I don't have great faith in Vega, but I don't necessarily hold Raja accountable just yet.

[Zion Halcyon]

First, I think Kyle's Article here is required reading:

https://www.hardocp.com/article/2016/05/27/from_ati_to_amd_back_journey_in_futility

In particular, note the complaints from Engineers and how Polaris fell well short of what they intended it to be.


Now consider that Polaris is part of the original Arctic Islands line (before Raja joined), which would culminate with Greenland, which is essentially Vega. Now consider that the team that developed the Arctic Islands was in place and conceived the line before Raja joined in 2013, and when he did join, Greenland would already be in the planning stages and therefore there would be significant loss of R&D $$ (likely millions) to scrap the thing and change course.

One thing Raja would be in time to do, however, is work on some newer tech for some of the other pieces of the card that could conceivably be put into Navi, which might make sense despite differing architectures since the one of the few talking points of Navi from the first RTG chart was scalability. So Vega might be used as a testing ground for some of the newer technologies they want for Navi, but at the core, it still is an Arctic Islands GPU.

I have no idea what Vega will be. I don't have high hopes for it to compete with the top end of NVidia simply because it seems like the Arctic Islands architecture is having scaling issues. I could be wrong.

But I really don't think we will know 100% what Raja can do for the RTG group until Navi - something that will be 100% on Raja. Until that time, it seems they are just trying to do the best they can with the lemons they were dealt to make lemonade.

I will say this - Raja better deliver with Navi. For as much as people keep saying that excuses are being made for AMD, that card will be 100% Raja, and he will have no excuse after that one...

(P.S. - I have already had some fun explaining this to the goofs in the AMD Reddit echo chamber, which is fun - 50% blind AMD fanbois, and 50% Intel/Nvidia shills, each pissed off for opposite reasons - it's kind of amusing!)

 

[Araxie]

problem with Vega it's strictly related to being more than one year later than Nvidia's main Vega Competitor which will be the GTX 1080 at best, a card launched in may 2016. at this point doesn't matter how good or bad can vega be, it will be irrelevant as nvidia already launched Volta pro segment, they are a full generation ahead AMD, even if Vega result to be a winner vs pascal, it have nothing to do vs Volta. so yeah, you are correct to not have great faith in Vega.

 

[Zion Halcyon]

problem with Vega it's strictly related to being more than one year later than Nvidia's main Vega Competitor which will be the GTX 1080 at best, a card launched in may 2016. at this point doesn't matter how good or bad can vega be, it will be irrelevant as nvidia already launched Volta pro segment, they are a full generation ahead AMD, even if Vega result to be a winner vs pascal, it have nothing to do vs Volta. so yeah, you are correct to not have great faith in Vega.

True in a sense.

The thing is, design is everything. We tend to rate things by generations and competition therein, and right now, AMD is a slave to the Arctic Islands design, and the cost already spent investing in it - they need to clear the books and recoup their R&D costs before they can actually move forward.

What little we know of Navi, is that it is supposed to focus on Scalability. In other words, comparing to Arctic Islands, that architecture didn't scale for shit - Polaris proved that. Navi is a blank slate. As you said, NVidia already launched their architecture - so they are locked into it for a while due to the R&D costs.

In other words, if AMD can hit an absolute home run with a fresh architecture design in Navi, NVidia suddenly finds themselves where AMD is now, but in a worse spot because they are in the BEGINNINGS of their new Architecture. There is SIGNIFICANT R&D cost in NVidia's cards right now, and while their company can absorb massive losses better than AMD can, if they choose that route, all that will do is put AMD on even more equal terms with NVidia, as it takes about 5 years from conception to taping out of a new graphics architecture.

So literally, everything hinges on Navi. And its entirely on Raja at that point. If Navi fails, RTG is DOA. Because whatever the new architecture is, AMD is locked into it for a significant portion of time, and going back to the article, the RTG may get their wish, but not because of how they planned it by becoming successful and splitting off, but because the Processor side of things is killing it with Ryzen and Threadripper, and the RTG is dragging the company down to the point where selling them off for cash is better for AMD proper.

I'm sure despite what I said, people will still be weeping and gnashing teeth over Vega if it ends up capturing the gaming mid-level like I think it will, but not challenging for the top spot. Still, I will withhold judgment until Navi for RTG for all the reasons I mentioned. RTG just needs to get out from under the Arctic Islands architecture for now, as quickly as possible (which is why I think they don't mind the crypto-miners buying up all their supply - they just want to recoup losses rather than stake their name on Polaris and the upcoming Vega, imho).

 

[Araxie]

"scalability" is a damn pretty looking marketing word.. do you remember the first marketing with the HD 7970 and GCN? "scalability" was the key word there, and of course it worked, scalability for AMD mean, they can make a base architecture and works through there according the market needs and apply refinements to that architecture year to year, and that's exactly what GCN did with all of their revisions from the 7970 (GCN 1.0) up to Polaris (GCN 1.4) and Vega which it's the same Polaris refined.. that's what scalability means for AMD.

 

[razor1]

Raja has been there for almost 5 years now as lead for both architecture and software, he is constrained by AMD's decisions in the past but Vega is his baby, and he will get blamed for it, if it doesn't measure up.

Also have to understand at this day and age there is really no such thing as a brand new architecture, its an evolutionary process, even Volta is just an evolution of Maxwell/Pascal.

GCN was made for DX11, and DX12 and other future API's, the API's are law lol, they are the ones that impact architectural differences from gen to gen. Now each IHV can and have created their own architectures which meet those API needs, and that is what we see as differences. The cost of creating these architectures which are constrained by the API's to a certain degree, and the cost of creating drivers are split through out the generations of using that architecture. nV is able to do pronounced architectural differences in their 1.5 years gen to gen (just have more R&D to push), which gives them certain advantages in the long run and in the short term can focus on specific games in a generation. AMD/ ATi has always been in the position of making less money (outside of the fx line from nV) and that is why they keep their major architectural revisions to 5 years per gen. With AMD's down slide recently, its looks like they stretched out that GCN architecture for a couple more years, bringing it to 7 years for one generation of architecture. In nV terms that is 4 generations of architectures. And this is why we see scaling problems, If the don't gut GCN and redo its pipelines, the through put issues just won't get fixed. And if they are going to gut GCN its not going to be GCN anymore.

GCN has the shader prowess to keep up and prior to Pascal beat nV's architectures, but games/applications just didn't take advantage of that.

Just a cursory look at Volta, Vega won't be able to keep up in shader performance with Volta, then we know Volta 2.0 is coming, out sometime next year, its probably still going to be behind Volta in TFlops.

 

[Zion Halcyon]

"scalability" is a damn pretty looking marketing word.. do you remember the first marketing with the HD 7970 and GCN? "scalability" was the key word there, and of course it worked, scalability for AMD mean, they can make a base architecture and works through there according the market needs and apply refinements to that architecture year to year, and that's exactly what GCN did with all of their revisions from the 7970 (GCN 1.0) up to Polaris (GCN 1.4) and Vega which it's the same Polaris refined.. that's what scalability means for AMD.

Not arguing that its a marketing catchphrase. However that doesn't mean they aren't going to attempt to do what I said.

Just remember that there is a different crew now in charge of RTG, so its not fair to judge them by their predecessors.

I'll liken it to an NFL team that just fired their GM and head coach, and brought in a new GM and Head Coach.

For the first couple years, both guys are going to be working to turn over the roster, but they will have to go to war, so to speak, with some of the crap left over from the previous regime. Whether they say the same things the previous regime did or not doesn't matter. What will matter is, once they do get the roster turned over, can "their" team that they put together compete? If they can, it won't matter what kind of NFL GM speak or Coachspeak those guys use - fans won't care because the team can win. However, if they get their guys and still suck, both those guys will be fired.

Navi is "Raja's team". The previous Arctic Islands architecture is the previous team. I won't say you shouldn't be skeptical, but it is kind of silly to hold someone accountable for the mistakes of the guy he replaced (and who was replaced because he already was held accountable for those mistakes).

Sure, scalability might be generic and market-y. But I think its too early for people to be all "same old AMD" - we won't know that until Navi.

 

[razor1]

I think Navi is AMD's last hope to maintain some sort of equality in performance with nV, if it doesn't happen with Navi, AMD isn't going to come back unless nV screws up.

 

[Phelptwan]

I think Navi is AMD's last hope to maintain some sort of equality in performance with nV, if it doesn't happen with Navi, AMD isn't going to come back unless nV screws up.

It's honestly at the point where it's pretty against the odds that even a breakout performance card in Navi is going to get them back in the game significantly...mostly because they just can't seem to meet the production demands.

At this point, I honestly believe RTG is going to drop out of the high end game in the future, and focus solely on mid range cards and workstation/pro/server cards.

 

[lolfail9001]

Navi is "Raja's team". The previous Arctic Islands architecture is the previous team

I am 99% positive that Arctic Islands (that btw span 2 different architectures, mind you) and Navi will be made by the same Shanghai team.

So, whether it is Raja at helm or not is at this point irrelevant.

 

[Zion Halcyon]

I am 99% positive that Arctic Islands (that btw span 2 different architectures, mind you) and Navi will be made by the same Shanghai team.

So, whether it is Raja at helm or not is at this point irrelevant.

I'd like to rebut this with the point that Bulldozer and Ryzen were made by mostly the same engineers too. The difference was Jim Keller was brought in from the outside to design Ryzen. So despite it being mostly the same teams, what really matters is who they have overseeing the design.

 

[razor1]

It's honestly at the point where it's pretty against the odds that even a breakout performance card in Navi is going to get them back in the game significantly...mostly because they just can't seem to meet the production demands.

At this point, I honestly believe RTG is going to drop out of the high end game in the future, and focus solely on mid range cards and workstation/pro/server cards.

That would be a bad move cause that means they will also pretty much stop working on architectures for the AI/ HPC market.

 

[Zion Halcyon]

It's honestly at the point where it's pretty against the odds that even a breakout performance card in Navi is going to get them back in the game significantly...mostly because they just can't seem to meet the production demands.

At this point, I honestly believe RTG is going to drop out of the high end game in the future, and focus solely on mid range cards and workstation/pro/server cards.

Completely and entirely possible. However if you take a look at the reason why they can't meet production demands, it hinges entirely on them betting on the wrong horse in terms of memory. The companies who can produce HBM and HBM 2 memory keep running into yield issues, which in turn leads to supply issues for AMD.

For Navi, they said "Next Gen Memory". Its obviously a marketing term that could mean anything, but what should also be noted is it also means they aren't tied to HBM for Navi. My gut tells me that they will go with whatever the fastest memory is that ALSO can be produced en masse for Navi. I highly doubt they go back to HBM after being burned by supply problems for both versions 1 and 2 of the RAM.

 

[Sprayingmango]

I don't understand how ANYONE could have faith in AMD....they consistently over promise and under deliver year after year. I wish they could deliver a home run, I would switch my entire system to AMD if so. Sadly Intel and Nvidia just destroy them.

 

[Phelptwan]

That would be a bad move cause that means they will also pretty much stop working on architectures for the AI/ HPC market.

I considered "Server" cards to be part of the HPC/AI market.

Completely and entirely possible. However if you take a look at the reason why they can't meet production demands, it hinges entirely on them betting on the wrong horse in terms of memory. The companies who can produce HBM and HBM 2 memory keep running into yield issues, which in turn leads to supply issues for AMD.

For Navi, they said "Next Gen Memory". Its obviously a marketing term that could mean anything, but what should also be noted is it also means they aren't tied to HBM for Navi. My gut tells me that they will go with whatever the fastest memory is that ALSO can be produced en masse for Navi. I highly doubt they go back to HBM after being burned by supply problems for both versions 1 and 2 of the RAM.

Who knows, but hopefully Navi is flexible enough that they don't get hammered into an issue with a single vendor/type. At this point, I'm very curious if the Vega delays have pushed back Navi significantly though.

 

[Zion Halcyon]

I don't understand how ANYONE could have faith in AMD....they consistently over promise and under deliver year after year. I wish they could deliver a home run, I would switch my entire system to AMD if so. Sadly Intel and Nvidia just destroy them.

Well, I feel like AMD more or less delivered with Ryzen. Still some growing pains but its competitive with Intel and cheaper, and a hell of a deal for people who need the cores.

On the graphics front, I agree. I don't have any faith in AMD's RTG group to right the ship. But if it makes sense, I don't doubt them either. In other words, I have no opinion one way or the other, because I want to wait until Navi, so that I can feel like I made an informed decision on the direction AMD is going with their graphics, rather than a kneejerk one based on emotion.

What makes me sad about if they fail though, is I think the marketing and business side of AMD is killing it right now. From making their API open source, to partnerships with Kronos and Microsoft that ensure oft ignored architecture in their cards gets used in DX12 and Vulkan (which led to some surprising performance boosts), to their deal with Bethseda, and so on. They even were able to "spitshine a turd" as Kyle put in his article to sell Polaris to the low end - which, while you may not like that they had to do it, the fact that they were able to says a lot about their ability to keep that side of things solvent.

At this stretch, it feels like they are doing EVERYTHING just about right EXCEPT putting out a quality graphics card. And I can understand that if they have an Arctic Islands boat anchor around their neck they have to get rid of before moving forward. I get that. But that also means they better hope Navi is a hot air balloon and not another boat anchor, because if it is the latter, this time they are going to drown...

 

[Zion Halcyon]

I considered "Server" cards to be part of the HPC/AI market.



Who knows, but hopefully Navi is flexible enough that they don't get hammered into an issue with a single vendor/type. At this point, I'm very curious if the Vega delays have pushed back Navi significantly though.

Possibly. Just the vibe I get, but I think they know the stakes with Navi. They want to get it just right. I think they are eager to sell off the stock for Vega (and Polaris before it) - I wouldn't be surprised to hear that they would consider breaking even on those cards a win.

 

[Magic Hate Ball]

I'm still fairly certain that Navi will be their Epyc play for GPU instead of CPU.

They're stitching together smaller CPUs together in a single socket to compete with monolith CPUs from Intel. If they can improve the Infinity Fabric communications, I foresee multi-GPU acting as a single GPU being their future with Navi.

Smaller pieces of silicon put together to make something say 4x as large? You can have a 1x chip for low end, 2x for midrange, and a 4x for high end.... exactly the way Zen is being rolled out. And you'll only need to make the same small chip for all of them. With the way production works with silicon, even having an interposer will be cheaper with this method than large die manufacturing.

 

[Zion Halcyon]

I'm still fairly certain that Navi will be their Epyc play for GPU instead of CPU.

They're stitching together smaller CPUs together in a single socket to compete with monolith CPUs from Intel. If they can improve the Infinity Fabric communications, I foresee multi-GPU acting as a single GPU being their future with Navi.

Smaller pieces of silicon put together to make something say 4x as large? You can have a 1x chip for low end, 2x for midrange, and a 4x for high end.... exactly the way Zen is being rolled out. And you'll only need to make the same small chip for all of them. With the way production works with silicon, even having an interposer will be cheaper with this method than large die manufacturing.

Yep, seeing the same here.

 

[Zion Halcyon]

Raja has been there for almost 5 years now as lead for both architecture and software, he is constrained by AMD's decisions in the past but Vega is his baby, and he will get blamed for it, if it doesn't measure up.

Also have to understand at this day and age there is really no such thing as a brand new architecture, its an evolutionary process, even Volta is just an evolution of Maxwell/Pascal.

GCN was made for DX11, and DX12 and other future API's, the API's are law lol, they are the ones that impact architectural differences from gen to gen. Now each IHV can and have created their own architectures which meet those API needs, and that is what we see as differences. The cost of creating these architectures which are constrained by the API's to a certain degree, and the cost of creating drivers are split through out the generations of using that architecture. nV is able to do pronounced architectural differences in their 1.5 years gen to gen (just have more R&D to push), which gives them certain advantages in the long run and in the short term can focus on specific games in a generation. AMD/ ATi has always been in the position of making less money (outside of the fx line from nV) and that is why they keep their major architectural revisions to 5 years per gen. With AMD's down slide recently, its looks like they stretched out that GCN architecture for a couple more years, bringing it to 7 years for one generation of architecture. In nV terms that is 4 generations of architectures. And this is why we see scaling problems, If the don't gut GCN and redo its pipelines, the through put issues just won't get fixed. And if they are going to gut GCN its not going to be GCN anymore.

GCN has the shader prowess to keep up and prior to Pascal beat nV's architectures, but games/applications just didn't take advantage of that.

Just a cursory look at Volta, Vega won't be able to keep up in shader performance with Volta, then we know Volta 2.0 is coming, out sometime next year, its probably still going to be behind Volta in TFlops.

I know Raja will get the brunt for Vega, but I really don't think it will be all his fault - he still is being saddled with a poorly designed Arctic Islands architecture in Greenland.

I do agree with the overall sentiment in your later post that Navi is the make or break. You may be right that tech tends to be more evolutionary than revolutionary, Zen actually shows us the path to how you can still make something that differs architecturally from its predecessor that stands up on its own and allows you to catch up. AMD already identified the reason for the performance gap now between Zen and the Intel chips is purely Clocks and GhZ, and they are working on upping those for Zen 2. Navi needs to be Zen for AMD's RTG group. And I agree - if Navi bombs, if I were Lisa Su, I'd sell off RTG to intel if Intel still wants them.

 

[lolfail9001]

Smaller pieces of silicon put together to make something say 4x as large? You can have a 1x chip for low end, 2x for midrange, and a 4x for high end.... exactly the way Zen is being rolled out. And you'll only need to make the same small chip for all of them. With the way production works with silicon, even having an interposer will be cheaper with this method than large die manufacturing.

What you fail to realize is that ultimately, GPUs are a single pipeline with massively parallel stages, not multiple separate pipelines like a multi core CPU. Zen MCM play simply will not work with GPUs.

 

[Magic Hate Ball]

What you fail to realize is that ultimately, GPUs are a single pipeline with massively parallel stages, not multiple separate pipelines like a multi core CPU. Zen MCM play simply will not work with GPUs.

With a fast enough interconnect and a low enough latency, I don't see how single package mGPU is out of the realm of possibilities.

 

[Zion Halcyon]

What you fail to realize is that ultimately, GPUs are a single pipeline with massively parallel stages, not multiple separate pipelines like a multi core CPU. Zen MCM play simply will not work with GPUs.

Truth is though, none of us know what AMD will have under its hat with Navi until they debut it. No one ever sees the next big thing coming until it actually shows up and blows everyone away, and people have gone poor/insane trying to predict the "next big thing" because if it were that easy, companies would do it more often.

I don't blame someone who is skeptical of AMD delivering with Navi - AMD has a very poor track record, and only just started reversing that with Zen (and who knows if they can capitalize with future revisions on that architecture). I also don't pick on someone who is hopeful that Navi will be a leap in performance to bring it in line with NVidia's offerings, because thus far, some of the business decisions made by the RTG have been downright savvy and smart, and perhaps they feel that leadership shows confidence in where they are going in the future.

Point blank, I am reserving judgment either way, even if Vega is a flaming turd. If Raja thinks he's such hot shit, let's see what he does with his first pure offering with AMD in Navi. If it succeeds, I'll give the man props and his self-confidence is warranted. If it fails, I'll lead the bandwagon for AMD to sell off RTG to Intel. And then wonder if there's possibly anyone out there who can design something that can compete with NVidia...

 

[lolfail9001]

With a fast enough interconnect and a low enough latency, I don't see how single package mGPU is out of the realm of possibilities.

Think again, how that would work. Did? Great. Now think why SFR is never used.

 

[Magic Hate Ball]

Think again, how that would work. Did? Great. Now think why SFR is never used.

Oh the guy who said Ryzen could never clock to 4ghz on air agrees with you.

All I ever see from your posts is nay saying. Perhaps some real contributions?

 

[razor1]

I know Raja will get the brunt for Vega, but I really don't think it will be all his fault - he still is being saddled with a poorly designed Arctic Islands architecture in Greenland.

I do agree with the overall sentiment in your later post that Navi is the make or break. You may be right that tech tends to be more evolutionary than revolutionary, Zen actually shows us the path to how you can still make something that differs architecturally from its predecessor that stands up on its own and allows you to catch up. AMD already identified the reason for the performance gap now between Zen and the Intel chips is purely Clocks and GhZ, and they are working on upping those for Zen 2. Navi needs to be Zen for AMD's RTG group. And I agree - if Navi bombs, if I were Lisa Su, I'd sell off RTG to intel if Intel still wants them.

True I don't think its Raja's fault either, unfortunate for him he will be in the hot seat.

Now AMD will not have the consoles to push developers towards Vega's special features, Xbox Scorpio and PS4 pro, both are still on Polaris type architectures. That is going to be another area they have to spend money on and work with developers. If Vega falls short, Navi's R&D would likely stay similar to Vega, that isn't enough to match up with nV. Normally I am a firm believer in the capability of creating a great product is in the hands and not in the wallet. But in this case they need both, They might have the minds, or might be able to get the minds, but the amount of money to create these products, AMD just doesn't have enough. The longer they stay with out making profits, the more they will fall into a bottomless pit. Zen was a good step to get them out of a hole, but by itself, doesn't look like it can do much for another 2 or 3 quarters maybe longer. By then Navi should be near completion of design (according to possible AMD timelines). And as you stated the future of Zen processors is based on their own intangible process changes and Intel on the other side has woken up.

Maybe its a bit too late already for Navi. And this is why they are doing Vega 2.0. What doesn't make sense in all this, with AMD's reduced budgets, why have two separate IP's for their line up? That actually increases R&D costs, not reduce. What it does do is reduce risk vs potential profits.

Now they did the same thing with Polaris and Vega, having two separate IP's. Why did this happen, probably because big Polaris couldn't match up with Pascal, which AMD knew that it wouldn't when they saw Maxwell. So AMD didn't bother with it and went straight for Vega, which coincidentally is coming a year and a bit more after Polaris, hence according to AMD timelines, its expedited about 1 to 1.5 quarters, which is the maximum amount of time they can get when expediting future architectures.

 

[razor1]

Oh the guy who said Ryzen could never clock to 4ghz on air agrees with you.

All I ever see from your posts is nay saying. Perhaps some real contributions?

Multi die processors on GPU's are not cost effective yet and won't be till interposers and manufacturing of them and fabricating them come down to reasonable prices along with the high speed interconnects too. For AMD they would need to use what something in the neighborhood of x4 or x5 the infinity fabric speeds they have on their server parts? Maybe more, haven't looked into the numbers much but at least that.

Just a little bit of latency in this causes what kinda of performance hit on CPU's, think about graphics processors which operate at much faster through put.

 

[Zion Halcyon]

Oh the guy who said Ryzen could never clock to 4ghz on air agrees with you.

All I ever see from your posts is nay saying. Perhaps some real contributions?

Frankly, I think it's too early to fight over something that might not even be in Navi's architecture. Seems like wasted energy to me to argue over hypotheticals based on as of yet nonexistent architecture.

The reason is, based on perspective, you both could be right. Up until now, the architecture you are describing hasn't worked. That doesn't mean some very smart engineer isn't going to figure out a better way that makes it work using a new design. However, as I said before, predicting how or when this will happen is impossible, therefore better to save the brawls for when someone else actually owns up to attempting the layout you are describing in a graphics card.

 

[Ieldra]

It wasn't 'next gen' memory for navi. It was nexgen memory. That's a company name, it was bought by AMD a few years back.

Scalability is extremely vague and I think OP is just trying too hard to spin it into a theory about future GPUs.

What does it mean when you say Polaris scales badly? Scales with what? Resolution? Memory bandwidth?

Polaris performs far better per flop than Fiji, they've alleviated some of the front end bottlenecks but introduced new ones with only 32 ROPs.

Weve not yet seen a GCN gpu with more than four SEs, even Vega will retain the same geometry throughput. Hope navi will fix that

 

[razor1]

Scalability is extremely vague and I think OP is just trying too hard to spin it into a theory about future GPUs.

What does it mean when you say Polaris scales badly? Scales with what? Resolution? Memory bandwidth?

Polaris performs far better per flop than Fiji, they've alleviated some of the front end bottlenecks but introduced new ones with only 32 ROPs.

Weve not yet seen a GCN gpu with more than four SEs, even Vega will retain the same geometry throughput. Hope navi will fix that

Its all about die size, I think it would be hard for them to increase their SE's and keep the die size respectable for Vega.

Polaris performs better per flop than Fiji true (much better), but is still less performance per transistor than a r390x when considering over all applications, of course consideration of power too, Polaris is much better.

Shintai pointed out something very interesting with "scalability", what he thinks it means is top to bottom line up without HPC and DL parts, which is what nV has gone to. Split their architecture and brands based on target markets. I think with Fermi nV learned a valuable lesson, to take on all three markets with identical architectures, created many problems specially on cost and power usage. AMD with GCN hit the same road block.

 

[KazeoHin]

If Vega is a turd, it is still Raja's fault. They had working silicon playing games at 1080 levels last year.

Even if Raja is not to blame for the underlying architecture, he is wholeheartedly to blame for this product being 6 months late.

 

[WorldExclusive]

I would delay Vega if I where Raja too.
People are going to get fired right after its launch, so why not keep you friends on the books for as long as possible. It's the least he can do.

 

[Maddness]

I would delay Vega if I where Raja too.
People are going to get fired right after its launch, so why not keep you friends on the books for as long as possible. It's the least he can do.

Do you really think it will be that bad, or do you mean because it is so late to the fight, or a bit of both.

 

[polydiol]

there are a lot of hurdles for Vega but I hope Vega does well because its looking like its still going to be far behind.

 

[SighTurtle]

Would they really fire Raja for this? Seems more circumstance than anything else. Unless Raja was the main promoter of HBM for GPUs.

Edit: Assuming of course that the HBM shortage is the main reason AMD cant release Vega.

 

[Zion Halcyon]

Would they really fire Raja for this? Seems more circumstance than anything else. Unless Raja was the main promoter of HBM for GPUs.

Edit: Assuming of course that the HBM shortage is the main reason AMD cant release Vega.

They aren't going to fire Raja over HBM shortages- that's silly talk.

As I said, if Navi tanks, that is a different story. But if that happens, they should consider selling off the entire RTG division to Intel. Even not as a direct NVidia competitor, RTG is still light-years ahead of Intel's own integrated solutions.

 

[Zion Halcyon]

there are a lot of hurdles for Vega but I hope Vega does well because its looking like its still going to be far behind.

I think Vega will land somewhere between the 1070 and 1080 if I had to guess...

 

[Maddness]

I think Vega will land somewhere between the 1070 and 1080 if I had to guess...

Really, that bad. I would think it would have to best the 1080 to be a success. Price wise it will more than likely be expensive considering the HBM supply issue. At least at launch anyway.

 

[Vader1975]

True in a sense.

The thing is, design is everything. We tend to rate things by generations and competition therein, and right now, AMD is a slave to the Arctic Islands design, and the cost already spent investing in it - they need to clear the books and recoup their R&D costs before they can actually move forward.

What little we know of Navi, is that it is supposed to focus on Scalability. In other words, comparing to Arctic Islands, that architecture didn't scale for shit - Polaris proved that. Navi is a blank slate. As you said, NVidia already launched their architecture - so they are locked into it for a while due to the R&D costs.

In other words, if AMD can hit an absolute home run with a fresh architecture design in Navi, NVidia suddenly finds themselves where AMD is now, but in a worse spot because they are in the BEGINNINGS of their new Architecture. There is SIGNIFICANT R&D cost in NVidia's cards right now, and while their company can absorb massive losses better than AMD can, if they choose that route, all that will do is put AMD on even more equal terms with NVidia, as it takes about 5 years from conception to taping out of a new graphics architecture.

So literally, everything hinges on Navi. And its entirely on Raja at that point. If Navi fails, RTG is DOA. Because whatever the new architecture is, AMD is locked into it for a significant portion of time, and going back to the article, the RTG may get their wish, but not because of how they planned it by becoming successful and splitting off, but because the Processor side of things is killing it with Ryzen and Threadripper, and the RTG is dragging the company down to the point where selling them off for cash is better for AMD proper.

I'm sure despite what I said, people will still be weeping and gnashing teeth over Vega if it ends up capturing the gaming mid-level like I think it will, but not challenging for the top spot. Still, I will withhold judgment until Navi for RTG for all the reasons I mentioned. RTG just needs to get out from under the Arctic Islands architecture for now, as quickly as possible (which is why I think they don't mind the crypto-miners buying up all their supply - they just want to recoup losses rather than stake their name on Polaris and the upcoming Vega, imho).

Well said sir.

 

[Captain Newmackwa]

Its all about die size, I think it would be hard for them to increase their SE's and keep the die size respectable for Vega.

Polaris performs better per flop than Fiji true (much better), but is still less performance per transistor than a r390x when considering over all applications, of course consideration of power too, Polaris is much better.

Shintai pointed out something very interesting with "scalability", what he thinks it means is top to bottom line up without HPC and DL parts, which is what nV has gone to. Split their architecture and brands based on target markets. I think with Fermi nV learned a valuable lesson, to take on all three markets with identical architectures, created many problems specially on cost and power usage. AMD with GCN hit the same road block.

With regards to "scalability", that's what I thought so too. Nvidia's focus on efficiency has allowed them the flexibility to offer chips for a variety of applications across different markets and industries. Looking at their GeForce/Titan line of GPUs for example, their perf-per-watt advantage has allowed them to have high performance chips in gaming laptops all the way up to "money is no object" halo graphics cards while AMD/RTG is seemingly struggling to fill in the gaps in their product stack at the moment with present architectures. I believe this is the "scalablity" issue that RTG wants to address going forward.

 

[WorldExclusive]

Do you really think it will be that bad, or do you mean because it is so late to the fight, or a bit of both.

I think it's just late. AMD tends to wrry about things that will matter four years from now, while Nvidia focuses on the current market.
HBM is great, but we still have a ways to go before making it useful for desktop users. Design decisions that doesn't enhance core performance and lead to delays will get anyone fired.

Raja maybe safe because no one will want to replace him. But many people will have to fall on their own sword to save him.