AMD "Greenland" Vega10 Silicon Features 4096 Stream Processors?

[Kaapstad]

http://www.anandtech.com/show/8526/nvidia-geforce-gtx-980-review

Why did the 680 have a smaller memory bus than the 580, why did the 980 have a smaller memory bus than the 780ti yet both faster than the predecessor? Who says the new midrange card will be faster than Fury X?

Where is AMD abandoning HBM, it's going to be on their and Nvidia's high end cards. It's expensive memory still, midrange cards have less need to push for 4k performance and midrange cards are cheaper and smaller meaning they have less space for a memory controller than a card over double the size, they have to fit into a cheaper price bracket. On top of all that, every generation brings efficiency improvements. Just because X shaders need Y bandwidth in one generation doesn't mean they do in the next generation.

Historically almost every single midrange card has had similar/higher performance to the previous gen midrange card yet less bandwidth and a cheaper narrower memory bus.

But something that has been completely standard for the past 15 years of new generations of graphics cards for AMD and Nvidia you decided to phrase as AMD abandoning HBM... sure, you have no agenda, no bias.

You are having a laugh.

This physical size of the chip is not important, it is the performance which determines whether a particular memory setup is needed or not.

The new AMD cards should at least be faster than the Fury X but by how much no one can say.

AMD have realised there are benefits to not using HBM for this card at this performance level so why can't you.

 

[AlamoX]

AMD is banking on VR, and they said they will bring 390 performance to a lower price tag than 350$, so yea you can expect them to cut cost on a lot of stuff gddr instead of hbm, memory bus and pcb cost, personally i think polaris 10(R9 480) would be around 390 perf and under 299$ price tag, polaris 11 (R9 470 or 460)

 

[TheRealDeal]

You are having a laugh.

This physical size of the chip is not important, it is the performance which determines whether a particular memory setup is needed or not.

The new AMD cards should at least be faster than the Fury X but by how much no one can say.

AMD have realised there are benefits to not using HBM for this card at this performance level so why can't you.

No AMD realise that 4gb is not what the paying public want and will most likely go for 8gb of cheaper gddr5 on there mid/high polaris cards. The only benefits i can see for Gddr5 is its most likely cheaper and can be used in higher amounts compared to HBM1 more easily.

 

[Kaapstad]

No AMD realise that 4gb is not what the paying public want and will most likely go for 8gb of cheaper gddr5 on there mid/high polaris cards. The only benefits i can see for Gddr5 is its most likely cheaper and can be used in higher amounts compared to HBM1 more easily.

So HBM1 has been a failure if AMD are back on GDDR5 ?

 

[Mauller]

So HBM1 has been a failure if AMD are back on GDDR5 ?

Production of HBM 1.0 was required to gear up a whole new manufacturing process that has many more stages to it.

It is just unfortunate that HBM 1.0 was limited to 1GB per stack. In no way was it a failure since it was a required step, one which Nvidia have not taken so will need to figure out all of the production issues themselves when it comes to using HBM 2.0.

 

[TheRealDeal]

So HBM1 has been a failure if AMD are back on GDDR5 ?

In what way has it been a failure. They have gained experience from it but people now want more than 4gb which is probably not viable for a mid end gpu if AMD price it as such. HBM2 will take over so no HBM is not a failure it's just not viable for what's coming. They are still sticking with HBM (HBM2) for the big guns and so are Nvidia. If it was the failure that you believe it to be then Nvidia would not be switching to it.

 

[AlamoX]

So HBM1 has been a failure if AMD are back on GDDR5 ?

i will take you up on this comment, once the first APU shows up

 

[Kaapstad]

Production of HBM 1.0 was required to gear up a whole new manufacturing process that has many more stages to it.

It is just unfortunate that HBM 1.0 was limited to 1GB per stack. In no way was it a failure since it was a required step, one which Nvidia have not taken so will need to figure out all of the production issues themselves when it comes to using HBM 2.0.

In what way has it been a failure. They have gained experience from it but people now want more than 4gb which is probably not viable for a mid end gpu if AMD price it as such. HBM2 will take over so no HBM is not a failure it's just not viable for what's coming. They are still sticking with HBM (HBM2) for the big guns and so are Nvidia. If it was the failure that you believe it to be then Nvidia would not be switching to it.

HBM1 has been a failure as AMD are no longer prepared to use it and have gone back to GDDR5 for a new card that is going to be faster than a Fury X.

HBM2 and beyond will be a different story with larger memory sizes and higher clockspeeds.

 

[Mauller]

HBM1 has been a failure as AMD are no longer prepared to use it and have gone back to GDDR5 for a new card that is going to be faster than a Fury X.

HBM2 and beyond will be a different story with larger memory sizes and higher clockspeeds.

For the price that they want to release the Polaris cards within it does not merit the use of HBM. And the limited capacity being another issue. Production of it was still required to gain experience with interposer technology and packaging, so it was not a failure but a stepping stone.

Although it is being left behind rather quick, it does not mean it was a failure. But of course your own view on HBM is already flawed.

 

[r7slayer]

Nothing points to them being next year, in the slide Vega was shown hovering over 2017 in the timeline, so they could come late this year or early next year.



Might be worth waiting to see how the 8 core zen comes up, would be a better upgrade if the 8 core has the same price as a top end, but not extreme edition, i7.

So on the road map Polaris being mid 2016 and vega hovering over 2017 means its likley to come out end of 2016? Thats not the impression I get nor logically makes sense to me. When has Nvidia or Amd released thier full range of cards allmost in one go? 2017 vega will be thats for sure.

 

[Mauller]

So on the road map Polaris being mid 2016 and vega hovering over 2017 means its likley to come out end of 2016? Thats not the impression I get nor logically makes sense to me. When has Nvidia or Amd released thier full range of cards allmost in one go? 2017 vega will be thats for sure.

Same as i said before, but them stating Polaris out mid year for back to school adds weight to the possibility of Vega out end of year or early next. That and AMD have had more time to play with 14nm so it is all within the realm of possibility that they can have Vega out sooner than later.

Also have to couple together that AMD have had far more experience with HBM parts production so will have a HBM 2.0 part out sooner than Nvidia.

 

[Freddie1980]

I'm a bit surprised we haven't had a 14nm flusher card from AMD. Sometimes when there is a new process and a you have a wide range of existing products you take one of your mid rage cards and try it out on the new process to learn about the new process, it looks like there going straight in with a new design and new process.

 

[eddyr]

Kaap, If the use of gddr5 signals failure of hbm1. Had hbm2 launched last year on Fiji, or even become available shortly after Fiji and yet Polaris this year still featured gddr5, would would hbm2 also be considered a failure?

 

[h4rm0ny]

i will take you up on this comment, once the first APU shows up

By then something else will be proof that AMD have failed.

 

[drunkenmaster]

You are having a laugh.

This physical size of the chip is not important, it is the performance which determines whether a particular memory setup is needed or not.

The new AMD cards should at least be faster than the Fury X but by how much no one can say.

AMD have realised there are benefits to not using HBM for this card at this performance level so why can't you.

Again, a 980 is faster than a 780ti... but uses a smaller memory bus... explain why this is... because the examples I gave you were precisely based on PERFORMANCE as well as size. Once again you utterly fail to understand a basic comparison that makes your post pointless.


From one generation to the next you gain improvement in efficiency in every part of the gpu, including memory bandwidth usage, as Kepler had over Fermi, as Maxwell had over Kepler, as Pascal will have over Maxwell.... as Polaris will have over Fiji.

Just because Fiji required X amount of bandwidth for a given performance level doesn't mean Polaris requires the same amount of bandwidth for the same performance.

Again, basically every Nvidia and AMD new gen midrange has used a smaller bus yet usually offered more performance than the previous gen high end... explain in your holy wisdom of "I've taken the heatsink off a card and you haven't" why this is.

 

[Mauller]

Also the use of instruction prefetching and further improved memory compression should mean a smaller bus has less of an impact on GCN 4.0 than it would on earlier GCN. Also the Larger L2 cache helps this.

 

[Boomstick777]

Nothing points to them being next year, in the slide Vega was shown hovering over 2017 in the timeline, so they could come late this year or early next year.



Might be worth waiting to see how the 8 core zen comes up, would be a better upgrade if the 8 core has the same price as a top end, but not extreme edition, i7.

Yeah might be but would mean new mobo etc as well, so will likely just grab the Kaby Lake and then move onto whatever comes in the later Zen generations as time goes on. Less hassle that way.

 

[spikerules]

Kaap you seem to be on a HBM hate crusade, which isn't warranted or required. Every other post I have read catching up here has had you either being mentioned for hating HBM, or you yourself just linking to some article which apprently states HBM was a failure. You then try to shovel your opinion down everyone's throats! It's tiresome.

I have a Nano and LOVE IT. The form factor alone warrants HBM for me. It's a beast of a card and in many ways I prefer it to my 980Ti. Is it as powerful? No! But performance to size ratio it IS the more powerful card.

 

[r7slayer]

Kaap is quite focused on memory recently. He was on a bit of a crusade telling everyone how newer games have x usage like it was warranting the need of a titan x to play newer games on max or near max settings lol. I think last time i heard ROTTR was breaking 6Gb on 1080p lol. Here is me palying this game on 1440p on near max settings(just ambient occlusion not max) and its sitting nicley at around 5Gb. This dosn't mean you need atleast 5Gb to play at these settings at 1440p it's just what my usage was at. Could possibly work fine on cards with 4Gb of memory.

 

[Ian Evey]

1 card 4gb is the norm and fine below 4k in recent games, with 2 cards 4gb is not enough regardless of hbm or gddr unless your happy paying double the cost with little benefit to IQ.

I know I would not be happy if I could not play a game at ultra settings with 2 cards because it had 4gb without any difference in IQ to having 1 card.

That's where 2 8gb 390's shine.