Vega/Navi Rumors (Updated)

[crisium]

Polaris 12 vs GTX 1050 doesn't make sense, unless Polaris 12 is nothing more than a refined Polaris 11. Fully enabled P11 (1024 cores) comes reasonably reasonably close to GTX 1050 anyway.

I don't know why AMD would spend resources designing a brand-new chip just to sell it below $120 in this day and age. Not unless Apple wants it.

On another note, I do not like AMD's re-branding to the same number tier again if 480 -> 580 with overclock is true. It is disingenuous, and I feel it hurts the brand image. I prefer the GTX 680 -> 770 style.

I forgave AMD last time because they shifted the naming scheme to add 2 cards above 390X. Surely we will have at least 2 cards above the 580. Many expect up to 4 cards. 590, Vega, Vega X? Will we keep this up to where Polaris 10 is RX 680, but don't worry we will have 6 cards above?

Another thing not to like about that rumour: RX 580 still to use 8 Gbps (GTs, GHz, whatever) GDDR5. New name, same memory speed. Why not match your competitor who is moving to 9 Gbps?

 

[Glo.]

Its because of Vega architecture, which will be separate line of GPUs, containing two dGPUs and APU GPU.

 

[w3rd]

Nope, they can. HBM will act just like "normal memory" not like just a cache for Data. It is important to understand the difference here, because it changes approach, but does not make one better or worse form the other.

Not even sure what you are saying, or if you understand. Nvidia has not shown, nor possesses a nonvolatile, cohesive memory controller, or cache.

 

[Krteq]

Well, we still haven't seen any technical details about IMC and HBC Controller. How can you know how exactly it's working?

 

[w3rd]

Watch the video with AMDs chip engineer. Where he explains AMDs technology and direction..!

 

[Glo.]

Not even sure what you are saying, or if you understand. Nvidia has not shown, nor possesses a nonvolatile, cohesive memory controller, or cache.

It appears that its not me who does not understand .

What you do believe is that you need some magical hardware on silicon to exploit all of capabilities of the HBM. No. Simply no. HBCC allows to increase the throughput of the data, and management of how it is delivered to particular CU when its needed. Its like "streaming" the data to CUs, rather than storing it in the memory. HBM2 in GP100 acts like a storage. Data in Vega GPU will be "streamed". Different approaches.

But it does not mean that one is better than the other. For Vega, definitely HBCC will be better option. For GP100 better was Storage, because the GPU is targeted at HPC market. Which is completely different and has different requirements.

 

[tential]

Watch the video with AMDs chip engineer. Where he explains AMDs technology and direction..!

Why do we have a like system where we can't down vote? Utterly useless system.

If you're going to ask someone to watch a video, then link it.

 

[Krteq]

Watch the video with AMDs chip engineer. Where he explains AMDs technology and direction..!

I saw it a few times, but he explained a basis only. We don't know how it really works.

HBC and HBCC looks really promising, let's wait for more technical details... meanwhile be careful with such statements about it

 

[w3rd]

It appears that its not me who does not understand .

What you do believe is that you need some magical hardware on silicon to exploit all of capabilities of the HBM. No. Simply no. HBCC allows to increase the throughput of the data, and management of how it is delivered to particular CU when its needed. Its like "streaming" the data to CUs, rather than storing it in the memory. HBM2 in GP100 acts like a storage. Data in Vega GPU will be "streamed". Different approaches.

But it does not mean that one is better than the other. For Vega, definitely HBCC will be better option. For GP100 better was Storage, because the GPU is targeted at HPC market. Which is completely different and has different requirements.

I understand.

You are saying one is not better than the other, because you are truly missing the subtle hints. Which is the whole industry will inevitably be moving to speedless caches, where they can be design/configured to handle ANY bandwidth. Because DATA never essentially moves.

BTW, something like AMD's non-volatile memory controller cache, is the future and a pre-cursor before the industry starts moving into memristor tech. AMD is ahead of Intel and Nvidia on this front. Look at what technologies AMD has gobbled up in the last 10 years & are investing into right now (ie: HSA)

Cheng mentions a few key hints as to what AMD has in mind. Don't discount the things he mentioned. The technology comes from their APU pursuits.

 

[richaron]

AFAIK they call it a "non-volatile memory controller cache" because it directly refers to the greatest advantage of the "HBCC"; that is it's ability to directly reference the computers virtual memory subsystem (RAM, SWAP/Page, and all SSD's/HDDs). Being able to directly access SSD's/HDD's (non-volatile memory) opens up huge possibilities in the HPC and Scientific computing areas. This is on top of the GPU having direct access to system RAM.

For example the conventional thinking in the Molecular Dynamics community, even if they've been converted to GPU acceleration, is to max out system RAM. And this RAM limit directly affects the scope of the calculations possible (such as number of atoms). But conventionally the GPU can't even directly access system RAM so there's redundant data copies and added software complexity. A GPU directly accessing RAM would obviously be more efficient, but with barely any more complexity the GPU could also directly access something like an SSD array or massive HDD array. Obviously this would not be as fast as RAM, but it opens up the possibilities when considering the scope of the data sets available.

 

[Valantar]

Polaris 12 vs GTX 1050 doesn't make sense, unless Polaris 12 is nothing more than a refined Polaris 11. Fully enabled P11 (1024 cores) comes reasonably reasonably close to GTX 1050 anyway.

Overclocked P11 comes reasonably close to the GTX 1050. Fully enabled P11 beats it pretty much across the board.

I don't know why AMD would spend resources designing a brand-new chip just to sell it below $120 in this day and age. Not unless Apple wants it.

Probably for the mobile market. Apple alone is a huge design win, but so far I haven't seen Polaris based dGPUs in a single other laptop (the R7 4xxm and R9 4xxm are rebrands of older chips, after all). I for one would love to see 900p ultra/1080p medium-to-high-capable dGPUs start showing up in laptops, especially thin-and-light ones.

On another note, I do not like AMD's re-branding to the same number tier again if 480 -> 580 with overclock is true. It is disingenuous, and I feel it hurts the brand image. I prefer the GTX 680 -> 770 style.

I forgave AMD last time because they shifted the naming scheme to add 2 cards above 390X. Surely we will have at least 2 cards above the 580. Many expect up to 4 cards. 590, Vega, Vega X? Will we keep this up to where Polaris 10 is RX 680, but don't worry we will have 6 cards above?

Completely agree here. Even with process improvements leading to 10%+ increased clocks, rebranding on the same naming level implies performance stagnation, and hurts the Radeon brand. While the R9 390(X) -> Fury (X) naming scheme makes some sense, AMD is muddling their naming scheme something fierce. One thing is their "flagship" mid-range product having the same ending as the competition's high-end offering (not that AMD can affect Nvidia's branding, but I've seen surprising amounts of "OMG AMD sucks the 1080 is twice as powerful as the 480!" posts around the web), another entirely is seemingly abandoning their high-end naming (where are the RX 4/590 (X) cards?) for what would almost need to be more than two cards in a seemingly separate Radeon Vega series. After all, they need to compete at the very least with the 1070 and 1080, but I'd say they need something noticeably faster than the 1080 too. Also, AMDs current largest GPU die is slightly more than 200mm2. That's quite small. There's easily room for two larger dies in there, although I hope they don't have to go 500mm2+ this early in the lifetime of the 14nm process. That would indicate some serious performance deficits in their architecture.

Any rebrand of the 480, even with higher clocks and a price drop, should be called the 570. This would leave two tiers for "mid-range" Vega cards (with possible X models down the line after yield/process improvements), and the RX Vega (or Fury? Vega Fury?) tier for high-end cards.

Not even sure what you are saying, or if you understand. Nvidia has not shown, nor possesses a nonvolatile, cohesive memory controller, or cache.

Do we have any indication that consumer Radeon GPUs will use any kind of non-volatile cache? Do we know of any scenario where consumer usage would gain from this? And how would they implement this without significantly increasing costs?

We know this has uses in big data processing, finance and the like. But gaming? Not likely.

 

[Glo.]

I understand.

You are saying one is not better than the other, because you are truly missing the subtle hints. Which is the whole industry will inevitably be moving to speedless caches, where they can be design/configured to handle ANY bandwidth. Because DATA never essentially moves.

BTW, something like AMD's non-volatile memory controller cache, is the future and a pre-cursor before the industry starts moving into memristor tech. AMD is ahead of Intel and Nvidia on this front. Look at what technologies AMD has gobbled up in the last 10 years & are investing into right now (ie: HSA)

Cheng mentions a few key hints as to what AMD has in mind. Don't discount the things he mentioned. The technology comes from their APU pursuits.

I know all of this. The problem is that I know the reality of this industry you are talking about. It is the most resistant to change. It doesn't want it. The change you are talking about is 10 years from now.

 

[lobz]

Why do we have a like system where we can't down vote? Utterly useless system.

If you're going to ask someone to watch a video, then link it.

exactly. Wccftech is so much better, right?

 

[w3rd]

I know all of this. The problem is that I know the reality of this industry you are talking about. It is the most resistant to change. It doesn't want it. The change you are talking about is 10 years from now.

Correct^, & that era STARTS in 2 months when Vega is released...

 

[w3rd]

AFAIK they call it a "non-volatile memory controller cache" because it directly refers to the greatest advantage of the "HBCC"; that is it's ability to directly reference the computers virtual memory subsystem (RAM, SWAP/Page, and all SSD's/HDDs). Being able to directly access SSD's/HDD's (non-volatile memory) opens up huge possibilities in the HPC and Scientific computing areas. This is on top of the GPU having direct access to system RAM.

For example the conventional thinking in the Molecular Dynamics community, even if they've been converted to GPU acceleration, is to max out system RAM. And this RAM limit directly affects the scope of the calculations possible (such as number of atoms). But conventionally the GPU can't even directly access system RAM so there's redundant data copies and added software complexity. A GPU directly accessing RAM would obviously be more efficient, but with barely any more complexity the GPU could also directly access something like an SSD array or massive HDD array. Obviously this would not be as fast as RAM, but it opens up the possibilities when considering the scope of the data sets available.

Thank you.

+

 

[Bacon1]

Any rebrand of the 480, even with higher clocks and a price drop, should be called the 570. This would leave two tiers for "mid-range" Vega cards (with possible X models down the line after yield/process improvements), and the RX Vega (or Fury? Vega Fury?) tier for high-end cards.

Yeah it's odd because they named the Ryzen chips to directly compete with Intel's while offering more cores/threads.... R7 vs i7, R5 vs i5, R3 vs i3. Yet they name 480 vs 1060...

 

[Valantar]

Yeah it's odd because they named the Ryzen chips to directly compete with Intel's while offering more cores/threads.... R7 vs i7, R5 vs i5, R3 vs i3. Yet they name 480 vs 1060...

AMD has had a 1-tier performance difference with Nvidia in terms of graphics card names for quite some time - the 290(x) competed with the 780(Ti), and so on. I believe AMD named the 480 as they did because they knew they weren't coming out with a higher-end model for a while, and having a X70 card as your top-end model does look a lot worse. When that card is just competitive with the 1060, that's when they start getting in trouble - and they should really take the opportunity to correct this if the 5XX rebrand is happening. A 2-tier performance deficit just looks bad, regardless if there are higher-end models without numbers. Truth be told, this could be a sign of how unprepared AMD were for the performance jump of the GTX 10XX-series (with the 1060 nipping at the heels of the 980 and the 1070 really close to the 980 Ti with a little OC...).

 

[Bacon1]

AMD has had a 1-tier performance difference with Nvidia in terms of graphics card names for quite some time - the 290(x) competed with the 780(Ti)

Yeah their naming has just been off though...

290 vs 780, 290x vs 780 ti

but then 390 vs 970, 390x vs 980

And they knew Polaris was $200 price point for a long time, but still called it 480 against what they knew would be a x60 series since 970 was $300+ at launch.

But yeah I get not having any high end... but just odd naming for the cards and makes them look so much weaker in comparison.

 

[Valantar]

but still called it 480 against what they knew would be a x60 series since 970 was $300+ at launch.

I don't think they knew. Not many people really expected the 1060 to encroach on the 980's territory. 970 or a little more was reasonable, but the 9XX-10XX jump was big.

 

[Bacon1]

I don't think they knew. Not many people really expected the 1060 to encroach on the 980's territory. 970 or a little more was reasonable, but the 9XX-10XX jump was big.

Not really though, 480 compares to 390x and 1060 compares to 980 which compares to 390x.

I mean 970 ~= 780 Ti, 1070 ~= 980 Ti.

The only oddity was AMD's naming, could have just been like previous 470 and 470x but they changed it to RX so RX 470x would sound odd.

The 1060 naming was also odd, as the 1060 6gb should have been a 1060 Ti and 1060 3gb been the sole 1060 "non-ti / regular".

That would have fit their usual roles but who knows these companies are nuts ... Now to see what they end up calling Vega stuff :O

 

[tential]

Yeah it's odd because they named the Ryzen chips to directly compete with Intel's while offering more cores/threads.... R7 vs i7, R5 vs i5, R3 vs i3. Yet they name 480 vs 1060...

R7 was dumb. It is clearly meant to go against the i7s from the hedt platform and not the 7700k

The whole naming scheme was stupid. R5 should have been named to go against consumer i7.

Trying to name to directly mirror the competition is stupid.

They should create compelling products and explain them better. The fact that people compare R7 to the 7700k shows amd can't do this.

With gpus luckily got amd its easier. I

I prefer the X naming system from amd only because its easy to pick out the full and cut down chips and to just get the cut down one easily.

With the current naming scheme it requires a little more research.

 

[Bacon1]

R5 should have been named to go against consumer i7.

The fact that people compare R7 to the 7700k shows amd can't do this.

So you wanted people to compare the R5 series to compete against the 7700k instead of the 1700?? That would be a blowout though. I think naming them R4, R6 and R8 would have made more sense though since thats their core counts...

Yeah I'm hoping we end up with Vega YYY and Vega YYYx again or some other qualifier like Nvidia has with YYY Ti

 

[tential]

So you wanted people to compare the R5 series to compete against the 7700k instead of the 1700?? That would be a blowout though. I think naming them R4, R6 and R8 would have made more sense though since thats their core counts...

Yeah I'm hoping we end up with Vega YYY and Vega YYYx again or some other qualifier like Nvidia has with YYY Ti

Blowout by who? I think Ryzen 5 is a far better deal for gaming than Ryzen 7 or the 7700k.

If you fix smt with Windows, it's not even a fair fight. The 7700k doesn't stand a chance.

Ryzen 5 is the king of gaming for price / perf.

Funny I actually think your naming scheme is the ideal one and the one I was going to post before.

That's why I'm stoked for Vega. I think amd will be aggressive. They look like they want to gain market share as much as possible to start creating a real relationship with consumers.

Ryzen 6 being cheaper than a 7700k, named to be "weaker" than that, but stronger than an i5, but still competing with the 7700k... That would have been the intelligent move.

Too many people here are thinking games will take advantage of all 16 threads when the real move is to get the 12 threaded cpu if you care about value.

If Vega cut down chips come in at the typical cut down amd goodness, combined with Ryzen 6(ive wanted to call it this for awhile. I'm just going to start I don't care. The real name is stupid don't name your processors based on Intel. It's pathetic), then amd will have the best budget setups.

I think all users on here are way too obsessed with trying to use an Octocore for gaming when they can't provide any proof that the Octocore is a better value for gaming than the hexacore from amd.
. Someone even posted how well a straight up quad-core does from amd....

You don't lose much dropping cores, and THAT'S NOW. Wait till Windows gets some updates for smt just like it needed updates for hyper threading...

This story has only just begun. Vega 11+ryzen6 for the price conscious. Hedt + 1080ti for those who aren't.

 

[Bacon1]

Ryzen 5 is the king of gaming for price / perf.

Oh I agree, I just meant as far as raw framerates was concerned.

Ryzen 6 being cheaper than a 7700k, named to be "weaker" than that, but stronger than an i5, but still competing with the 7700k... That would have been the intelligent move.

Yeah I was confused by the R5/R7 naming at first until I realized they were trying to make it like i5/i7... Even more confusing is 1700 vs 1800(x) when they are just clock speeds but same chip :'(

If Vega cut down chips come in at the typical cut down amd goodness, combined with Ryzen 6(ive wanted to call it this for awhile. I'm just going to start I don't care. The real name is stupid don't name your processors based on Intel. It's pathetic), then amd will have the best budget setups.

Yeah I'm interested in seeing what the 2nd tier Vega will be, as often with launch you can do partial or even full unlock into the full tier, but at hundred(s) less.

 

[Glo.]

I am actually staggered that nobody spotted a rumor about Vega GPUs, that appeared on another forum, few days ago...