[VC]AMD Radeon R9 390X WCE Speculation Thread

[el etro]

I have a very much unofficial rumor claiming DX12.1 will be for Nvidia Maxwell and ahead only(along with another rumor saying all GCN is DX12.0 full compliant)...

...But as far as i know, DX12.1 will be like the DX11.1 and DX11.2 was, so it will not have so big acceptance on the market.

 

[csbin]

290X supports DX12.0, it doesnt support DX12.1.

AMD also list it as DX12 on their own site. AIBs list its as DX12.

The slide funny enough list the 290X as OpenGL 4.4, while it only supports OpenGL 4.3.

If it was at AIBs only and the 390X supported DX12.1, they would note it.

http://www.anandtech.com/show/7955/amd-posts-catalyst-144-release-candidate-drivers

As we briefly noted in our review of the R9 295X2, Catalyst 14.4 is from a new driver branch (14.100), and as is usually the case for new branches AMD has rolled in some new features with this driver. Of note, these drivers bring support for OpenGL 4.4 to AMD’s GPUs, which among other additions sees support for sparse textures (partially resident textures) and buffer storage objects for use with processors that have unified memory between the CPU and the GPU (Kaveri, etc).


AMD Catalyst 14.4 Brings Full OpenGL 4.4 Support To Linux
http://www.phoronix.com/scan.php?page=news_item&px=MTY2OTY

 

[el etro]

And I believe very much these rumors(about the R300 newer slides) are true.


Too much i believe Fiji and company will really be the GCN2 cards.

 

[ShintaiDK]

http://www.anandtech.com/show/7955/amd-posts-catalyst-144-release-candidate-drivers

As we briefly noted in our review of the R9 295X2, Catalyst 14.4 is from a new driver branch (14.100), and as is usually the case for new branches AMD has rolled in some new features with this driver. Of note, these drivers bring support for OpenGL 4.4 to AMD’s GPUs, which among other additions sees support for sparse textures (partially resident textures) and buffer storage objects for use with processors that have unified memory between the CPU and the GPU (Kaveri, etc).


AMD Catalyst 14.4 Brings Full OpenGL 4.4 Support To Linux
http://www.phoronix.com/scan.php?page=news_item&px=MTY2OTY

Yet AMDs own site and AIBs says 4.3 for OpenGL and DX12 for DirectX.

http://www.amd.com/en-us/products/graphics/desktop/r9#
http://www.sapphiretech.com/presentation/product/product_index.aspx?pid=2060&lid=1

And slides says DX11.2 and 4.4. The inconsistency...

 

[Gikaseixas]

Not a big deal Shintai, sites get updated all the time, maybe they forgot a detail or two

 

[el etro]

I feel this rumor\slides is legit. IMO is a PR counterattack against TitanX launch event.

 

[3DVagabond]

HBM only has 1GB per stack, the fact that AMD has to use dual-link interposing and even mentions HBM2 not needing dual-link interposing in the slide means that this is not an optimal solution but Nvidia forced their hand.

Is that what it means? lol

How about the resolution(s) they are promoting made it necessary? If it's true, of course.

 

[boozzer]

I feel this rumor\slides is legit. IMO is a PR counterattack against TitanX launch event.

it works if the leaked info is true :thumbsup:

 

[sontin]

If you don't understand it than here is an oversimplified version:
CR: this is an easy target because it is basically a rasterization form but with different interpolation. Most GPUs uses dedicated units for interpolation, but not GCN. The interpolation is already emulated in this architecture, so the only thing is needed is a new compiler unroll routine. Nothing more, nothing less. This is also possible with other architectures, but these are not designed for manual interpolation, so the other IHVs needs some new dedicated units. For example Maxwellv2 has new SFU blocks for this.

And yet AMD hasnt done it for opengl. There doesnt exit any documentation about CR with AMD hardware. So no, they dont support any form of CR through the hardware. And the interpolation still needs additional geometry and pixel shader operations to do CR.

ROVs: this is just a memory view ordering option. On GCN the GPU can wait for GDS changes preventing the further shader execution until it will be safe. AMD already support this in OpenGL with GL_INTEL_fragment_shader_ordering. ROVs will be the same.

The only documentation about the performance is from g-truct:

Without careful profiling, using this extension will be particularly slow on AMD hardware.

http://www.g-truc.net/doc/Candidate features for OpenGL 5.pdf

This is Intel's presentation from the GDC 2015:

Hardware in the ROP is responsible for enforcing this ordering requirement.

https://software.intel.com/sites/default/files/managed/4a/38/OIT-to-Volumetric-Shadow-Mapping.pdf

It is pretty useless for real time rendering. And i bet that Microsoft will force dedicated hardware support within the ROPs.

 

[raghu78]

Yep.

Another oddity is the memory speed. 1.25Ghz. Hynix still only sells 1Ghz. While 1.25Ghz could be in the works, why put it on the card besides for the halo effect.

1Ghz=512GB/sec. 1.25Ghz=640GB/sec To compare a 290X is 320GB/sec. And the 290X isnt exactly lacking bandwidth.

you realize dual link interposing is some kind of memory interleaving design. So its not similar to having 4GB at 512 GB/s. Here due to the sharing of a 1024 bit HBM controller by 2 HBM stacks the actual bandwidth will be lesser. So it makes sense to drive up the HBM chips to 1.25 Ghz to feed the GPU without bottlenecking when the entire 8 GB capacity is used.

 

[ThatBuzzkiller]

And yet AMD hasnt done it for opengl. There doesnt exit any documentation about CR with AMD hardware. So no, they dont support any form of CR through the hardware. And the interpolation still needs additional geometry and pixel shader operations to do CR.



The only documentation about the performance is from g-truct:


http://www.g-truc.net/doc/Candidate features for OpenGL 5.pdf

This is where conservative rasterization saves the day ...

With conservative rasterization, you can implement geometry binning at a per-tile basis which can help reduce the amount of resource contention going on ...

By doing this you can evaluate shading for each tile which means you can use the LDS instead of the GDS to perform programmable blending thus giving you something similar to GL_EXT_pixel_local_storage ...

If all else fails there's something else you can use called "linked lists" ...

 

[NostaSeronx]

you realize dual link interposing is some kind of memory interleaving design.

Can we just call it Clamshell?

 

[zlatan]

And yet AMD hasnt done it for opengl. There doesnt exit any documentation about CR with AMD hardware. So no, they dont support any form of CR through the hardware. And the interpolation still needs additional geometry and pixel shader operations to do CR.

Please read those documents to understand how does GCN works.

The only documentation about the performance is from g-truct:


http://www.g-truc.net/doc/Candidate features for OpenGL 5.pdf

This is Intel's presentation from the GDC 2015:

https://software.intel.com/sites/default/files/managed/4a/38/OIT-to-Volumetric-Shadow-Mapping.pdf

It is pretty useless for real time rendering. And i bet that Microsoft will force dedicated hardware support within the ROPs.

I don't understand this. We talk about UAV ordering here. This can't be implemented with pure dedicated hardware. Intel doing it almost the same way as AMD do. NVIDIA also.

 

[Cloudfire777]

With the recent performance chart from AMD I think Titan X and 390X will be extremely close in performance.

Since 390X TDP is 300W and Titan X TDP is 250W. With price being rumored $700 for R9 390 and "much higher" for R9 390X due to HBM, so perhaps both Titan X and 390X will cost $999. I think AMD will have the edge this time with watercooling and better 4K gaming performance than Titan X, because of HBM.

The only reason why AMD would lose sales, is if the cooling solution is noisy. But with cooler master that designed the cooling system, I cant see that happening.
Or price above $999, say $1200. But this is AMD. I cant see that happening either

As a Nvidia enthusiast, I can`t find a single reason to pick Titan X over 390X if everything above turns out true. Not a single one

 

[tviceman]

Agreed with cloudfire. Although personally I think my limit is <=$550 when it comes to a new card purchase, Nvidia clearly did a great job of raising the psychological price barrier for enthusiasts. If AMD is only 8-12 weeks behind GM200, and it can essentially outperform GM200 without the need of water cooling, AMD will have pulled a coup. Nvidia can take a page out of AMD's book and start bundling Titan with an AIO water solution edition and a little bit more vcore headroom.

 

[Subyman]

I was asking about this in the other thread, but why is memory speed showing on the chart as Gbps when it should be Ghz? It shows the 290X as being 5Gbps, but its supposed to say 5GHz.

 

[Cloudfire777]

I was asking about this in the other thread, but why is memory speed showing on the chart as Gbps when it should be Ghz? It shows the 290X as being 5Gbps, but its supposed to say 5GHz.

5Gbps/4 = 1250MHz

Its the 390X that should say GHz or 5Gbps

 

[Silverforce11]

As a Nvidia enthusiast, I can`t find a single reason to pick Titan X over 390X if everything above turns out true. Not a single one

1. 12GB vram
2. CUDA
3. NV ecosystem & "better drivers"
4. GameWorks
5. GPU PhysX

There's a few, surely it will apply for many NV enthusiast, even if 390X ends up faster, they won't switch. I saw a few posts on other forums, one guy even claims he will never ditch NV for AMD unless his NV GPU rapes his wife and burns down his house. That is what it would take for the switch. I had to lol.

 

[Subyman]

5Gbps/4 = 1250MHz

Its the 390X that should say GHz or 5Gbps

Yeah, my point is the chart is wrong, so it makes it questionable.

 

[Bobisuruncle54]

1. 12GB vram
2. CUDA
3. NV ecosystem & "better drivers"
4. GameWorks
5. GPU PhysX

There's a few, surely it will apply for many NV enthusiast, even if 390X ends up faster, they won't switch. I saw a few posts on other forums, one guy even claims he will never ditch NV for AMD unless his NV GPU rapes his wife and burns down his house. That is what it would take for the switch. I had to lol.

None of those reasons are worth it for a gamer. Brand loyalty is not a virtue of an intelligent consumer.

 

[nvgpu]

But a lot of people still bought AMD GCN products eventhough they were inferior, more power hungry, runs hot and slower than Nvidia products, that's clearly brand loyalty there.

 

[Bobisuruncle54]

But a lot of people still bought AMD GCN products eventhough they were inferior, more power hungry, runs hot and slower than Nvidia products, that's clearly brand loyalty there.

Not if they're cheaper for the performance.

 

[Skurge]

But a lot of people still bought AMD GCN products eventhough they were inferior, more power hungry, runs hot and slower than Nvidia products, that's clearly brand loyalty there.

You forgot to mention cheaper.

 

[lilltesaito]

But a lot of people still bought AMD GCN products eventhough they were inferior, more power hungry, runs hot and slower than Nvidia products, that's clearly brand loyalty there.

I think more people bought them for bitcoin over gaming.

 

[wand3r3r]

But a lot of people still bought AMD GCN products eventhough they were inferior, more power hungry, runs hot and slower than Nvidia products, that's clearly brand loyalty there.

It's just plain old brand agnostic consumers choosing value ($/performance). That wording isn't very brand agnostic.