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

[GodisanAtheist]

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

- I know, but thank you for the write-up nevertheless.(Edit: rereading this it sounds sarcastic, that isn't my intent).

"Missing the RTX train" was intended to convey that AMD likely will show up with the same general purpose compute units rather than any sort of specialized ray tracing hardware which is going to be less performant than specialized ray tracing hardware NV included in Turing.

As such, that isn't the hill AMD should die on and should instead focus on the traditional HEDT 4k120hz market.

Recall AMD getting slaughtered during the early tessellation craze because NV over engineered their tessellation hardware then hammered AMD on that front.

 

[Despoiler]

- I know, but thank you for the write-up nevertheless.(Edit: rereading this it sounds sarcastic, that isn't my intent).

"Missing the RTX train" was intended to convey that AMD likely will show up with the same general purpose compute units rather than any sort of specialized ray tracing hardware which is going to be less performant than specialized ray tracing hardware NV included in Turing.

As such, that isn't the hill AMD should die on and should instead focus on the traditional HEDT 4k120hz market.

Recall AMD getting slaughtered during the early tessellation craze because NV over engineered their tessellation hardware then hammered AMD on that front.

Compute focused raytracing APIs and AMD's compute oriented GPUs will probably work out well enough for ray tracing. I found some chatter over the weekend of AMD's response to RTX possibly being to leverage the bountiful amount of cores in Ryzen, and now Intel chips, to help speed up the raytracing process. If that is the case we all may be running Threadripper levels of core counts in the nearer future than people might suspect. Meanwhile Nvidia will be adding specialized cores to try to hit those benchmark marketing points.

As far as AMD 7nm, I think going to TSMC will be a boon. GF was never that good for high power or it took them awhile to refine the process to work better. AMD should tackle the power use which is part the process and part their architecture. They need more fine grain control in power. Some of that is down to board design. AMD cards themselves don't know how much total power is getting used and by what because of where they measure power. Really low hanging fruit. If they can measure better they can control better. It's possible we will see the Project Scorpio DX12 invocations added. It's possible they could expand the concept to include Vulkan. If there is a reason to free up CPU horsepower on PC, and if they are leveraging CPU cores for raytracing, it could be a big edge. It might be easier/cheaper just to push higher core counts on PC though. We know AMD will have NGG fully exposed in the 7nm chips. That should get them some better throughput in some of their weaker areas.

 

[crisium]

What they should do:

6x shader engines
4224 shaders
96 ROPs
Generous memory bus (at least 384-bit for GDDR6 or 3072-bit HBM2)

(credit to Head1985)

Hawaii still has better performance-per-flop than Polaris or Vega because the 2816:64:512-bit ratio gives the shaders enough to work with while not exceeding 4x shader engine limitations. They need to scale it by 50% including the shader engines. With high clock rates from 7nm and the IPC improvements of Polaris+, it should compete with TU102 just fine.

 

[Glo.]

What they should do:

6x shader engines
4224 shaders
96 ROPs
Generous memory bus (at least 384-bit for GDDR6 or 3072-bit HBM2)

(credit to Head1985)

Hawaii still has better performance-per-flop than Polaris or Vega because the 2816:64:512-bit ratio gives the shaders enough to work with while not exceeding 4x shader engine limitations. They need to scale it by 50% including the shader engines. With high clock rates from 7nm and the IPC improvements of Polaris+, it should compete with TU102 just fine.

They do not have to increase amount of cores/ROPs, etc, if they can find a way to feed the cores with work.

http://www.freepatentsonline.com/20180144435.pdf

This patented by AMD tech is one of implementations of it. This tech saves memory bandwidth, reduces power consumption, and allows AMD to increase core clocks.

And is DIRECTLY connected to this patent, from similar time:
http://www.freepatentsonline.com/20180121386.pdf

 

[Head1985]

And why not both?4x shader engines just dont work.With 6x shader engines they gain +50% geometry performance and +50% pixel fillrate and afr better shader utilization.And what if they add 512bit like with hawaii?
4224sp
96rops
6xhader engines
all is +50% vs hawaii
with 512bit and GDDR6 14Ghz it will have 896gb/s bandwidth

But i think navi will be small GPU(still GCN and 256bit) around GTX 1080 performance competing with GTX2060.

 

[Glo.]

And why not both?4x shader engines just dont work.With 6x shader engines they gain +50% geometry performance and +50% pixel fillrate and afr better shader utilization.And what if they add 512bit like with hawaii?
4224sp
96rops
6xhader engines
all is +50% vs hawaii
with 512bit and GDDR6 14Ghz it will have 896gb/s bandwidth

But i think navy will be small GPU(still GCN and 256bit) around GTX 1080 performance competing with GTX2060.

http://www.freepatentsonline.com/20180121386.pdf
Quote from the patent:

A super single instruction, multiple data (SIMD) computing structure and a method of executing instructions in the super-SIMD is disclosed. The super-SIMD structure is capable of executing more than one instruction from a single or multiple thread and includes a plurality of vector general purpose registers (VGPRs), a first arithmetic logic unit (ALU), the first ALU coupled to the plurality of VGPRs, a second ALU, the second ALU coupled to the plurality of VGPRs, and a destination cache (Do$) that is coupled via bypass and forwarding logic to the first ALU, the second ALU and receiving an output of the first ALU and the second ALU. The Do$ holds multiple instructions results to extend an operand by-pass network to save read and write transactions power. A compute unit (CU) and a small CU including a plurality of super-SIMDs are also disclosed.

This patent allows AMD to schedule more job with each cycle, because the amount of instructions, are smaller in size, and more easy for the GPU execute, and fill the gaps in workload distribution more efficiently. It saves a lot of Register File Size, memory bandwidth and power, however to fully function it requires second bit:
http://www.freepatentsonline.com/20180144435.pdf
Direct quote from the patent:

The GPU coprocessor includes an inter-lane crossbar and intra-lane biased indexing mechanism for a vector general purpose register (VGPR) file. The VGPR file is split into two files. The first VGPR file is a larger register file with one read port and one write port. The second VGPR file is a smaller register file with multiple read ports and one write port. The second VGPR introduces the ability to co-issue more than one instruction per clock cycle.

Both of those patents actually give you pretty nice look at what actually was the bottleneck in graphics pipeline of AMD GPUs, apart from Geometry performance.

If this is bases of Navi, it will be interesting to see how it affects performance. At the very least GPUs should be more efficient on architectural level.

And remember guys. 4 Shader Engine, 2304 GCN core chip, with high enough core clocks(1.9 - 2.1 GHz) will be in the ranges of GTX 1080. And with 80 mTR/mm2 from TSMC 7 nm HPC 100 mm2 die GPU will have 8 Bln transistors. Almost 50% more than Polaris 10 had with the same core count. AMD will have a ton of transistors to burn on the silicon design.

Balancing 4096 GCN core chip still can provide amazing effects, if it can be scheduled efficiently.

I am actually baffled that nobody is concerned about RTX 2070 performance which will register only 3 triangles with each clock, because it has only 3 GP Clusters, compared to TU104 which has 4, and to TU2 which has 6. Why it won't have problems? Because Nvidia GPUs already have the tech AMD patented recently since Maxwell architecture.

 

[Head1985]

2070 have 3x gpc just like 1070.2070 have biggest buff of all cards.Fe or oc version will be like 10% slower than 1080TI FE.But AIB 1080TI will crush it and will be close to 20% faster.
Btw 6x shader engines will be always better than some patents.50% geometry performance and pixel fillrate will work in every game no matter what.They need to do that no matter what(atleast for big SKU)
Its like Nv would use 64rops and 4x GPC on 1080TI and 2080TI.

But i think navi will be small GPU something like rx480...it will compete with gtx1080/2060.

 

[Glo.]

Both Navi GPUs are small. The question is: which one is launching first: "small" or "smaller".

Smaller one will have somewhere between 2048 and 2304 GCN cores, with 128 bit GDDR6 memory bus. Bigger one will have 4096 GCN cores, with 256 Bit GDDR6 memory bus.

I wonder how big there is a chance for 6 shader engines, with 12 CU's in each SE.

If you will come to think about it: Those two patents I pointed out may actually allow AMD to not make dramatic redesign to structure of the GCN scheduling BUT at the same time allow them to get out of the hard 4 SE limit? Because those two patents are to increase the amount of work scheduled on smaller scale - properly feeding the cores.

Navi is supposed to be from PS5 right? PS4 had 1152 GCN cores. PS4 Pro - 2304 cores. PS5 could have perfectly straight again doubling the amount of cores: 4608. Which would fit perfectly with those 6 shader Engines, and scalability of the architecture.

But that is just my hypothesis.

 

[GodisanAtheist]

Compute focused raytracing APIs and AMD's compute oriented GPUs will probably work out well enough for ray tracing. I found some chatter over the weekend of AMD's response to RTX possibly being to leverage the bountiful amount of cores in Ryzen, and now Intel chips, to help speed up the raytracing process. If that is the case we all may be running Threadripper levels of core counts in the nearer future than people might suspect. Meanwhile Nvidia will be adding specialized cores to try to hit those benchmark marketing points.

- I hope their gameplan doesn't revolve around CPU core counts... if you thought Nvidia pushing specialized hardware on their cards was bad, imagine needing a whole new PC! Maybe they can push CPU game physics as well while they're at it

I would be amazed if compute based raytracing came close to the solution NV implemented. NV is no slouch in GPU compute tasks (Volta, at the very least) and if the solution to real time ray tracing was as simple as have gpu compute handle the problem they could have "easily" built a monster 6K shader core chip to just brute force it in the space they used for specialized RT/TN hardware and still had nothing to sweat from AMD and done a better job with the launch day titles.

If the thought goes "they put together all this specialized hardware to push for vendor lock-in" they could have accomplished much the same thing with an RTX software ecosystem built on CUDA with a deliberately gimped OpenCL/DXR path to make their solution seem like the better one. Would cost less at the same time.

Regarding core counts, I feel like AMD needs to move to 6 engines or mix up its core count if for no other reason than another 4096 shader card is going to absolutely break the souls of whatever diehard AMD fans are left after the disappointment that number has elicited on its last two outings.

AMD hasn't been playing the same game NV has in the dgpu space since Fury X and I don't expect that to change with 7nm as much as I wish it would (not to say the moves AMD is making aren't the right business moves, but it can leave an enthusiast of the tech feeling out in the cold).

 

[beginner99]

- I hope their gameplan doesn't revolve around CPU core counts... if you thought Nvidia pushing specialized hardware on their cards was bad, imagine needing a whole new PC! Maybe they can push CPU game physics as well while they're at it

Why not? With the different AVX instructions and high core counts currently going unused, I would actually like that.

 

[zlatan]

Both Navi GPUs are small. The question is: which one is launching first: "small" or "smaller".

Smaller one will have somewhere between 2048 and 2304 GCN cores, with 128 bit GDDR6 memory bus. Bigger one will have 4096 GCN cores, with 256 Bit GDDR6 memory bus.

I'm pretty sure that Navi don't use GDDR6. The minimum bandwith will be 1 TB/s, and it is not possible with GDDR. Also DXR requires bandwith, a lot of bandwith, so the industry have to go over a TB/s to achive good ray tracing performance. This is the only parameter that matters a lot. With around half TB/s ray tracing we will never go over Full HD.

Navi is supposed to be from PS5 right? PS4 had 1152 GCN cores. PS4 Pro - 2304 cores. PS5 could have perfectly straight again doubling the amount of cores: 4608. Which would fit perfectly with those 6 shader Engines, and scalability of the architecture.

But that is just my hypothesis.

PS5 is vastly different from Navi. But I can't talk about it.

 

[Glo.]

I'm pretty sure that Navi don't use GDDR6. The minimum bandwith will be 1 TB/s, and it is not possible with GDDR. Also DXR requires bandwith, a lot of bandwith, so the industry have to go over a TB/s to achive good ray tracing performance. This is the only parameter that matters a lot. With around half TB/s ray tracing we will never go over Full HD.




PS5 is vastly different from Navi. But I can't talk about it.

Feature wise - yes. But AMD has a particular addiction to reusing the core count and shader engine layout from Semi-Custom consoles: PS4Pro - Polaris 10 count.

What I meant by that, is that if PS5 has 4608 GCN cores, it will be most likely the core count of Navi 10, but reduced by the specific custom features like... the Polaris 10 had compared to PS4Pro, but had the same core count.

 

[Muhammed]

PS5 is vastly different from Navi. But I can't talk about it.

The next Xbox is also vastly different from GCN, I will say no more though

 

[Kenmitch]

The next Xbox is also vastly different from GCN, I will say no more though

With your extensive knowledge of the industry and all things nvidia why don't you help this guy out with his question?

https://forums.anandtech.com/thread...-usable-to-train-our-own-ai-networks.2554449/

 

[Headfoot]

The next Xbox is also vastly different from GCN, I will say no more though

We all see what this post is doing, except Zlatan has a history of high quality posts, accurate information, good perspective. Ask yourself if the quoted post achieves the same level of quality.

 

[darkswordsman17]

I'm pretty sure that Navi don't use GDDR6. The minimum bandwith will be 1 TB/s, and it is not possible with GDDR. Also DXR requires bandwith, a lot of bandwith, so the industry have to go over a TB/s to achive good ray tracing performance. This is the only parameter that matters a lot. With around half TB/s ray tracing we will never go over Full HD.




PS5 is vastly different from Navi. But I can't talk about it.

Where are you getting minimum bandwidth claim from? I really cannot fathom AMD thinking that Polaris level GPU suddenly needs 4 times the bandwidth. Double? Sure. And they can get more than double with GDDR6 pretty easily. Unless the issues (price, complexity) with HBM goes down significantly, I just don't see that being used. Not in a mainstream video card, and I doubt it in a console. Although since the console will probably sell for more ($399-499, granted it has other components that make it inherently cost more than a video card, there's still some extra margin they could work with), and because its sales numbers will likely be higher perhaps economies of scale make it more feasible there, and a console would benefit from the potential power savings and size reduction.

We'll see how much DXR capability they plan for it to have. And since ray-tracing is going to be limited for a while, they'll likely address shortcomings like that on newer GPUs. They might would overbuild for console because they probably won't be updating that as soon.

Are you speaking purely about graphics portion? Because well, the PS4 was GCN based but quite a bit different simplye seeing as it had CPU cores, audio processing blocks, its own media processing block. And then the PS4 Pro was based on Polaris but had some differences. But its not night and day, but more than enough for them to say its substantially different. And then Scorpio had differences so it wasn't Polaris, nor was it Vega, nor was it PS4 Pro.

And if its based on the current dev kits I wouldn't read much at all into that as its not using Navi based chips at all yet.

I can definitely see there being pro/Instinct based Navi that have HBM, but consumer? I'm quite skeptical, and even moreso that they'd be pushing for 1TB/s worth of HBM. Does HBM3 (or is it possible with HBM2) to increase the bus width without needing to go for more stacks?

 

[DrMrLordX]

What will AMD's 7nm GPUs look like?

A lot like John Cena, since we won't see them. Not us consumers anyway. Until, like, Q4 2019.

 

[jpiniero]

The next Xbox is also vastly different from GCN, I will say no more though

Have to think it's very likely they stick with AMD, especially with Sony staying. The only other option for them is Intel basically.

 

[darkswordsman17]

Have to think it's very likely they stick with AMD, especially with Sony staying. The only other option for them is Intel basically.

I don't know why you're even speculating about that. I don't think the other person was insinuating such a thing with that comment. AMD has explicitly said they're working with Microsoft on future console.

 

[jpiniero]

I don't know why you're even speculating about that. I don't think the other person was insinuating such a thing with that comment. AMD has explicitly said they're working with Microsoft on future console.

AMD hasn't really confirmed anything, only that they are pursuing it. It's a safe bet that MS will be using AMD, but things could happen you know.

 

[darkswordsman17]

AMD hasn't really confirmed anything, only that they are pursuing it. It's a safe bet that MS will be using AMD, but things could happen you know.

Oh please. There's been zero rumor or anything to indicate they're working with Intel and/or Nvidia (or anyone else), and AMD keeps saying they're working with both Sony and Microsoft on consoles (Lisa Su once again specifically mentioned that earlier this month). I guess we can't rule out that they're looking at using Power, so IBM could be in play. I bet Sony might even go back to Cell! And this time it'll be the GPU too! Technically we can't rule out that AMD is going to start selling NVidia GPUs, so its possible that they'll have AMD CPU and Nvidia GPUs!

 

[Kenmitch]

Technically we can't rule out that AMD is going to start selling NVidia GPUs, so its possible that they'll have AMD CPU and Nvidia GPUs!

We should start our own rumor mill....We could use the recent performance gains for the 2990wx as the first signs of the joint collaboration. It's a stretch, but just gotta throw in a couple of anonymous sources in the industry. lol

 

[HurleyBird]

Feature wise - yes. But AMD has a particular addiction to reusing the core count and shader engine layout from Semi-Custom consoles: PS4Pro - Polaris 10 count.

That's only because the PS4 Pro has 4 disabled CUs.

 

[Olikan]

They do not have to increase amount of cores/ROPs, etc, if they can find a way to feed the cores with work.

http://www.freepatentsonline.com/20180144435.pdf

This patented by AMD tech is one of implementations of it. This tech saves memory bandwidth, reduces power consumption, and allows AMD to increase core clocks.

And is DIRECTLY connected to this patent, from similar time:
http://www.freepatentsonline.com/20180121386.pdf

Btw... Parallel micropolygon rasterizers
https:///patents.justia.com/patent10062206