AMD: Navi Speculation, Rumours and Discussion [2017-2018]

Status
Not open for further replies.
I think your observation about node-change years is very interesting. Has someone done an NVidia history like this? The history you've drawn up does seem to paint a clear picture of AMD's retreat from graphics since 2013.

(One could put the console GPUs in that history, but I really think that's off-topic).

But what if, in future, there is literally never more than 1 chip per year. "Scalable" chips, like we're expecting Navi, onwards, to be? It might not be a problem if 4 or more cards can be derived from a single chip at the relevant performance-pair (pro and XT) levels. Perhaps covering mainstream, performance and enthusiast segments with 6 configurations (1, 2 and 4 chips - all three in pro and XT variants).

In a similar fashion to how Fiji was the "HBM pipe cleaner", I'm expecting Navi to be the "MCM pipe cleaner". So what Wang says might merely align with that plan. I'm not trying to suggest it's a non-subtle hint by him or some kind of confirmation, merely pointing out that what he said might not be coincidence. And that MCM scalability is what makes this "not a coincidence".

It may be that either one of "scalability" and "nextgen memory" is lost in Navi. e.g. it only results in 1 and 2 chip variants (has scalability) and uses GDDR6 (so not nextgen memory). To me GDDR6 is just kicking the can down the road while waiting for HBM to get cheaper, so doesn't qualify as "nextgen".
 
Aside from process shrink years, haven't we already had basically one major chip per year since GCN? […]
2012 - 3 - Tahiti, Pitcairn & Cape Verde (28nm debut)
2013 - 3 - Hawaii & Bonaire + Oland
2014 - 1 - Tonga
2015 - 1 - Fiji
2016 - 2 - Polaris 10 & 11 (14nm debut)
2017 - 2 - Vega 10 + Polaris 12
2018 - 2 - Vega Mobile & Vega 20

I'd say 2 GPUs per year since the launch of GCN (at the average).
 
I know it's a bit silly because it's not based on much at all, but I have a bad feeling about Navi. We had lots of hype about Polaris, lots of hype about Vega, and ultimately they were decent, at best, but underwhelming. And now, with Navi? We barely know anything at all. AMD is dead silent. I have a hunch that AMD is treating GCN as it was treating the Bulldozer family while working on Zen: just doing the bare minimum to keep the lights on until the real product can be released, which would be the post-Navi generation.

I hope I'm wrong but I fear we may not get any real competition until 2020, at best.
 
In a similar fashion to how Fiji was the "HBM pipe cleaner", I'm expecting Navi to be the "MCM pipe cleaner". So what Wang says might merely align with that plan. I'm not trying to suggest it's a non-subtle hint by him or some kind of confirmation, merely pointing out that what he said might not be coincidence. And that MCM scalability is what makes this "not a coincidence".
That would make sense with Vega20 seemingly being a MCM-ish design with xGMI. Gaming Vega should have been able to do the same, but I'm guessing there were issues with the geometry bottleneck that Navi may address. AMD likely still needs multiple chip variants unless they roll out FP64 on all consumer cards. Vega20+Navi might suffice until GCN is phased out. At that point we may be looking at x86 instruction sets as opposed to GCN with far more integrated memory systems.
 
I know it's a bit silly because it's not based on much at all, but I have a bad feeling about Navi. We had lots of hype about Polaris, lots of hype about Vega, and ultimately they were decent, at best, but underwhelming. And now, with Navi? We barely know anything at all. AMD is dead silent. I have a hunch that AMD is treating GCN as it was treating the Bulldozer family while working on Zen: just doing the bare minimum to keep the lights on until the real product can be released, which would be the post-Navi generation.

I hope I'm wrong but I fear we may not get any real competition until 2020, at best.
I hope Sony/MS hold enough sway to force AMD to put in more effort.
 
AMD guy Wang says they'll be launching 1 GPU per year. Perhaps we should take this literally: a single chip per year. And then it's a question of whether all chips, from now on, will be seen in single/double etc. (?) configurations on cards.

This is Wang's exact quote on the matter on PCWorld's article:

Nonetheless, in a small roundtable following AMD’s presentation, Wang confirmed that AMD would be bringing out a new graphics product every year, via a new architecture, process changes, or “maybe incremental architecture changes.
He specifically mentions a new architecture or incremental architecture upgrade or process change.
His following comments were about AMD's graphics releases gaining momentum to guarantee release schedules that are "more fun".

As mentioned in a previous post, AMD starting to release a single chip per year would result in AMD actually losing momentum because they've been releasing more than one graphics product per year on average (especially if we count with iGPUs).
There are also statements claiming that RTG had "some wastage on the GPU roadmap" while Raja was there, though the reason for that seems to navigate between "Raja's will" and Sony demanding too many changes to Navi.



So given the information above, I think it's safe to assume that AMD is planning to increase the number of GPUs being released every year, not decrease as one-GPU-per-year would imply.
Though a "tick" or "tock" per year probably doesn't mean an entire family of GPUs every year either. Up until 2012 AMD had a pretty decent track record of 3/4 distinct GPU chips per year.
Even the 2012 GCN launch lineup "only" brought 3 distinct GPU chips (Cape Verde, Immortal Pitcairn and Tahiti), with older Terascale 2 chips filling the gaps on the low end. In 2013 AMD released another 2 distinct chips now on GCN2 (Bonaire and Hawaii).
And then Bulldozer's losses started making an effect on the whole company, and the severe R&D cuts had an impact on GPU releases.



  • 2017 - 1 - Vega 10
  • 2018 - ? - (Does Vega 20 "count" if it's not in consumer desktop cards?)
Vega 10 and Raven Ridge Vega for 2017; Vega M and Vega 20 for 2018.
Why wouldn't they count if it's not a consumer desktop card? Regardless, it seems Vega 20 will indeed see a consumer release according to the PCWorld article (though maybe not in 2018?).


  • this midrange solution should reach the performance level of a GeForce GTX 1170/2070 and relies on GDDR6 memory (AMD works on GDDR6 interfaces)
  • speculatively this solution can reach the performance of a Radeon RX Vega 64 (+X%) for the power consumption of a Radeon RX 480 (150W)
Unless the 1070 -> 1170 performance upgrade is substantially smaller than what we saw in previous generations, these bullet points seem mutually exclusive.
To be honest, I don't expect the Navi RX670/680 to be on the level of nvidia's 70/80 series. AMD's x70/x80 have been countering nvidia's x60 for 4 years now, since the GTX 960 was positioned against the R9 285.
 
Vega 10 and Raven Ridge Vega for 2017; Vega M and Vega 20 for 2018.
Why wouldn't they count if it's not a consumer desktop card? Regardless, it seems Vega 20 will indeed see a consumer release according to the PCWorld article (though maybe not in 2018?).

I'm not trying to measure AMD's total engineering capacity. I'm trying to find a situation where the "1 chip/year comment" makes sense.

If you count "everything", then you blow through 1 chip/year easily.
  • AMD's semi-custom business does an average of 1 chip/year by itself since its inception.
    • 2013 - 2 - XB1 & PS4
    • 2014 - 0
    • 2015 - 0
    • 2016 - 3 - PS4Pro, PS4S & XB1S
    • 2017 - 1 - XB1X
    • 2018 - 1 - Vega M
  • And Raven Ridge isn't the only APU in the last half decade. Do all of these "count"?
    • 2011 - 2 - Llano & Brazos
    • 2012 - 1 - Trinity
    • 2013 - 2 - Richland & Kabini
    • 2014 - 2 - Kaveri & Beema
    • 2015 - 2 - Carrizo & Carrizo-L
    • 2016 - 2 - Bristol Ridge & Stoney Ridge
    • 2017 - 1 - Raven Ridge
    • 2018 - 0
  • Then we have to count all of the little OEM-friendly GPUs (e.g. Oland, Polaris 12, etc) that we see every couple years. No way those ever go away.
There's simply no realistic way for AMD to get down to 1 chip/year unless they are talking about only major dGPU releases, hence the criteria for my original list.
 
Then we have to count all of the little OEM-friendly GPUs (e.g. Oland, Polaris 12, etc) that we see every couple years. No way those ever go away.
With the upcoming Chromebooks, KabyG, and AMD APUs the "OEM-friendly" designs could get phased out in favor of more compact designs. At least now that APUs have comparable performance. Really don't need the dGPU outside of the top tiers and workstations. Even then it wouldn't be surprising to see a Threadripper/Epyc APU just for compute acceleration with the added memory channels. Some of the Vega20 and Infinity roadmap already suggest that. May be a 2nd socket, but a 2S consumer board may be practical.
 
Are they planning a new console before the post-navi generation?

No, Sony just recently announced the PS5 would come in 2021. "Coincidentally" a brand new, post GCN architecture was announced by AMD, to hit by 2021, around a month before that.

Whatever the PS5 uses, it's promised to be a completely new architecture. Would love to see a concentration on fast raytracing and splatting performance, Polys and Raster are getting too expensive for making games. In contrast voxels and related data structures can be a lot easier to work with from an art perspective, not to mention things like volumetrics and GI performance, and programming, is much, much easier. In fact programming with tracing in general is easier.
 
No, Sony just recently announced the PS5 would come in 2021. "Coincidentally" a brand new, post GCN architecture was announced by AMD, to hit by 2021, around a month before that.

Whatever the PS5 uses, it's promised to be a completely new architecture. Would love to see a concentration on fast raytracing and splatting performance, Polys and Raster are getting too expensive for making games. In contrast voxels and related data structures can be a lot easier to work with from an art perspective, not to mention things like volumetrics and GI performance, and programming, is much, much easier. In fact programming with tracing in general is easier.
When did Sony announce 2021?
 
Are they planning a new console before the post-navi generation?
Post navi is expected around 2020, Sony will want final dev kits well before post navi is ready.
At best we can expect Navi with next gen special sauce.
 
Interesting article that sounds legit to me.. or at least good guesswork.

https://wccftech.com/exclusive-amd-navi-gpu-roadmap-cost-zen/

If accurate, it confirms my fears: no high-end before 2021 is just… Bulldozer all over again. Still, I thought Navi was supposed to be the last iteration of GCN, not the brand new architecture. So which is it? I suppose the whole chiplet thing could be real, in which case a dual-Navi 10 could be pretty good, but I'm not optimistic.
 
When did Sony announce 2021?

About 2 weeks ago. Straight to The Wallstreet Journal from the CEO of their Entertainment division.

Supposedly the next Xbox could arrive in 2020. No doubt because MS isn't doing very well this generation.

That one could be Navi. Or the first run of AMD's new gen, done pre consumer GPU release just for MS. They did it before with the 360 so there's precedent.

Or it could be Nvidia, the Switch shows they're finally willing to do consoles on more, generous terms. Or heck Qualcomm for all we know. Or just a false rumor. Either way the Sony thing seems like a far better bet with the planned release date for both "just so happening" to line up, and with Sony doing very well this gen there's all the reason in the world to keep with what works.

Also yeah the Wccftech thing is false as hell. They've seemingly been making up shit off the top of their heads recently. No "faster Vega 64" showed up at Computex, exactly when you'd expect such to launch, among other things they seem to have gotten patently wrong. And having an officially announced new graphics architecture launching in the same time frame as the PS5 makes the "Navi being built specifically for the PS5" laughable.
 
Or it could be Nvidia, the Switch shows they're finally willing to do consoles on more, generous terms.
In this instance I don't think it indicates any such desire from Nvidia to become more involved in consoles. Their involvement with the switch is basically to provide a shit load of Tegra X-1 cores they had doing nothing in their warehouse.
 
About 2 weeks ago. Straight to The Wallstreet Journal from the CEO of their Entertainment division.

Supposedly the next Xbox could arrive in 2020. No doubt because MS isn't doing very well this generation.

That one could be Navi. Or the first run of AMD's new gen, done pre consumer GPU release just for MS. They did it before with the 360 so there's precedent.

Or it could be Nvidia, the Switch shows they're finally willing to do consoles on more, generous terms. Or heck Qualcomm for all we know. Or just a false rumor. Either way the Sony thing seems like a far better bet with the planned release date for both "just so happening" to line up, and with Sony doing very well this gen there's all the reason in the world to keep with what works.

Also yeah the Wccftech thing is false as hell. They've seemingly been making up shit off the top of their heads recently. No "faster Vega 64" showed up at Computex, exactly when you'd expect such to launch, among other things they seem to have gotten patently wrong. And having an officially announced new graphics architecture launching in the same time frame as the PS5 makes the "Navi being built specifically for the PS5" laughable.
You may want to actually read the quote and its context before saying Sony announced ps5 in 3 years :)
 
Interesting article that sounds legit to me.. or at least good guesswork.

https://wccftech.com/exclusive-amd-navi-gpu-roadmap-cost-zen/
Even if it does have some truth in it, I seriously doubt it's the whole truth, especially the future part.
AMD has previously taken input from both MS and Sony for GCN development and those changes have appeared in same generation, it's highly unlikely they'd suddenly start developing for example Navi only for Sony's needs (MS is going to launch next Xbox around the same time frame anyway, so they'll surely have something to say about what they want in next gen designs. Vega clearly wasn't developed only for Apple either, as parts of Vega design were first introduced in PS4 Pro (RPM) long before full blown Vegas.
This is also the first time anyone is claiming that Navi isn't GCN, I think?
Oh, and the claimed schedule has to be wrong, I doubt AMD would still show roadmap where there's 2 new architectures before end of 2020 if Navi was going to come in 2020 or even late 2019
 
Even if it does have some truth in it, I seriously doubt it's the whole truth, especially the future part.
AMD has previously taken input from both MS and Sony for GCN development and those changes have appeared in same generation, it's highly unlikely they'd suddenly start developing for example Navi only for Sony's needs (MS is going to launch next Xbox around the same time frame anyway, so they'll surely have something to say about what they want in next gen designs. Vega clearly wasn't developed only for Apple either, as parts of Vega design were first introduced in PS4 Pro (RPM) long before full blown Vegas.
This is also the first time anyone is claiming that Navi isn't GCN, I think?
Oh, and the claimed schedule has to be wrong, I doubt AMD would still show roadmap where there's 2 new architectures before end of 2020 if Navi was going to come in 2020 or even late 2019

Well, Forbes is saying something very similar:

https://www.forbes.com/sites/jasone...nys-playstation-5-vega-suffered/#12f01b4624fd
 
Status
Not open for further replies.
Back
Top