[Speculation] What will AMD's upcomming 7nm GPUs look like?

[piesquared]

Lot's of possibilities with the jump 7nm is over 14nm. Will they increase compute resources to make them even more of a compute power house, or will they shrink the existing down to make it small with excellent efficiency? Or maybe both. We know AMD has diverted some engineers working in the CPU division to increase clocks, so if they have done that then it would be quite a high performing GPU. Really, over the last generations the main reason nvidia was ahead was due to higher clocks. If AMD manage to increase clocks on 7nm another few hundred MHz while taking advantage of the density improvements to add more cores..

 

[JDG1980]

If AMD manage to increase clocks on 7nm another few hundred MHz while taking advantage of the density improvements to add more cores..

More cores on GCN won't help. The underlying problem is that the architecture is limited to 4 triangles/clock no matter how many CUs there are. This means that once you get much above Hawaii levels (40-48 CUs), the card will be bottlenecked in many scenarios and the extra shaders will go to waste.

Higher clock speeds and more memory bandwidth *can* help. First-generation Vega had decent clock rates (though it was still running above its sweet spot), but was hamstrung by HBM clock rates being lower than expected; due to the reduction of HBM to two stacks, Vega actually had less memory bandwidth than Fiji. Any new version of Vega should benefit on this front, since HBM yields and clocks are now considerably better than they were in mid-2017.

 

[prtskg]

Since AMD is making Vega 20 especially for AI use, I think AMD will separate graphics and compute gpu from 7nm onward.

 

[ozzy702]

No clue, my crystal ball doesn't go 3 years into the future. LOL

 

[maddogmcgee]

Wasn't the main issue with clocks the Glofo process it was built on?

You also have to remember that higher clocks result in a disproportionate increase in power. While AMD is trying to create gpu's that do compute and gaming it is going to be very hard to overtake Nvidia at the high end.

Despite all this I am hoping AMD get some smaller, cheaper to make Navi parts out for the mid and low range.

 

[alcoholbob]

I'm guessing it'll end up being about 10% faster than a 1080 Ti if we are talking about rasterization ability given the die size, assuming no IPC gains.

 

[prtskg]

I'm guessing it'll end up being about 10% faster than a 1080 Ti if we are talking about rasterization ability given the die size, assuming no IPC gains.

Navi?

 

[Qwertilot]

Immediately it'll be a die shrunk Vega aimed at compute uses, which I wouldn't touch it with a barge pole for gaming.
(Not just performance, drivers etc.).

Then, much more speculatively, something like a rerun of polaris - mostly R&D for the next generation of consoles - which might be decent if & only if they can produce a really big jump in how efficient their architecture is.

With all the extra money coming in now from Ryzen and Epyc they might manage a proper comeback - top to bottom line up, somewhere near NV etc - after that. R&D takes time.

 

[Glo.]

There is some speculation that for Navi this patent is really important:
http://www.freepatentsonline.com/20180144435.pdf

 

[Glo.]

Immediately it'll be a die shrunk Vega aimed at compute uses, which I wouldn't touch it with a barge pole for gaming.
(Not just performance, drivers etc.).

Then, much more speculatively, something like a rerun of polaris - mostly R&D for the next generation of consoles - which might be decent if & only if they can produce a really big jump in how efficient their architecture is.

With all the extra money coming in now from Ryzen and Epyc they might manage a proper comeback - top to bottom line up, somewhere near NV etc - after that. R&D takes time.

Vega 20 - 326 mm2 die size, with 80 mTR/mm2. Over two times higher density of Vega 10, with over 25 bln transistors.

Vega 20 is NOT a die shrink of Vega 10, especially when it is GCN 5.1, and Vega 10 was GCN 5.0.

 

[Glo.]

Also, there is a strong rumor that Polaris 30 GPU is coming next month. 12 nm refresh.

 

[Mopetar]

Vega 20 is NOT a die shrink of Vega 10, especially when it is GCN 5.1, and Vega 10 was GCN 5.0.

It’s still GCN though. While that’s likely okay for compute (at least for what AMD charges), as JDG pointed out the architecture is beyond long in the tooth and they need something new if they hope to be competitive in the consumer space going forward.

NVidia won’t keep their jacked up prices forever and won’t have a problem taking low margins at the mid-level as long as AMD can’t touch the their high-margin top end.

I wouldn’t be surprised to see a replacement for Polaris that uses GDDR6 and has more cores than Polaris, but fewer than Vega. Essentially a good mid-range card that will compete against the 2060 and/or 3060. Almost something like a Vega 48, but with better clocks and a few other minor improvements.

 

[Glo.]

It’s still GCN though. While that’s likely okay for compute (at least for what AMD charges), as JDG pointed out the architecture is beyond long in the tooth and they need something new if they hope to be competitive in the consumer space going forward.

NVidia won’t keep their jacked up prices forever and won’t have a problem taking low margins at the mid-level as long as AMD can’t touch the their high-margin top end.

I wouldn’t be surprised to see a replacement for Polaris that uses GDDR6 and has more cores than Polaris, but fewer than Vega. Essentially a good mid-range card that will compete against the 2060 and/or 3060. Almost something like a Vega 48, but with better clocks and a few other minor improvements.

3072 GCN core chip with 256 bit GDDR6? It would fit perfectly in the rumored GTX 1070 levels of performance...

 

[Olikan]

There was some new instances for gpuopen drivers for navi... mostly related for color compression and tiled rendering

"Wave break" is hard, no real ideia of what it is

https://github.com/GPUOpen-Drivers/...77b/src/core/hw/gfxip/gfx9/settings_gfx9.json

Gfx9UseDcc - "Bitmask of cases where DCC (delta color compression) surfaces will be used":

New Option:

Gfx10UseDccUav - "Shader writable surfaces(UAV)"

----

Gfx10UseCompToSingle - "Whether we need to set first pixel of each block that corresponds to 1 byte in DCC memory, and don't need to do a fast clear eliminate based on image type."

Options:

Gfx10UseCompToSingleNone: "Use comp-to-reg all image type."
Gfx10UseCompToSingle2d: "Use comp-to-single for 2d image."
Gfx10UseCompToSingle2dArray: "Use comp-to-single for 2d array image."
Gfx10UseCompToSingleMsaa: "Use comp-to-single for MSAA image."
Gfx10UseCompToSingle3D: "Use comp-to-single for 3d image."
Gfx10DisableCompToReg: "If set, comp-to-reg rendering will be disabled for images that were cleared with comp-to-single."

----

SdmaPreferCompressedSource - "Affects tiled-to-tiled image copies on the GFX10 SDMA engine where both images are compressed Set to true to leave the source image in a compressed state, set to false to leave the dest image in a compressed state."

----

Gfx10ForceWaveBreakSize - "Forces the size of a wave break; over-rides client-specified value."

Options:

Gfx10ForceWaveBreakSizeNone: "No wave breaks by region."
Gfx10ForceWaveBreakSize8x8: "8x8 region size."
Gfx10ForceWaveBreakSize16x16: "16x16 region size."
Gfx10ForceWaveBreakSize32x32: "32x32 region size."
Gfx10ForceWaveBreakSizeClient: "Use client specified value."
Gfx10ForceWaveBreakSizeAuto: "Let PAL decide."

----

And then there are some GFX9 options referencing a PAL_BUILD_NAVI10_LITE in addition to a PAL_BUILD_GFX10.

 

[Kenmitch]

Only thing I can add to the thread is one way or another it'll be jaw dropping. Just not sure why at this time. What is a given is it'll spawn many threads, arguments, infractions, a couple of bands, and most likely move the metrics once again.

 

[AtenRa]

Only thing I can add to the thread is one way or another it'll be jaw dropping. Just not sure why at this time. What is a given is it'll spawn many threads, arguments, infractions, a couple of bands, and most likely move the metrics once again.

ehehehe

 

[piesquared]

It will be interesting to see if this GPU will get a Frontier Edition card but i'm thinking it will be a strictly Radeon Instinct compute card. They have said as much, but that doesn't preclude another card coming out after the compute card for the prosumer market I suppose.

 

[Mopetar]

Only thing I can add to the thread is one way or another it'll be jaw dropping. Just not sure why at this time. What is a given is it'll spawn many threads, arguments, infractions, a couple of bands, and most likely move the metrics once again.

That’s almost true if any card though.

 

[Kenmitch]

That’s almost true if any card though.

True, but it's amplified with AMD as the majority favors nvidia.

 

[Qwertilot]

Nah, just means people rant the other way.

Its exceedingly silly

 

[Kenmitch]

Nah, just means people rant the other way.

Its exceedingly silly

I guess it's what makes the comment sections of reviews more fun to read than skimming the result images.

 

[GodisanAtheist]

AMD definitely has the potential to release a monster of a card given that a lot of the architecture groundwork has been put in place with Vega (30-40% increases if they can sort tiled rendering and some other features out) and there is likely a solid 30-40% performance just hiding in remixing the GCN cluster ratios and "fixing" GCN to scale beyond 4 Graphics Processing Engines.

Clock/Power/Density increases from 7nm only amplify all that.

AMD are likely going to miss the RTX party so they will need to show up strong for the 4k120hz crowd to capture their niche.

That being said, AMD does not need another hype train. I fully expect a ~250-300mm2 max part from AMD to kick off the 7nm generation.

The only question is if it will perform like a 1080 or if it will perform like a 2080+...

 

[frozentundra123456]

True, but it's amplified with AMD as the majority favors nvidia.

Seriously? You must not have been reading these forums a couple of years ago when the hype train was in full runaway mode for the 470 and 480.

 

[AtenRa]

AMD definitely has the potential to release a monster of a card given that a lot of the architecture groundwork has been put in place with Vega (30-40% increases if they can sort tiled rendering and some other features out) and there is likely a solid 30-40% performance just hiding in remixing the GCN cluster ratios and "fixing" GCN to scale beyond 4 Graphics Processing Engines.

Clock/Power/Density increases from 7nm only amplify all that.

AMD are likely going to miss the RTX party so they will need to show up strong for the 4k120hz crowd to capture their niche.

That being said, AMD does not need another hype train. I fully expect a ~250-300mm2 max part from AMD to kick off the 7nm generation.

The only question is if it will perform like a 1080 or if it will perform like a 2080+...

RTX is a proprietary RayTracing Technology by NVIDIA, there is also the DXR RayTracing Technology from Microsoft in DX-12 that all cards can run.
Since AMDs VEGA architecture is compute oriented, it is highly reasonable to expect that next AMD cards will continue to be heavily Compute oriented and perform good in DXR.

https://blogs.msdn.microsoft.com/directx/2018/03/19/announcing-microsoft-directx-raytracing/

https://www.anandtech.com/show/12547/expanding-directx-12-microsoft-announces-directx-raytracing

Also check Futuremark DXR demo

https://www.youtube.com/watch?v=81E9yVU-KB8&feature=youtu.be

 

[Krteq]

RTX is just a NV's backend to Microsoft's DXR API, nothing more

In conjunction with Microsoft’s new DirectX Raytracing (DXR) API announcement, today NVIDIA is unveiling their RTX technology, providing ray tracing acceleration for Volta and later GPUs. Intended to enable real-time ray tracing for games and other applications, RTX is essentially NVIDIA's DXR backend implementation.

AnandTech - NVIDIA Announces RTX Technology: Real Time Ray Tracing Acceleration for Volta GPUs and Later

AMD is working on it's own DXR backend and it could be unveiled on December this year