Vega/Navi Rumors (Updated)

[Mopetar]

Well GF is using the 14 nm LPP developed by Samsung, but it's based of their LPE process so it's designed to be efficient at the expense of overall performance. Samsung has announced a newer 14 nm LPU process that's supposed to offer improved performance, but I have no idea if Global Foundries is getting it. At least AMD has the option of using other companies to fab their chips now after changes to their agreement with GF. Beyond that wait for HBM2, AMD may have delayed Vega until they could get better fab tech if the 14 nm at GF wasn't cutting it.

Thinking back on it, around last March the timeline for Vega was late 2016 / early 2017, depending on how you want to interpret the image:

Even Polaris itself may have been somewhat delayed as they had earlier announcements that had hinted at launching much sooner than they actually did. My guess is that what they were getting back from GF either made them delay Vega or scrap using the 14 nm LPP process at GF to make it because it wasn't good enough. HBM2 is ready by now as evidence by its use in GP100 products, so the only reason Vega won't be out until around mid-2017 has to be due to issues with the GF fab not being adequate.

 

[JDG1980]

If the Vega samples are running at only 1200 MHz and the design goal is >1500 MHz, that goes a long way towards explaining why the displayed performance on game demos (like that Doom 4K demo that everyone keeps citing) has been only about on par with GP104 and not GP102.

Ryzen saw a lot of clock speed boosts at a relatively late stage - at one time, people weren't expecting it to get much over 3.0 GHz. Vega was designed specifically with higher clock speeds as a goal, and AMD is going to need to push this as hard as they can to compete with Nvidia. Pascal tops out around 2000 MHz; I don't expect Vega to go that high, but the closer AMD can go, the better.

It's surprising to see the leaked benchmarks indicate only 8 GB of HBM2, when the leaked slides from a while back said 16 GB. Maybe 16 GB will be for the professional cards only? I don't see this being a major limitation in the real world, though.

 

[JDG1980]

It wasn't until railven brought up gf that I wondered why more people didn't worry about Ryzen and Polaris. Both would have been far better products if made elsewhere....

Ryzen is incredibly efficient in the applications AMD really cares about (servers and laptops). When clocked modestly - 3.0 GHz or less - it actually beats Intel's offerings in the same range in terms of perf/watt. A version on TSMC 16FF+ might have been better optimized for enthusiast clocks (4.0+ GHz), but the truth is that desktop gaming enthusiasts are a fairly small niche and it would have been economically foolish for AMD, with their limited resources, to focus on those needs when it might mean a worse product for bigger and more important markets.

Polaris is a different story. The primary reason it was a disappointment is that it was overhyped - not surprisingly, since it was basically AMD's only new product for 2016. Polaris is a slightly refined version of GCN; it would not be unfair to call it GCN 1.3. It was also a "pipe cleaner" product used to test the market readiness of the GloFo 14nm process and work out kinks. In other words, it was largely a placeholder so AMD would have something at least modestly competitive until Vega was ready. All other versions of GCN have their "sweet spot" in terms of clocks around 800-900 MHz and so did Polaris. 1266 MHz was pushing it, and AMD only did that so that the product would be within spitting distance of Hawaii in terms of performance. (They wanted to discontinue Hawaii since it was a massive die that didn't help them fill their WSA and required large, expensive boards.) I'm not at all convinced that fabbing Polaris on TSMC 16FF+ would have provided a huge advantage. It might have given some marginal benefits, but it was always going to require substantial effort on AMD's part to break the clock speed wall for GCN, and Polaris just did not incorporate those efforts.

 

[Dave2150]

I really don't understand why AMD has been messing around with HBM & HBM2 for the last few years. It appears to have hampered their ability to compete at all, due to the massively increased costs of developing and implementing such a radical new memory solution for consumer cards.

Nvidia is having massive success with GDDR5X - why on earth haven't AMD jumped on this bandwagon? Do NVIDIA have exclusive access to it?

Perhaps this is why Vega appears to have been delayed, maybe they are making a GDDR5X 11 Gbps version of the GPU instead? Though I guess it's far too late for this, unless Vega was designed to have both types of memory controller on it, which I doubt it possible/feasible.

Either way it looks likely that I'll be getting a AIB custom 1080ti when they are in stock (1-2 months I guess, due to insane demand) instead of Vega. I wish AMD could compete, though it seems their R&D and other budgets have all been neutered to launch Ryzen. Sad times.

 

[amenx]

Whenever there are big delays, you know things dont look good.

 

[richaron]

Perhaps this is why Vega appears to have been delayed, maybe they are making a GDDR5X 11 Gbps version of the GPU instead? Though I guess it's far too late for this, unless Vega was designed to have both types of memory controller on it, which I doubt it possible/feasible.

Either way it looks likely that I'll be getting a AIB custom 1080ti when they are in stock (1-2 months I guess, due to insane demand) instead of Vega. I wish AMD could compete, though it seems their R&D and other budgets have all been neutered to launch Ryzen. Sad times.

Apparently Vega supports both HBM2 and GDDR5. Not sure about the GDDR5X, but from what I read it would be a substantial redesign over non-X and this probably wouldn't be financially viable vs putting resources into HBM2 (and I assume GDDR6) controllers. Also I'm not sure if they meant each Vega chip supports both, or if the two Vegas have a different controller.

One could also say it's sad to assume performance without seeing evidence, no?

 

[Glo.]

The real reason for the delay was simply launching the GPUs in the same year as Ryzen GPUs, to reestablish AMD brand.

 

[JDG1980]

I really don't understand why AMD has been messing around with HBM & HBM2 for the last few years. It appears to have hampered their ability to compete at all, due to the massively increased costs of developing and implementing such a radical new memory solution for consumer cards.

AMD is interested in planning for the future, not just creating a cheap and dirty solution to beat Nvidia right this minute. HBM allows all kinds of potential designs that can't be done with GDDR5X. It's also a lower-power solution, which is vitally important in the professional and laptop markets. In the future, we will be seeing APUs with Zen+ cores and Vega GPUs sharing a couple gigabytes of HBM as a cache, providing all-around performance in a tiny, ultra-low-power package that nothing from Intel and Nvidia can match.

Yes, if you only care about gaming and are willing to replace your rig in 3 years or so, you're probably better off going with a 7700K CPU and 1080 Ti GPU. AMD has their sights set higher than that, and on much more lucrative markets.

 

[SpaceBeer]

I doubt Vega is delayed (if we can say that at all, since it was always H1) due to HBM2. AMD had released gaming card with HBM 20 months ago. So they know how to make it. Of course, there are some differences in architecture comparing to Fiji, but that was expected and planned. Regarding HBM volume - that might be the issue, but if we know this is 30% of market and AMD's share is 30% of that, well, I think SK Hynix shouldn't have any problems to make that qty.

 

[w3rd]

I really don't understand why AMD has been messing around with HBM & HBM2 for the last few years. It appears to have hampered their ability to compete at all, due to the massively increased costs of developing and implementing such a radical new memory solution for consumer cards.

Nvidia is having massive success with GDDR5X - why on earth haven't AMD jumped on this bandwagon? Do NVIDIA have exclusive access to it?

Perhaps this is why Vega appears to have been delayed, maybe they are making a GDDR5X 11 Gbps version of the GPU instead? Though I guess it's far too late for this, unless Vega was designed to have both types of memory controller on it, which I doubt it possible/feasible.

Either way it looks likely that I'll be getting a AIB custom 1080ti when they are in stock (1-2 months I guess, due to insane demand) instead of Vega. I wish AMD could compete, though it seems their R&D and other budgets have all been neutered to launch Ryzen. Sad times.

Correct.
AMD worked with and has exclusive rights to HBM2, while Nvidia sign a deal with Crucial? for their GDDR5x. GDDR5x is just a stop gap because Nvidia can't release a HBM2 part, until after AMD has theirs out in the market.

So AMD can take all the time it wants, it has an uber advantage.

 

[jpiniero]

AMD worked with and has exclusive rights to HBM2,.

Except for GP100 products of course?

 

[96Firebird]

AMD worked with and has exclusive rights to HBM2...

What? Nvidia has been using HBM2 with their GP100 cards for year now, 16GB per card.

You are completely wrong, there is no exclusivity rights for any of the current VRAM technologies.

 

[amenx]

So AMD can take all the time it wants, it has an uber advantage.

Meanwhile, Nvidia continues to rake in the cash from a over a year of no competition for above mid-range segments of their product line. Its not just lost sales for AMD, but lost or losing traditional customer base who want big performance NOW.

 

[tential]

I really don't understand why AMD has been messing around with HBM & HBM2 for the last few years. It appears to have hampered their ability to compete at all, due to the massively increased costs of developing and implementing such a radical new memory solution for consumer cards.

Nvidia is having massive success with GDDR5X - why on earth haven't AMD jumped on this bandwagon? Do NVIDIA have exclusive access to it?

Perhaps this is why Vega appears to have been delayed, maybe they are making a GDDR5X 11 Gbps version of the GPU instead? Though I guess it's far too late for this, unless Vega was designed to have both types of memory controller on it, which I doubt it possible/feasible.

Either way it looks likely that I'll be getting a AIB custom 1080ti when they are in stock (1-2 months I guess, due to insane demand) instead of Vega. I wish AMD could compete, though it seems their R&D and other budgets have all been neutered to launch Ryzen. Sad times.

It's part of the corporate philosophy of the company to be innovative. You can tell they prioritize this above the logical "efficient" move. So you can't expect AMD to continue to do the efficient move when they run themselves in that manner.

 

[richaron]

It's part of the corporate philosophy of the company to be innovative. You can tell they prioritize this above the logical "efficient" move. So you can't expect AMD to continue to do the efficient move when they run themselves in that manner.

Actually it's probably much more efficient to skip GDDR5X, since it's basically a half generation with about a quater generation lifetime at the top (or less). It will be the fastest GDDR for less than 2 years before GDDR6 comes out. Compared to GDDR5 being on the market since 2008.

AMD had already committed to HBM2 for the top end and there are already other GDDR options coming out before the end of this year, so why exactly is it less efficient to skip GDDR5X? Sure they'll be loosing some revenue from performance, and they'll gain some from cheaper products and less work designing the X memory controller.

This seems like the definition of efficiency: Less work and less money output to achieve (at least) similar financial position...

 

[Mopetar]

Apparently Vega supports both HBM2 and GDDR5. Not sure about the GDDR5X, but from what I read it would be a substantial redesign over non-X and this probably wouldn't be financially viable vs putting resources into HBM2 (and I assume GDDR6) controllers. Also I'm not sure if they meant each Vega chip supports both, or if the two Vegas have a different controller.

I doubt that it's directly compatible with both, but it's hard to say with the HBCC which seems to imply that it can work with any memory in some way. If anything it's more likely that the two Vega chips use different memory instead of having support for both types baked in which seems like a waste of die space in general unless they figured out a slick implementation that doesn't involve having a lot of redundant silicon.

Correct.
AMD worked with and has exclusive rights to HBM2, while Nvidia sign a deal with Crucial? for their GDDR5x. GDDR5x is just a stop gap because Nvidia can't release a HBM2 part, until after AMD has theirs out in the market.

I don't know what you're talking about here. The Tesla (P100) has HBM2 memory and has been available for a while now. AMD doesn't have exclusive rights to it (but they do own patents on some of the technology if that's perhaps what you're thinking about) so it's pretty clear that Vega delays have nothing to do with HBM2 memory availability.

 

[richaron]

I don't know what you're talking about here. The Tesla (P100) has HBM2 memory and has been available for a while now. AMD doesn't have exclusive rights to it (but they do own patents on some of the technology if that's perhaps what you're thinking about) so it's pretty clear that Vega delays have nothing to do with HBM2 memory availability.

Some time ago there were reports of AMD having "priority" with HBM2 since they worked with SK Hynix for quite a while during development. Unsubstantiated but this seems entirely possibly to me.

 

[Sweepr]

Two stack HBM2 with only a minor bandwidth advantage compared to cut down GP102 (8GB 512GB/s vs 11GB 484GB/s) isn't really impressive - Fury X had 50% more bandwidth than GTX 980 Ti. Meanwhile rumor says the competition is embracing 16GB GDDR6 at >14Gbps for their next generation.

 

[w3rd]

Some time ago there were reports of AMD having "priority" with HBM2 since they worked with SK Hynix for quite a while during development. Unsubstantiated but this seems entirely possibly to me.

Correct. Why else would Nvidia sign on for 1 year of GDDR5x for consumer cards, if they already have HBM2 cards..?

Because AMD has release rights for HMB2 on desktop.

 

[SunnyNW]

The HBM2 that Nvidia has been sourcing for the P100 is from Samsung not SK Hynix. So AMD might indeed have timed exclusivity of HBM2 but only from SK Hynix, and they might not even have that.

Edit: To answer w3rd above ^^ - Nvidia is using GDDR5x instead of HBM2 on their consumer cards for the time being because the transition is Much easier from GDDR5 to 5x, AMD might use HBM memory for power savings but Nvidia doesn't have as much of an efficiency problem and are doing rather well when it comes to perf/w, and using GDDR5x is a lot cheaper and easier to work with in regards to production (no extra silicon interposer, memory is cheaper because no TSVs, and no HBM assembly steps). Also Nvidia is getting almost as much memory bandwidth with 1080ti/Titan X as AMD will have with Vega.

Edit2: Ugh spelled it HMB and more than once facepalm.

 

[Arachnotronic]

Correct. Why else would Nvidia sign on for 1 year of GDDR5x for consumer cards, if they already have HBM2 cards..?

Because AMD has release rights for HMB2 on desktop.

Because HBM2 is expensive to implement, is expensive itself, makes it prohibitively expensive for large memory configurations, and is just generally a poor solution for consumer products when a GDDR-follow on that can offer similar performance is available.

It has nothing to do with "release rights."

 

[Glo.]

Two stack HBM2 with only a minor bandwidth advantage compared to cut down GP102 (8GB 512GB/s vs 11GB 484GB/s) isn't really impressive - Fury X had 50% more bandwidth than GTX 980 Ti. Meanwhile rumor says the competition is embracing 16GB GDDR6 at >14Gbps for their next generation.

Those two stacks will consume 8W of power, total, while those 11 memory cells will consume at best 35W(low clocked) and over 40W for high clocked, and still not able to achieve 100% of HBM2 bandwidth.

HBM technology was designed not for DGPUs in the first place, but to be used in other situations as well, which you completely forgot guys. It was used, because of manufacturing costs.

HBM technology, because of its complexity was expensive, HBM2, when you consider scale used to achieve the same/better effect than HBM - does not have to be.

 

[CatMerc]

I really don't understand why AMD has been messing around with HBM & HBM2 for the last few years. It appears to have hampered their ability to compete at all, due to the massively increased costs of developing and implementing such a radical new memory solution for consumer cards.

Nvidia is having massive success with GDDR5X - why on earth haven't AMD jumped on this bandwagon? Do NVIDIA have exclusive access to it?

Perhaps this is why Vega appears to have been delayed, maybe they are making a GDDR5X 11 Gbps version of the GPU instead? Though I guess it's far too late for this, unless Vega was designed to have both types of memory controller on it, which I doubt it possible/feasible.

Either way it looks likely that I'll be getting a AIB custom 1080ti when they are in stock (1-2 months I guess, due to insane demand) instead of Vega. I wish AMD could compete, though it seems their R&D and other budgets have all been neutered to launch Ryzen. Sad times.

AMD had no reason to use GDDR5X. RX 480 wouldn't benefit enough from it to justify costs, and that was their highest end card for a while.

Vega is an architecture, not a specific chip, they can have different chips with different memory controllers as they see fit, but there won't be both on the same chip as that's incredibly wasteful of die space.

HBM2 is the future, and AMD is investing in it. The more they use it, the better they can implement it, and the cheaper it gets due to economics of scale. It is more efficient than GDDR of any kind, and allows for form factors that aren't otherwise possible. Think about how their R&D into HBM2 on video cards can benefit their other endeavours like APU's.

Apparently Vega supports both HBM2 and GDDR5. Not sure about the GDDR5X, but from what I read it would be a substantial redesign over non-X and this probably wouldn't be financially viable vs putting resources into HBM2 (and I assume GDDR6) controllers. Also I'm not sure if they meant each Vega chip supports both, or if the two Vegas have a different controller.

One could also say it's sad to assume performance without seeing evidence, no?

The architecture does, but the specific RX Vega chip does not.

GDDR5X is also not a substantial redesign at all, in fact Polaris ALREADY supports GDDR5X. There's just no reason for them to use it, the card is too weak to justify GDDR5X. At most a refresh would have 9Gbps GDDR5.

Correct.
AMD worked with and has exclusive rights to HBM2, while Nvidia sign a deal with Crucial? for their GDDR5x. GDDR5x is just a stop gap because Nvidia can't release a HBM2 part, until after AMD has theirs out in the market.

So AMD can take all the time it wants, it has an uber advantage.

AMD has priority access for HBM2 from SK Hynix, but that does not mean exclusivity, and NVIDIA has Samsung.

 

[tential]

Ryzen is incredibly efficient in the applications AMD really cares about (servers and laptops). When clocked modestly - 3.0 GHz or less - it actually beats Intel's offerings in the same range in terms of perf/watt. A version on TSMC 16FF+ might have been better optimized for enthusiast clocks (4.0+ GHz), but the truth is that desktop gaming enthusiasts are a fairly small niche and it would have been economically foolish for AMD, with their limited resources, to focus on those needs when it might mean a worse product for bigger and more important markets.

Polaris is a different story. The primary reason it was a disappointment is that it was overhyped - not surprisingly, since it was basically AMD's only new product for 2016. Polaris is a slightly refined version of GCN; it would not be unfair to call it GCN 1.3. It was also a "pipe cleaner" product used to test the market readiness of the GloFo 14nm process and work out kinks. In other words, it was largely a placeholder so AMD would have something at least modestly competitive until Vega was ready. All other versions of GCN have their "sweet spot" in terms of clocks around 800-900 MHz and so did Polaris. 1266 MHz was pushing it, and AMD only did that so that the product would be within spitting distance of Hawaii in terms of performance. (They wanted to discontinue Hawaii since it was a massive die that didn't help them fill their WSA and required large, expensive boards.) I'm not at all convinced that fabbing Polaris on TSMC 16FF+ would have provided a huge advantage. It might have given some marginal benefits, but it was always going to require substantial effort on AMD's part to break the clock speed wall for GCN, and Polaris just did not incorporate those efforts.

You're preaching to the choir if you want to talk about Ryzen merits. I've been the one crying the most about this. Ryzen doesn't fit my specific use case. Of course that makes me sad. The design choices they have specifically make it impossible for me to use the processor the way I want. This still doesn't change what I've said either way that Ryzen is the best budget choice.
I'm saying Ryzen discussions are stupid in general. Ryzen 7 has been compared to the 7700k. If you compare Ryzen 7 to a 7700k, then you don't get what Ryzen 7 is for. PERIOD. It's like comparing an SUV to a sports car. They are both cars, but they have very different capabilities. Ryzen 7 can do server tasks. 7700k CAN NOT.
So my main complaint with Ryzen 7 is the discussion of Ryzen 7 vs 7700k in gaming when I think that's one of the most low intelligence level conversations you can have about a processor. It should be Ryzen 5 and the 7700k, and Ryzen 5 is just better value for the average gamer period. It's not a discussion, it's just a fact that I don't feel like even debating.
I don't even want to talk about it any further, it's a frustrating conversation that's happening all over the internet and it's way too simple.

Polaris was overhyped by some people. It was always a R9 290 replacement at a lower pricepoint. That's what AMD said.... people took that and ran with it to ridiculous levels. If we're being honest, with every GPU release on here, people take things and start twisting what is said to make everything crazy.... so I don't look at the extreme things people say on here. Polaris could have been better. Global Foundries. It sucks. You fab this chip somewhere else, and it will have marginal benefits. The thing is, how large are those marginal benefits....
Until Global Foundries is a reliable place that has a proven track record, I won't put faith in them.

Because HBM2 is expensive to implement, is expensive itself, makes it prohibitively expensive for large memory configurations, and is just generally a poor solution for consumer products when a GDDR-follow on that can offer similar performance is available.

It has nothing to do with "release rights."

I take a much simpler approach....
Because Nvidia with a DDR5x chip will still have the FASTEST chip on the market for god knows how long now? Even if HBM2 made Nvidia chips faster, it's not like they need it right now....

 

[JDG1980]

Because HBM2 is expensive to implement, is expensive itself, makes it prohibitively expensive for large memory configurations

The above is all 100% pure speculation. No one has ever posted any concrete figures on the relative cost of HBM2 versus GDDR5X. It seems likely that HBM-based setups are more expensive than standard GDDR5 setups, but we don't know how much more GDDR5X costs. If the extra chip costs for GDDR5X were negligible, then why use a 192-bit GDDR5 bus on GTX 1060 instead of a 128-bit GDDR5X bus? That would have saved on board routing costs.

It's not at all clear that Vega's two-stack HBM2 layout will necessarily be more expensive than GP102's 11- or 12-chip GDDR5X layout. The HBM2 solution may have higher costs for RAM and interposer (though we don't know even this), but it will be offset by a much smaller and simpler PCB that needs fewer VRM phases (power savings for HBM2 over GDDR5X).