If you look at the specs, you'll see that RTX 2070/2080 have the same level of GP/s as GTX 1070/1080, while offering a significant performance boost (~35%) over the previous generation. RTX 2070 performs close to GTX 1080 Ti, and RTX 2080 way beyond it with far less rasterization performance. The rasterization performance of Vega is not that far off. Nvidia is not bottlenecked by rasterization performance, neither should AMD be.
You do realize that a Vega 56 overclocked to 1710 MHz (12.2 TFLOPS) is close to RTX 2080 Ti level theoretical performance? (11.8-13.4 TFLOPS) While the similarly performing RTX 2070 have just 6.5-7.5 TFLOPs! No, GCN can't survive in this battle, it simply doesn't scale.
The problem with GCN is not lack of rasterization hardware, it's not lack of computation performance (it has plenty) and it's not lack of memory bandwidth (plenty here too). It's siply lack of saturation of resources, sometimes 30-40% of the CUs are simply "starving". The reason why you can sometimes see like an overclocked Vega 56 scale better than an overclocked Vega 64 has to do with resource management, not total resources. Low-level scheduling is always done in the GPU, and some bins just strike a better balance of scheduling, cache utilization and register files.
You can see the same thing with GTX 970; while GTX 960, 980 and 980 Ti all perform nearly identical in performance per TFLOP, GTX 970 performs much better than it should "on paper". This is also the reason why many factory overclocked GTX 970s where close to stock GTX 980s, offering an incredible value. So why does GTX 970, which is a cut down GTX 980, scale so much better? It simply strikes a sweetspot in GPU resources for scheduling, cache management and register files.
You are wrong.
RX Vega 56 with
64 ROPS at 1710 Mhz and ~12.2 TFLOPS beaten Strix Vega 64 with 64 ROPS at 1590 Mhz with
13 TFLOPS! Lesson: RX Vega needs ROPS power.
At 1710 Mhz, this Vega 56 has faster rasterization, ROPS and geometry.
There's four Rasterizer units with four Shader Engines equipped GCNs. RB units has the ROPS units.
For classic rasterizer, compute TFLOPS means little without ROPS read/write units.
RTX 2080's 64 ROPS has
nearly 2000 Mhz clock speed and coupled with
4 MB L2 cache (with NVIDIA superior delta color compression).
At 1928 Mhz (stealth boost mode), RTX 2080 has 11.85 TFLOPS.
GTX 1080 Ti's
88 ROPS clock speed range from
1600 Mhz to 1800 Mhz and coupled with
3 MB L2 cache (with NVIDIA superior delta color compression). RTX 2080 is superior with 4MB L2 cache when compared to GTX 1080 Ti's 3MB L2 cache.
My MSI GTX 1080 Ti FE reference model goes beyond paper spec clock speed mentioned in techpowerup's GTX 1080 Ti FE database. My GTX 1080 Ti FE at 1600 Mhz has 11.5 TFLOPS and at 1800 Mhz has 12.9 TFLOPS which is dependent on thermals. Techpowerup's database with Pascal GPU's clock speed specs are bullshit since my stock GTX 1080 Ti FE is faster than the paper specs.
RTX 2080 Ti's 88 ROPS has
nearly 2000 Mhz clock speed and coupled with
6 MB L2 cache (with NVIDIA superior delta color compression).
At 1928 Mhz (stealth boost mode), RTX 2080 Ti has 16.78 TFLOPS.
NVIDIA GPUs has stealth boost mode features goes beyond paper specs.
RX Vega 64 has
64 ROPS at 1536 Mhz and coupled with 4 MB L2 cache (AMD's delta color compression is inferior)
ROPS coupled with multi-MB L2 cache is used for tile render methods.
For comparison, GP102's six rasterziation units for each GPC block.
GP102's ROPS range from 88 to 96 ROPS with clock speed range from 1600 Mhz to 1800 Mhz, or sometimes 1900 Mhz during cold winter. My GTX 1080 Ti FE usually has 1700 Mhz to 1800 Mhz clock speed range, but rarely drops to 1600 Mhz.
Vega 64 can't match my GTX 1080 Ti's
six rasterziation units and 88 ROPS with clock speed range between 1600 to 1800Mhz.
GP104 has four GPC blocks, hence four rasterziation units just like Vega 56/64's four rasterziation units.
GP104 has 64 ROPS, just like Vega 56/64's 64 RPS
GP104 has higher clock speed while Vega 56/64 has higher horizontal CU compute.
No brainier to why Vega 56/64 lands into GP104 level game results.
GCN lacks rasterization hardware while cryptocurrency workloads doesn't use rasterziation/ROPS hardware.
