The thing with the large GPU market is with 7nm cost of development, it needs a piggy back market of some sort to make the cost of developing the GPU worth it.
This means adding back the compute. From what I have seen in terms of luxmarks and other programs that test compute, Navi seems to be 2 steps forward for gaming architecture but one or two steps back as far as compute.
This makes sense because Navi seems to be a pure gaming architecture with the consoles in mind and now has IPC similar to Nvidia turing which is very impressive. It seems to be Maxwell + Async in the behavior of it's shaders which makes it no surprise that it IPC is similar to Turing now. However it appears to have taken a step back as far as compute ability goes
https://www.hardwareluxx.de/index.p...0-und-radeon-rx-5700-xt-im-test.html?start=11
To make a A 400mm, 96 rop, 4096 shader navi that pure gaming is possible but it wouldn't have the compute to be competitive with Nvidia next gen 7nm products and would mostly be more inline with Nvidia's volta or less if we look beyond pure tflops(double precision). This would reduce it's TAM in a very fast growing market which is less desirable for AMD in the long run.
I think AMD's next enthusiast card will have the compute added back on and AMD will try to replicate what Nvidia does where they have compute specific cards on one side of the fence and gaming cards one one side of the fence. I didn't think this was possible before with AMD budget, but with the shifting of labor to China for GPU development, some reverse engineering(Navi performs too similarly in regards to IPC in the same games as maxwell and Turing),the piggy backing of R and D development from the PS5 and Xbox next, it looks like AMD managed to do it.
In terms of beyond, AMD will need to look not at the present but to the future for the design of its next gen card. Looking at Nvidia's financials, half of their revenue is now provided by professional and business markets which means Nvidia's professional cards are going to be strong.
I could see coming from Nvidia for their first compute product to succeed Volta being a 600mm-700mm2, 6144 shader card clocks at a conservative(for 7nm standards)1800mhz. This translates into a 22.1tflops card. This is probably Nvidia's target since GV100 already does 14.7 tflops, so roughly a 50% increase is expected with a new node. This will be followed by a 104 series with 0.666 of the specs(This pattern has been repeated since maxwell). So in this case 4096 shaders with a 102 series being the same 6144 shaders but higher clocks. In higher clocks I am guessing 2.2 ghz for the 104 and 2.1ghz for the gv102 since these clocks can be achieved on air/water today on turing and typically new nodes improve on these clocks and become the new default clock. Assuming no changes in architecture and simply a moving of Turing to a new node, AMD will need something with more than 4096 shaders to compete with Nvidia's next gen.
So what specs will AMD need to compete with a 18tflop card(104) or a 25.8tflop card(100/102) using the above specs.
Assuming the clocks don't improve(and the chance of it improving is not good since the wider specs will eat into the clocks, along with compute and the possibility of double precision), this would mean AMD would need something with 5120 shaders to be competitive with Nvidia and perhaps beat the 104.
5120*2*1.8ghz = 18.4 tflops. This becomes tricky for AMD because such a card is a mammoth for AMD in terms of die size consider the space requirements of 2560 shaders at 251mm2. Add in the likely compute ability that needs to come in and AMD is looking at a 500mm2 plus card. With Nvidia, we are likely going to see a growth of die sizes compared to pascal because of how big turing cards are today, I am estimating ga104 being 350mm to 400mm, ga102 being 525 to 600mm and GA100 being600mm- 700mm2.
This playing around with die sizes shows that although RDNA was a good step in the right direction, AMD competitiveness is coming mostly as a result of the move to a new node vs Nvidia. AMD problem with RDNA is scalability and power consumption and this is why Nvidia is not in panic mode.
As is currently, power consumption reveals problems with scalability.
Nvidia's next gen ga106, should be 2036 shader(looking at trends with maxwell, pascal and turing). At 2.25ghz(these cards are clocked a tad higher because of small die and less leakage), we are looking at a 9.16 tflop card. Things look good when the 5700xt produces 9tflops so it initially looks like a competitive race. That is until you look at power consumption. There's a good chance GA106 will consume 120 watts, like the GTX 960, 1060 and 1660 ti before it. It also shows the scalability of the architecture. If 9.1tflops of power can produced for 120watts, there is the possibility of 18tflops at 200-225watts and 25.8tflops at 300watts.
With RDNA there are concerns when 9 tflops already takes 225watts. Where is the scalability to get 18 or 25.8tflops? The 7.6tflops at 180watts isn't much better from the 5700.
If Navi was a 150watt card at 9tflops(don't use a youtube video where someone undervolts, can't guarantee stability across hundreds of thousands of samples across all uses cases), AMD would be in a much more competitive position for scaling but at 225 watts, the outlook does not look good for scaling upward.
I wonder if AMD will use a chiplet approach for compute GPUs since they seem to have made Navi (and RDNA) for gaming.
Interviews with their engineers said that multiple chillers were fine for compute workloads, but it didn’t work well for gaming.
Sure the professional market can justify the high cost of massive compute dies, but the other side of chiplets is better yield and more product to sell.