Vega/Navi Rumors (Updated)

[tviceman]

What you are quoting is actually correct, but requires implementation of Primitive Shaders and Programmable Geometry pipeline into the game. Otherwise RX Vega will be just on GTX 1080 level.

Both those features can be implemented only in DX12 and Vulkan.

1) did you even read his quote? He said if Vega is 430-450mm2 then then nano version will be able to match Titan X pascal. Vega is bigger than 450mm2 and the nano version won't even be in the same ballpark as Titan X pascal, let alone a GTX 1080. RX Vega is going to end up slower than a GTX 1080, and only in a few select games here in there will it be noticeably faster (greater than 5%).

2) If you did read his quote, and you're still clinging to what you said, let me know so I can block you. I no longer wish to deal with people who peddle fantasies as reality.

 

[Glo.]

1) did you even read his quote? He said if Vega is 430-450mm2 then then nano version will be able to match Titan X pascal.
2) If you did read his quote, and you're still clinging to what you said let me know so I can block you.

Yes, that is exactly what I am referring to.

Let me explain this. Graphics Pipelines rely on geometry throughput. AMD GCN was always limited by this, compared to Nvidia. GCN was able to register 1 triangle each cycle per shader engine, and Nvidia CUDA architecture was able to register 1 triangle per SM. And Nvidia architecture was able to have more than 4 SM's. GTX 1080 Ti has 6 SM's, so it is able to register 6 triangles with each clock. And it is able to clock much higher than GCN4.

And Nvidia architectures had TileBased rasterization which improved maximum framerates, because it culled massive amounts of unused geometry.

GCN5 has all of those features. Primitive Shaders and Programmable Geometry Pipeline allow for GCN to register up to 10 triangles per 4 shader engines, increasing the geometry throughput possible to feed the cores in GCN5 architecture.

Here is very informative film about this feature:

The quote was correct. Putting Fiji on 1.6 GHz, even without any of GCN5 advancements, would make Vega tie with GTX 1080 Ti. Just architecture advancements, in software that gets used of those features make Vega PER CLOCK, and per core 40% faster than Fiji. So think that 150-175W Nano with 1.2 GHz, would be on par GTX 1080 Ti. But that requires using the primitive Shaders and Programmable Geometry Pipeline.

 

[Elixer]

"Random message".... right.

https://mobile.twitter.com/GFXChipTweeter?p=s

 

[Mopetar]

"Random message".... right.

https://mobile.twitter.com/GFXChipTweeter?p=s

Answer: Power Draw

 

[tamz_msc]

The biggest issue that AMD is faced with since Hawaii is that they're still stuck with 4 Tri/clock worth of geometry performance while NVIDIA has 6 Tri/clock since Maxwell, plus more due to tiled caching.

Geometry performance of 28nm GCN doesn't scale at all with increasing #of SPs. Fiji is 2X Tonga but still has the same geometry performance as the smaller chip:

Hawaii(290X) is the same. It's difficult to estimate the NVIDIA numbers due to tiled caching, but a lower bound of 6 Tri/clock is a reasonable estimate.

As you can see, Polaris improved in certain aspects quite drastically, but we'll have to see if Vega is a further improvement in this critical aspect.

I think it has been established that the 'over 2X' geometry improvement that AMD touted in its slides referred to primitive shaders, which is frankly misleading IMO.

Then there's this, from Ryan Smith:

Talking to AMD’s engineers, what especially surprised me is where the bulk of those transistors went; the single largest consumer of the additional 3.9B transistors was spent on designing the chip to clock much higher than Fiji. Vega 10 can reach 1.7GHz, whereas Fiji couldn’t do much more than 1.05GHz. Additional transistors are needed to add pipeline stages at various points or build in latency hiding mechanisms, as electrons can only move so far on a single clock cycle; this is something we’ve seen in NVIDIA’s Pascal, not to mention countless CPU designs. Still, what it means is that those 3.9B transistors are serving a very important performance purpose: allowing AMD to clock the card high enough to see significant performance gains over Fiji.

So Vega is supposed to have tons of fancy new features, but the largest part of the increased transistor budget offered by the new process node was spent on increasing clocks? This sounds ominously like Intel's failed Northwood-Prescott experiment.

 

[mohit9206]

1.Will Navi be a new architecture or will it also be based on GCN?
2. Will Navi also include sub $250 GPUs or is it only hi end like Vega?
3.When is the earliest expected release date?
4.Any engineering samples so far?

 

[Glo.]

The biggest issue that AMD is faced with since Hawaii is that they're still stuck with 4 Tri/clock worth of geometry performance while NVIDIA has 6 Tri/clock since Maxwell, plus more due to tiled caching.

Geometry performance of 28nm GCN doesn't scale at all with increasing #of SPs. Fiji is 2X Tonga but still has the same geometry performance as the smaller chip:

Hawaii(290X) is the same. It's difficult to estimate the NVIDIA numbers due to tiled caching, but a lower bound of 6 Tri/clock is a reasonable estimate.

As you can see, Polaris improved in certain aspects quite drastically, but we'll have to see if Vega is a further improvement in this critical aspect.

I think it has been established that the 'over 2X' geometry improvement that AMD touted in its slides referred to primitive shaders, which is frankly misleading IMO.

Then there's this, from Ryan Smith:

So Vega is supposed to have tons of fancy new features, but the largest part of the increased transistor budget offered by the new process node was spent on increasing clocks? This sounds ominously like Intel's failed Northwood-Prescott experiment.

Programmable Geometry Pipeline alongside with Primitive Shaders are increasing this number of Geometry Throughput to 10 Triangles per clock, but requires developers to implement this feature in games.

How big impact do you guys think this change will have on Vega performance?

Short version. Massive. Vega has very explicit culling techniques: Draw Stream Binning Rasterizer, and Primitive Shaders. Rasterizer is culling things that do not have to be rasterized, because they are not displayed, and Primitive Shaders are culling geometry which has not been displayed. And we have to remember that Polaris brought us Primitive Discard Accelerator, which also does not need Developer implementation, and already is culling triangles early in the pipeline.

Add to this that Programmable Geometry Pipeline is increasing lets say even for the sake of argument that we will get 8 triangles per clock with 4 shader engines. It is already two times higher value, and its affect on performance will be massive, and you have to bare in mind that core clock of the GPU has massive effect on fillrate performance. That is why I have said, that with proper implementation of software Vega will be 30% faster than GTX 1080 Ti.

Do not underestimate Vega performance based on initial benchmarks. It has all of the tricks that Nvidia hardware has, but also adds something more. It is inconvenience for developers, but the benefits are huge.

 

[Glo.]

1.Will Navi be a new architecture or will it also be based on GCN?
2. Will Navi also include sub $250 GPUs or is it only hi end like Vega?
3.When is the earliest expected release date?
4.Any engineering samples so far?

Navi will be just slightly updated GCN5 architecture. Vega is foundation for all next generation GCN architectures.

 

[mohit9206]

Navi will be just slightly updated GCN5 architecture. Vega is foundation for all next generation GCN architectures.

I'll take this as a bad news.

 

[Krteq]

Huh, why?

 

[mohit9206]

Huh, why?

Because its time to move on from GCN. It cannot match Pascal on efficiency or performance, then why is amd still deciding to stick with this inefficient architecture?
GCN is a proven dead end.
A brand new highly efficient architecture is needed to compete with Nvidia.
Dragging the corpse of GCN will not give amd the results they wish for.

 

[Krteq]

GCN is great, quite robust and universal uarch, it's not "dead end" at all.

Yes, there are some issues with power efficiency in latest iteration, but those issues can be solved in future.

 

[mohit9206]

GCN is great, quite robust and universal uarch, it's not "dead end" at all.

Yes, there are some issues with power efficiency in latest iteration, but those issues can be solved in future.

If that's the case why don't all companies keep using the same architecture.
Why Nvidia keeps making new architecture when they could of just stuck with Kepler and kept adding new features to it?

 

[Glo.]

Because its time to move on from GCN. It cannot match Pascal on efficiency or performance, then why is amd still deciding to stick with this inefficient architecture?
GCN is a proven dead end.
A brand new highly efficient architecture is needed to compete with Nvidia.
Dragging the corpse of GCN will not give amd the results they wish for.

Is this educated opinion, or not educated opinion?

 

[Glo.]

If that's the case why don't all companies keep using the same architecture.
Why Nvidia keeps making new architecture when they could of just stuck with Kepler and kept adding new features to it?

This is in fact what they are doing. Only changes on architecture layout is the core throughput.

Kepler was fed by 256 KB Register File size for each 192 cores.
Maxwell was fed by 256 KB RFS for each 128 cores.
Pascal was fed by 256 KB RFS for each 128 cores(Pascal is actually Maxwell architecture on 16 nm process).

Rest of architecture is the same as was with Kepler. Only what changed in meaningful way is the way cores are fed by RFS.

 

[Krteq]

If that's the case why don't all companies keep using the same architecture.
Why Nvidia keeps making new architecture when they could of just stuck with Kepler and kept adding new features to it?

New architectures? Cuda cores hasn't changed much since Fermi. Actually, almost all changes were in SM layout, caches, front/back-end... in fact, same changes in generation leaps for GCN.

//Glo was faster

 

[SirDinadan]

NVIDIA is doing a great job with it's naming scheme. It doesn't have an "official" name for their main microarchitecture / instruction set. Just dubbing every iteration with a scientist is a huge win. People with no insight may arrive at the conclusion that every new microarchitecture from NVIDIA is something brand new.

On the other hand, AMD's GCN moniker is working just the opposite way - again GCN, it's old, refinements only, AMD should ditch this train wreck already.

 

[Tup3x]

Programmable Geometry Pipeline alongside with Primitive Shaders are increasing this number of Geometry Throughput to 10 Triangles per clock, but requires developers to implement this feature in games.

How big impact do you guys think this change will have on Vega performance?

Short version. Massive. Vega has very explicit culling techniques: Draw Stream Binning Rasterizer, and Primitive Shaders. Rasterizer is culling things that do not have to be rasterized, because they are not displayed, and Primitive Shaders are culling geometry which has not been displayed. And we have to remember that Polaris brought us Primitive Discard Accelerator, which also does not need Developer implementation, and already is culling triangles early in the pipeline.

Add to this that Programmable Geometry Pipeline is increasing lets say even for the sake of argument that we will get 8 triangles per clock with 4 shader engines. It is already two times higher value, and its affect on performance will be massive, and you have to bare in mind that core clock of the GPU has massive effect on fillrate performance. That is why I have said, that with proper implementation of software Vega will be 30% faster than GTX 1080 Ti.

Do not underestimate Vega performance based on initial benchmarks. It has all of the tricks that Nvidia hardware has, but also adds something more. It is inconvenience for developers, but the benefits are huge.

If this and then that and if the developers do this and that and the drivers enable this and that and optimisations and this that... then 150W power starved Vega Nano beats GTX 1080 Ti. Maybe, in theory. Because Vega is so awesome alien tech architecture that even RTG's driver team doesn't know how to produce a driver that gives even 50 % of the potential at launch.

Did I get it right?

 

[Glo.]

If this and then that and if the developers do this and that and the drivers enable this and that and optimisations and this that... then 150W power starved Vega Nano beats GTX 1080 Ti. Maybe, in theory. Because Vega is so awesome alien tech architecture that even RTG's driver team doesn't know how to produce a driver that gives even 50 % of the potential at launch.

Did I get it right?

No.

Read something about the architecture of GCN5 and how it compares to GCN4 and 3. This type of post shows your competence level, or lack of it. Im interested in discussion, not refutes like this.

I already provided a lot of points of interest for people who are truly interested in understanding what is happening.

I will ask for you a simple question. How much faster will be Vega with implementation of Programmable Geometry Pipeline in games, when it will double the geometry performance? If you are able to answer this question, we can discuss more.

 

[mohit9206]

No.

Read something about the architecture of GCN5 and how it compares to GCN4 and 3. This type of post shows your competence level, or lack of it. Im interested in discussion, not refutes like this.

I already provided a lot of points of interest for people who are truly interested in understanding what is happening.

I will ask for you a simple question. How much faster will be Vega with implementation of Programmable Geometry Pipeline in games, when it will double the geometry performance? If you are able to answer this question, we can discuss more.

Aka magic drivers?
First wait for Vega and now wait for the implementation of programmable geometry pipeline? How long does amd expect people to keep waiting?
RX470/570 are great cards and Vega 56 probably has some chance of being good option also but day 1 performance is important as well.

 

[Glo.]

Aka magic drivers?
First wait for Vega and now wait for the implementation of programmable geometry pipeline? How long does amd expect people to keep waiting?
RX470/570 are great cards and Vega 56 probably has some chance of being good option also but day 1 performance is important as well.

Performance of Vega in current state has little to do with drivers.

Drivers only expose what is available on hardware to the application, if it is designed to use those features. One of reasons for example why Titan Xp got big boost in SpecPerf application with latest driver update was because Nvidia stopper artificially gimping its performance by disabling Antialiased Lines feature in non Quadro hardware. The applications are designed to use this feature, but its drivers job to show it to the application.

Current software is not designed with Programmable Geometry Pipeline and Primitive Shaders in mind.

 

[raghu78]

Vega is a EPYC disaster . Competes with GTX 1080 for almost 300w TDP. Only AMD can fail so hard. After 10 years AMD have a true successor to HD 2900XT. Luckily for AMD Zen turned out to be a lifesaver.

https://videocardz.com/71771/final-radeon-rx-vega-64-3dmark-performance

 

[zlatan]

If that's the case why don't all companies keep using the same architecture.
Why Nvidia keeps making new architecture when they could of just stuck with Kepler and kept adding new features to it?

Pascal is basically a Kepler with new features.

 

[CatMerc]

If that's the case why don't all companies keep using the same architecture.
Why Nvidia keeps making new architecture when they could of just stuck with Kepler and kept adding new features to it?

That IS what they're doing. A different name doesn't change the fact that it's iterative.

There hasn't been a clean sheet design from NVIDIA since Tesla. Same for AMD and GCN.

 

[mohit9206]

Vega is a EPYC disaster . Competes with GTX 1080 for almost 300w TDP. Only AMD can fail so hard. After 10 years AMD have a true successor to HD 2900XT. Luckily for AMD Zen turned out to be a lifesaver.

https://videocardz.com/71771/final-radeon-rx-vega-64-3dmark-performance

Synthetic benchmarks do not always translate to better gaming performance.