Well what GPU do you think they could cram into a CPU without it being a problem, going beyond VEGA 11?
Let's look at Intel, for a moment. On their 10 nm process (which is pretty equivalent to the 7 nm TSMC process that AMD is using) the Gen11 iGPUs max out at 64 EU. Each EU has 8 of what AMD tearms "shaders". So, that works out to
roughly what AMD would term a 512-shader iGPU.
A Vega CU (which I guess they call NCU, because it was "new", at one point) has 64 shaders. So, a 11-CU Vega iGPU has 704 shaders. That suggests that Intel's Gen11 is mostly a catch-up exercise. Of course, it's a bit simplistic to compare GPUs on shaders, alone. Anyone following the AMD vs. Nvidia race could tell you that raw compute isn't the whole story. Even within AMD's products, Navi shows that Vega didn't use its compute very efficiently. But, I digress.
So, we see a rough convergence around 8-11 CUs. Maybe that suggests that memory is a bottleneck, if you go much higher. That shouldn't be surprising, if you look at the compute vs. bandwidth ratio of their dGPUs. Below are numbers derived from the base clocks of the respective reference card specs.
Polaris (GDDR5):
| Model | Bus Width | GFLOPS | GB/sec | FLO/B |
|---|
| RX 550 | 128 | 1126 | 112 | 10.1 |
| RX 560 | 128 | 2406 | 112 | 21.5 |
| RX 570 | 256 | 4784 | 224 | 21.4 |
| RX 580 | 256 | 5792 | 256 | 22.6 |
| RX 590 | 256 | 6769 | 256 | 26.4 |
Vega (HBM2):
| Model | Bus Width | GFLOPS | GB/sec | FLO/B |
|---|
| RX Vega 56 | 2048 | 8286 | 410 | 20.2 |
| RX Vega 64 | 2048 | 10215 | 484 | 21.1 |
| Radeon VII | 4096 | 11136 | 1028 | 10.8 |
Navi (GDDR6):
| Model | Bus Width | GFLOPS | GB/sec | FLO/B |
|---|
| RX 5500 XT | 128 | 4703 | 224 | 21.0 |
| RX 5700 | 256 | 6751 | 448 | 15.1 |
| RX 5700 XT | 256 | 8218 | 448 | 18.3 |
Now, what patterns do we see? Well, first, we should probably ignore the RX 550 (indeed, there are supposedly 64-bit bus versions of that GPU, which have FLO/B right in line with its siblings). Also, we should ignore Radeon VII, since it wasn't originally intended to be a gaming GPU. As for the rest, we see FLO/B range from 15.1 to 26.4, with a definite clustering around 21.
So, what does this tell us about GPU bandwidth requirements? Well, at least for Vega, I think we can say that you probably don't want much more than 21 GFLOPS per GB/sec. Given that the Ryzen 5 3400G supports up to dual-channel DDR4-2933, which I think is good for somewhere in the neighborhood of 47 GB/sec, that would suggest about 987 GFLOPS of compute. Dividing that by 2 gives us a target shader * GHz product. With typical AMD iGPU clocks ranging from 1.1 to 1.4 GHz, that yields 449 to 353 shaders. So, that's actually between about 6 and 7 CUs. This suggests their APUs are
already compute-heavy.
For a sanity-check, consider that Ryzen 5 3400G's GPU reportedly delivers 1971 GFLOPS (though I'm not sure if the 1.4 GHz figure is technically base or boost clocks). Divided by 47 GB/sec, we get about 42 FLO/B. So, it's a real outlier. Very compute-heavy, for the amount of bandwidth it has. And it has to
share that bandwidth with the CPU! You can further validate this, by checking out various benchmarks people have run on it, where they vary the memory frequency.
In conclusion, I'd say that iGPUs are already rather maxed, unless/until you do something to alleviate memory bottlenecks. Maybe put some HBM2, in package. However, that would require a bigger package and significantly increase price. Anyway, once you can scale up the bandwidth,
then your bottlenecks become power and thermals.
Where I would expect such a product to
possibly make sense is in the laptop sector. There, you could potentially reap some cost savings from the higher-level of integration. What's tricky about that proposition is that high-end laptops are in the 16 - 32 GB range. Certainly, you want at least 8 GB for the GPU, so 16 GB would be a minimum. However, HBM2 isn't cheap and the CPU doesn't really need so much bandwidth. So, at that point, it's tempting just to move the GPU into its own package, with its own memory. Furthermore, the thinnest laptops, that place the highest premium on integration, are also the most power & thermally-restricted - so, not great candidates to feature an extremely high-powered iGPU. And in anything bigger, then it doesn't seem such an issue to have a dGPU.
As a bonus, let's try the same analysis on some recent consoles (both Polaris + GDDR5):
| Model | Bus Width | GFLOPS | GB/sec | FLO/B |
|---|
| PS4 Pro | 256 | 4198 | 218 | 19.3 |
| Xbox One X | 384 | 6001 | 326 | 18.4 |
This seems to confirm that you really want somewhere in the realm of 1 GB/sec of bandwidth, for every 20 GFLOPS of GPU horsepower.
Facebook
Twitter
Instagram