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

Status
Not open for further replies.
I highly suspect we will see something in the next quarter, if they want to release something in mid-2019.
Reading into the last update given last week at the Deutsche Bank telco, it seems the Navi GPUs will be late. The roadmap given says 2018 for Vega 20, early 2019 7nm EPYC aka Rome and after that they will push the 7nm desktop CPUs/APUs. After all this is out of the house, they will release Navi. This doesn't seem to be a mid-2019 release TBH.
 
there is talk navi not being gcn based.
https://wccftech.com/exclusive-amd-navi-gpu-roadmap-cost-zen/
whether their right or wrong is another matter, i wouldn't bet house their right personally.
If it is the case then what is 'next gen', that sounds like the one that shouldn't gcn based to me.
I have no clue where WCCFTech gets their ideas, but AMD hasn't said or shown anything suggesting Navi wouldn't be GCN

Wasn't Navi supposed to be the "Next-Gen" arch, completely different from GCN? I seem to remember so from slides back in early-mid 2016. Did something change?
I know my memory isn't perfect, but I follow these closely (even for work) and can't remember seeing any such slides ever, nor did I find any suggesting such either.
You sure you're not mixing it with the one that says "Scalability" and "Nexgen memory" on it?
 
Wasn't Navi supposed to be the "Next-Gen" arch, completely different from GCN? I seem to remember so from slides back in early-mid 2016. Did something change?

In several presentation slides, AMD presented Navi as being before "Next Gen", so I'm guessing it's the final gasp of GCN before they bring out the successor, probably in 2020/21. As shown here -

gpuroadmap_575px.png


07054015428l.jpg
 
I find it very hard to believe AMD isn't working on a ~1/2 size midrange 7nm chip as well as that giant monster.

They almost certainly are.

There are a couple relatively straightforward points that suggest that Navi 10 is a ~64CU ~150W 256-bit GDDR6 GPU and Navi 11 is roughly half of Navi 10.
  • Polaris 10 & 11 will be three years old next year and they both need to be replaced, especially Polaris 11 for use on mobile.
    • AMD doesn't have the R&D to do a full lineup refresh like Nvidia can. They use GPUs for about 3 years before replacing them.
  • GCN still has a 4 shader engine limit as of Vega, so it's just 64CU GPUs until then. Luckily, a 7nm "shrink" of Vega 10 would pretty much perfectly fit into 150W (roughly half of 300W), which is frighteningly convenient as that's Polaris 10's target.
    • Of course, AMD could change GCN to go past 4 shader engines, but they have their reasons for avoiding that for as long as possible. Anandtech inquired about this:
    Talking to AMD’s engineers about the matter, they haven’t taken any steps with Vega to change this. They have made it clear that 4 compute engines is not a fundamental limitation – they know how to build a design with more engines – however to do so would require additional work. In other words, the usual engineering trade-offs apply, with AMD’s engineers focusing on addressing things like HBCC and rasterization as opposed to doing the replumbing necessary for additional compute engines in Vega 10.
  • AMD loves selling 256-bit GDDR GPUs at the $200-300 space and the bandwidth math just works out perfectly for GDDR6.
    • 14Gbps GDDR6 is already seen in the new Turing GPUs. On a 256-bit bus, it would provide 448 GB/s, which is pretty close to the 484 GB/s that the 64CU Vega 64 needs. You assume some minor efficiency increases in necessary bandwidth (and/or slightly fewer than 64CUs) and you're perfect.
  • Finally, yields. 7nm yields initially won't be high enough for a large consumer die.
    • AMD survived in 2016 without a high end consumer GPU when they transitioned to 14nm. Why should we expect the 7nm transition to be any different?
 
Compared to 2080ti the frontend of Vega is not bad. I will get the same Drawcall Limits like the 2080 ti in the API Overhead test.
 
The narrative that AMD was caught by surprise is ridiculous. Will they implement something similar to Nvidia's RT cores (which technically aren't even "cores"..) ?
The described functionality of the RT cores is along the lines of an offloading co-processor or semi-independent sequencer. It's not entirely clear how they are physically integrated relative to the SIMD units or other blocks, but their behavior sounds closer to a core than the SIMD lanes, and might be nearer the polymorph or texturing units.
In that regard, perhaps AMD could benefit from exposing to the outside world a new sequencer domain in their SIMDs, like the texturing block which does a fair amount of computation on its own and runs on the far side of a set of data and command buses.
The scalar unit as we know it is linked more tightly to the scheduling and flow control elements that likely had something similar in hardware pre-GCN, just not exposed to the programmer.
Adding more such resources may allow more concurrency in non-vector work and take advantage of the elements in the pipeline that do things like coalesce accesses and handle long-latency fetch loops, while adding a level of programmability the RT cores lack, although this doesn't guarantee the compactness in hardware and may not play well with the much more confined caches in current GCN.


Saying they have a vision of cloud computing and are working with Microsoft is corporate-speak. It doesn't indicate the scope of the work, what part of the console space it encompasses, or how much has actually been committed to.
From https://seekingalpha.com/article/42...18-deutsche-bank-technology-conference?page=5 (some transcription irregularities, but the sentences in bold seem clear):

Devinder Kumar

We like the semicustom model a lot. Semicustom model is one of those; as you observe the game consoles, you win the designs; some of the engine and the expenses gets depraved by the input from the customers; we go ahead and get the chip out; and after that, it’s a mutually exclusive deal where you can predict revenue. Going back to 2012, 2013 timeframes, we’ve had predictably somewhere between $1.5 billion to $2 billion of revenues coming from the game console business, both Sony and Microsoft and that has allowed us to invest in exactly the roadmap that is delivering right now. We like that business a lot. We are competing for the next generation product. But Sony and Microsoft have to make their decisions and then taken we'll take it from there. But we like it a lot from an overall standpoint

In both cases, it's an executive that most likely does not want to speak for a partner, but also means they are not claiming that the consoles have been committed to be AMD hardware.
Competing for a contract and working together with the console makers are also not exclusive. Projects of this scope can have a lot of cooperative work between a hardware vendor and the platform holders, and that cooperation in setting down a candidate design could broadly count as "working with" Microsoft or Sony even if it is rejected. AMD's EPYC processors might be part of the cloud infrastructure for a console that is all or in part not AMD, and a worse scenario is working with Sony or Microsoft on a framework for backward/forward compatibility with a different vendor's CPU and/or GPU.

Hasn't AMD mentioned that Navi was going to be their first major GFX revision since GCN was introduced? IE - marking a significant move away from the derivative GCN architectures that we've seen since the 7970?
Possibly yes and no? Perhaps the most significant ISA change came with GCN3, where VI changed its encoding significantly for scalar memory writes, as well as data-parallel operations and sub-word addressing. Also, AMD touted Vega as the biggest jump ever, and probably had similar marketing of other transitions.


And the most anticipated features not working :eek:
One idea that came to mind is that with Nvidia's task and mesh shaders, both vendors have now offered a re-tooling of the geometry front end. There is some overlap of scope, and some of their decisions probably align because they are facing similar challenges. However, I think they also diverge in various parts in ways that may conflict. One of the vendors has gone ahead and committed to offering its new shaders with a more clear API extension for its methods, and the other vendor has for some reason let things drop. The possibility exists that even a functional NGG might be threatened if the design target it had was replaced due to outside factors.

Compared to 2080ti the frontend of Vega is not bad. I will get the same Drawcall Limits like the 2080 ti in the API Overhead test.
If Nvidia gets its way, however, there's going to be a path available for a significant fraction of those calls that would bypass that bottleneck.
 
In several presentation slides, AMD presented Navi as being before "Next Gen", so I'm guessing it's the final gasp of GCN before they bring out the successor, probably in 2020/21. As shown here -

gpuroadmap_575px.png


07054015428l.jpg

https://wccftech.com/amd-new-major-gpu-architecture-to-succeed-gcn-by-2020-2021/
AMD’s GCN architecture has proven to be one of the most successful macro architectures of its time debuting with the Southern Islands micro-architecture (1st Generation GCN) and lives on,
currently, with the Vega micro-architecture (5th Generation GCN). Navi, the upcoming next generation uArch from AMD in 2019 will constitute the 6th Generation GCN and will be the last and final breath of this overarching GCN series.

Navi will be the last GCN-based architecture, will be succeeded by brand new macro-architecture in 2020/2021 timeframe, execution of GCN successor being led by new RTG leadership
 
Vega 20 will be oonly a pro/research card, no ?
Correct.

However it wouldn't be ideal to make it immediately outdated. So the question is whether AMD pulls another Polaris - a "quick" architecture update for the sub-$300 market - or if they want to invest more heavily in a new top-to-bottom stack and/or major new features.

I could easily see them going either direction, so AMD is a very hard read right now. Especially as Vega Mobile remains AWOL.
 
So the question is whether AMD pulls another Polaris - a "quick" architecture update for the sub-$300 market - or if they want to invest more heavily in a new top-to-bottom stack and/or major new features.

I could easily see them going either direction, so AMD is a very hard read right now. Especially as Vega Mobile remains AWOL.
Yes, where is the Vega Mobile? It's been almost 10 months since Dr. Su held the chip on stage.

One might think it was the original "Polaris replacement" - Vega 11 which got reevaluated at some point. Then it was probably tweaked towards mobile and then again reevaluated and killed? The last hope are notebooks slated for the Xmas season.
 
Yes, where is the Vega Mobile? It's been almost 10 months since Dr. Su held the chip on stage.

One might think it was the original "Polaris replacement" - Vega 11 which got reevaluated at some point. Then it was probably tweaked towards mobile and then again reevaluated and killed? The last hope are notebooks slated for the Xmas season.
I think Vega mobile is just the least financially attractive part for AMD and given that GF is capacity constrained already it doesn't make sense to produce it, especially if this China SOC in manufactured at GF. Could they sell vega mobile for much more then Polaris, even enough to offset the interposer and HBM?
 
I highly suspect we will see something in the next quarter, if they want to release something in mid-2019.
I just hope AMD manages to keep their mouths shut until Navi actually releases, whenever that is.
80% of the "disappointment" surrounding Vega came from all the stuff they blabbed about during those 6 months before release.


I'm only pretty sure the previous PS5 rumors are totally bogus.
Which ones?



However it wouldn't be ideal to make it immediately outdated. So the question is whether AMD pulls another Polaris - a "quick" architecture update for the sub-$300 market - or if they want to invest more heavily in a new top-to-bottom stack and/or major new features.

I could easily see them going either direction, so AMD is a very hard read right now. Especially as Vega Mobile remains AWOL.

Vega 20 is a 1/2 FP64 throughput GPU, so Navi wouldn't completely replace it unless it has the same functionality.
After all, AMD still has the 5 year-old Hawaii for the FP64 HPC market, despite the GPU being outdated as "consumer flagship" since Fiji 4 years ago.
 
Last edited by a moderator:
To me it looks like Navi was supposed to come out sometime during 2018, possibly early 2018, since Polaris came out in mid 2016 and that is reflected by the slide.
 
To me it looks like Navi was supposed to come out sometime during 2018, possibly early 2018, since Polaris came out in mid 2016 and that is reflected by the slide.

Using Polaris as a reference point, Vega was due early 2017 and Navi a year later, early 2018, provided that these are supposed to be to scale.
 
Using Polaris as a reference point, Vega was due early 2017 and Navi a year later, early 2018, provided that these are supposed to be to scale.
Yes, this seems plausible. The Greenland chip aka Vega 10 was mentioned numerous times along with the Polaris chips. Thus Vega should have originally launched much early.
 
Status
Not open for further replies.
Back
Top