The Intel Kaby Lake-G, the first Intel chip to sport an AMD Radeon GPU was seen in action in the Intel Hades Canyon NUC. The benchmarks are not exactly groundbreaking, but the Intel NUC does prove to be a decent small factor machine for 1080p gaming. Reports have now started coming in that the Intel Kaby Lake-G Core i7-8809G used in the Hades Canyon NUC is packing a Polaris GPU instead of a Vega.
In tests, many apps are showing discrepancies as far as the power consumption and temperatures are concerned. Tom’s Hardware did some testing to check the validity of these claims and the results are more than intriguing.
Kaby Lake-G: Polaris or Vega
AIDA64, a third-party benchmark utility, identifies the Radeon GPU on the NUC as Polaris 22. AIDA also lists the architecture as AMD GCN4 (Polaris). SiSoftware Sandra on the other hand lists it as Radeon RX Vega M GH.
PC World has also discovered that the Hades Canyon Radeon graphics does not support DirectX 12.1 like Vega graphics chips. Instead, it only supports only DirectX 12.0, similar to Polaris. The AMD Ryzen 5 2400G APU featuring discrete Vega graphics however does support DirectX 12.1.
While replying on third party applications like AIDA64 and Sandra for identifying a GPU architecture is not very reliable, there are multiple other pointers that bolster the claim. The DirectX 12 feature level being limited to 12.0 being the most prominent one. Third-party utilities use PCI device IDs for identification. 694C and 694E are the device IDs for Kaby Lake-G Radeon graphics. These IDs also scream Polaris 22.
Sandra GP3D11 Testing and Rapid Packed Maths
One of the most impressive features of the Vega architecture is Rapid Packed Maths, as seen in the Far Cry 5 benchmarks. Far Cry 5 leverages Rapid Packed Maths, and as such performs really well with AMD Vega cards.
The Vega architecture’s RPM allows the GCN engine to pack two 16-bit operations into a 32-bit register, allowing 64 shaders to perform up to 256 16-bit ops per clock. This cuts the performance costs whenever FP16 compute can be used instead of FP32, as seen in Far Cry 5.
SiSoftware Sandra’s GPGPU Processing test can be used to test FP16 performance. It runs FP16 operations using the DX11 API.
In case of a Radeon RX 470 a.k.a Polaris, the FP32 and FP16 performance is identical. For the Radeon RX Vega 64 and 56, however, a 64.2% and 65.8% boost with FP16 is observed. Then coming to the Ryzen 5 2400G that also has an integrated Vega chip. The integrated GPU in the 2400G also showed a 61.8% delta. This concludes that Sandra’s utility supports Rapid Packed Maths.
And when the Intel Core i7-8809G (Kaby Lake-G) was tested guess what happened? It performed similar to the Polaris based GPUs. The FP32 and FP16 performances were rather similar.
Bottomline: So Is Kaby Lake-G Really Packing Polaris
As of now, nothing can be said for certain. However, one thing is for sure. The Radeon GPU on the Intel Core i7-8809G Kaby Lake-G is not the same Vega as the one the Ryzen 5 2400G. It doesn’t support Rapid Packed Maths or DX12.1 as of now. Whether it can be enabled via a patch or software level code update is also not certain.
The Kaby Lake-G chips are custom made parts specifically for Intel’s 8th gen CPUs. So it’s possible that the architecture was somewhat modified or adjusted to work with Intel’s chips given, it was originally never designed for the purpose. However, the fact is that the press should have been made aware of this discrepancy. Intel has stated that, “Radeon graphics solution built for Intel is similar to the desktop Radeon RX Vega solution with a high bandwidth memory cache controller and enhanced compute units with additional ROPs”
Not exactly forthcoming. We’ll just have to wait and see how this blows over and how Intel responds to the story in the coming days. Till then everything is pure conjecture.