR6 or both?

Kaapstad said:
Or play it safe and use GDDR5X, it is fast enough for a Pascal Titan so it should be fine for Navi.

If they'd gone with 5x or even the the GDDR5 found on Hawaii/Grenada I doubt Vega would be any worse than it is, Plus it'd be a lot cheaper and it would have released a lot earlier so I hope they do move away from HBM for Navi. One question, Is Navi a Polaris or Vega replacement?

The best thing that AMD could do is build the fastest GDDR5, 5x or 6 graphics card then can and then purposely nerf it so it can't mine, Then we'll get a reversal of fortunes with miners buying up all the Nvidia cards and AMD getting back into the gaming side of things, Then they can release a really fast mining version so every miner wants that over Nvidia's cards and then the market gets flooded with unwanted Nvidia cards making Nvidia's stock value slump.

Aaah we can all dream.

nashathedog · 20 Dec 2017 at 16:40

4K8KW10 said:
One Scalable Navi GPU To Rule Them All – The Legend Of The Wonder Glue https://wccftech.com/amd-navi-gpu-spotted-in-linux-drivers/

I've read that bit about Vega supporting Infinity fabric before but how exactly does it support it? As far as I know it's not used in Vega so why does it matter if it can use it if it doesn't?

LoadsaMoney · 20 Dec 2017 at 17:24

Just goes to show you how far ahead Nvidia are, if AMD are going to have glue about 5/6 GPUs together, just to get near to a single Nvidia one.

jigger · 20 Dec 2017 at 18:00

LoadsaMoney said:
Just goes to show you how far ahead Nvidia are, if AMD are going to have glue about 5/6 GPUs together, just to get near to a single Nvidia one.

Seems about 10-15% between Vega and Pascal in DX11. The problem cost not performance.

AllBodies · 20 Dec 2017 at 19:12

nashathedog said:
I've seen rumours regarding a 2018 Q4 or 2019 Q1 PS5, I haven't a clue if it's true(ish) but if it is I just hope they stay all AMD for that. Hopefully they will as AMD's the only one able to offer a complete APU package, I get the feeling we need both Microsoft and Sony to do both, I don't think the Switch matters as much so I hope Nvidia are content with that.

I imagine they're going to wait for late 2019-2020.

Both companies have barely brought out their 'Pro' consoles, and late 2019- would allow them to use the refined 7nm+EUV process, plus use HBM3 (or the cheap-HBM variant), plus potentially even get Zen3 and Navi's successor, instead of a Zen2+Navi APU.

Also late 2019 would coincide with a lot of development in VR, as I imagine both companies will want to push that properly in the next generation.

Bergkamp · 20 Dec 2017 at 19:14

jigger said:
Seems about 10-15% between Vega and Pascal in DX11. The problem cost not performance.

If you need to spend hugely more for the same performance then there is clearly a performance issue.

nashathedog · 20 Dec 2017 at 19:15

AllBodies said:
I imagine they're going to wait for late 2019-2020.

Both companies have barely brought out their 'Pro' consoles, and late 2019- would allow them to use the refined 7nm+EUV process, plus use HBM3 (or the cheap-HBM variant), plus potentially even get Zen3 and Navi's successor, instead of a Zen2+Navi APU.

Also late 2019 would coincide with a lot of development in VR, as I imagine both companies will want to push that properly in the next generation.

That'd definitely make a decent console, I'm not too sure about VR still having much relevance by then though.

jigger · 20 Dec 2017 at 19:18

Bergkamp said:
If you need to spend hugely more for the same performance then there is clearly a performance issue.

It's a market issue. The performance is fine, the tech is better it's just the cost. I'm half tempted to throw mine on ebay for 1k a pop cash on collection. Or maybe even offer 4 hour delivery.

AllBodies · 20 Dec 2017 at 19:35

nashathedog said:
That'd definitely make a decent console, I'm not too sure about VR still having much relevance by then though.

Mmm, and even if they went for low-cost HBM for savings they could still do 16GB with 800 GB/s bandwidth (4 stacks).

And on 7nm+EUV you could fit 6 Zen cores plus 4500-5000 Polaris cores all into less than 300mm2.

So unless Zen2/3 and Navi increase die size per core substantially, or MS and Sony go for tiny APUs, the next consoles should be genuinely very powerful.

nashathedog · 20 Dec 2017 at 19:41

AllBodies said:
Mmm, and even if they went for low-cost HBM for savings they could still do 16GB with 800 GB/s bandwidth (4 stacks).

At least AMD will of perfected HBM by then so hopefully they'll be getting good yields so they can offer decent chips.

Gregster · 20 Dec 2017 at 22:15

jigger said:
It's a market issue. The performance is fine, the tech is better it's just the cost. I'm half tempted to throw mine on ebay for 1k a pop cash on collection. Or maybe even offer 4 hour delivery.

lmao

Kaapstad · 21 Dec 2017 at 09:19

nashathedog said:
If they'd gone with 5x or even the the GDDR5 found on Hawaii/Grenada I doubt Vega would be any worse than it is, Plus it'd be a lot cheaper and it would have released a lot earlier so I hope they do move away from HBM for Navi. One question, Is Navi a Polaris or Vega replacement?

The best thing that AMD could do is build the fastest GDDR5, 5x or 6 graphics card then can and then purposely nerf it so it can't mine, Then we'll get a reversal of fortunes with miners buying up all the Nvidia cards and AMD getting back into the gaming side of things, Then they can release a really fast mining version so every miner wants that over Nvidia's cards and then the market gets flooded with unwanted Nvidia cards making Nvidia's stock value slump.

Aaah we can all dream.

It is interesting using the Titan Volta as it has HBM2 and more bandwidth than any other gaming card around.

The interesting bit is despite having so much memory bandwidth you still get noticeable performance gains when you overclock the HBM2.

The point I am trying to make is despite having so much bandwidth available HBM2 is still holding the card back for gaming.

I do hope that all future gaming cards use GDDR5(X) or 6 as I really do think it is better for gaming.

Geck0 · 21 Dec 2017 at 09:31

AMD couldn't have utilised GDDR5 or X for Vega, it would have blown the power envelope out of the park.

Rroff · 21 Dec 2017 at 09:38

Kaapstad said:
It is interesting using the Titan Volta as it has HBM2 and more bandwidth than any other gaming card around.

The interesting bit is despite having so much memory bandwidth you still get noticeable performance gains when you overclock the HBM2.

The point I am trying to make is despite having so much bandwidth available HBM2 is still holding the card back for gaming.

I do hope that all future gaming cards use GDDR5(X) or 6 as I really do think it is better for gaming.

I'm not sure its too reliable as a guide to next gen gaming cards though - the Volta core is designed for compute and just the whole way it operates is more closely tied to memory capabilities for want of a better way to put it without getting complex. Gaming Volta if such existed would likely be a little less memory sensitive though we are moving towards where faster than GDDR5 will be required.

LoadsaMoney said:
Just goes to show you how far ahead Nvidia are, if AMD are going to have glue about 5/6 GPUs together, just to get near to a single Nvidia one.

Advanced in substrate technology mean that soon we will no longer be restricted to a single monolithic die for the kind of bandwidth and latency, etc. required internally to a GPU to function at decent performance levels which opens up a lot of possibilities for spreading the sub-systems of a GPU out to different dies and allowing mix and match configuration depending on target market, etc. whether that is compute or a lower end consumer card and so on.

ubersonic · 21 Dec 2017 at 10:15

Geck0 said:
AMD couldn't have utilised GDDR5 or X for Vega

Midrange Vega cards will reportedly be GDDR5.

nashathedog · 21 Dec 2017 at 12:42

Kaapstad said:
It is interesting using the Titan Volta as it has HBM2 and more bandwidth than any other gaming card around.

The interesting bit is despite having so much memory bandwidth you still get noticeable performance gains when you overclock the HBM2.

The point I am trying to make is despite having so much bandwidth available HBM2 is still holding the card back for gaming.

I do hope that all future gaming cards use GDDR5(X) or 6 as I really do think it is better for gaming.

Same here, It sounds like a no brainer to me, especially considering how AMD's HBM 2 seems to be inferior to the HBM 1 used with the Fury range in every way except quantity which makes it sound like HBM isn't really ready for gaming yet. Hopefully they'll get it right with Navi.

EDIT: Don't forget to put my Nitro+ on the Vega 64 ROH please mate.

Kaapstad · 21 Dec 2017 at 14:00

nashathedog said:
EDIT: Don't forget to put my Nitro+ on the Vega 64 ROH please mate.

Fixed.

TheRealDeal · 21 Dec 2017 at 14:02

nashathedog said:
Same here, It sounds like a no brainer to me, especially considering how AMD's HBM 2 seems to be inferior to the HBM 1 used with the Fury range in every way except quantity which makes it sound like HBM isn't really ready for gaming yet. Hopefully they'll get it right with Navi.

EDIT: Don't forget to put my Nitro+ on the Vega 64 ROH please mate.

Where are you getting that the hbm 2 on Vega is inferior to the hbm 1 on Fury. I think you are confused a little here. Fury X needed 4 stacks just to get 4gb and slightly more bandwidth. Vega only needed 2 stacks of Hbm2 to achieve this. Had Vega used 4 stacks like Fury X then it would have had nearly double the bandwidth of Fury X and 4 x the memory.

nashathedog · 21 Dec 2017 at 15:17

TheRealDeal said:
Where are you getting that the hbm 2 on Vega is inferior to the hbm 1 on Fury. I think you are confused a little here. Fury X needed 4 stacks just to get 4gb and slightly more bandwidth. Vega only needed 2 stacks of Hbm2 to achieve this. Had Vega used 4 stacks like Fury X then it would have had nearly double the bandwidth of Fury X and 4 x the memory.

Hi, I explained my opinion porely sorry, I meant what we actually got not what we would have had if they had given us more stacks, What matters is that they didn't give us 4 stacks and even the 16gb Frontier editions only has 2 stacks and no improvement on bandwidth over the 8gb RX? Sure we could have got better but we didn't so while quantity is up bandwidth is down compared to Fiji.

Kaapstad said:
Fixed.

Cheers.