Today we're finally going to explore Kaby Lake G, one of the most unexpected and interesting hardware components that has been from Intel for some time. This is the first chip that combines an Intel CPU and an AMD GPU on a single piece of silicon, making it essentially the fastest APU-style processor on the market.
It's not quite the same as an APU or SoC, but it combines a powerful CPU and a high-performance GPU on a single, compact chip.
In the meantime, you may have reported a bit about Intel's Kaby Lake G NUC, also known as Hades Canyon. This was the first Kaby Lake G product to hit the market, but as a NUC it is a niche product and most reviewers only have access to the top-end model with the fastest Kaby Lake G SKU.
Given that Kaby Lake G was primarily designed for mobile devices like laptops where the top end SKU is unlikely to be used, we wanted to have the processor in our hands in an actual laptop as a more common SKU to check this chip.
The processor we have available for review is the Core i7-8705G, one of the 65 W Kaby Lake G variants, in contrast to the full 100 W units of the NUC. There are already two laptops that use the 65 W model. If more laptops use Kaby Lake-G in the future, they are far more likely to opt for a 65W model as we test it today. The benchmarks we will be reaching shortly should therefore give a good idea of how Kaby Lake G will develop for most users.
We tested the 8705G in the Dell XPS 15 2-in-1, a slim and light 15-inch laptop that also contains 16 GB of two-channel DDR4-2400. Since the best use cases for Kaby Lake G are these slim ones, we get a good idea of how the chip behaves in a typical usage scenario.
Let’s explore Kaby Lake G, the line-up, and its specs before we look at performance numbers. The chip itself contains both an Intel CPU and an AMD GPU in one package, but is not a combination of the two in one chip. Instead, the package contains two chips: a quad-core Intel CPU with integrated HD graphics and a Radeon RX Vega M-GPU with up to 24 processing units and 4 GB HBM2 memory.
The two chips are connected via eight PCIe 3.0 lanes, and the GPU is connected to its HBM2 via the so-called embedded multi-die interconnect [EMIB]. In this picture you can see exactly how the chips are arranged on the chip: The Intel CPU is on the right, the AMD GPU and HBM2 are on the left. The HBM2 chip, as you can see right now, is separate from the GPU and is between the chips that Intel EMIB brings into play.
There are a total of five Kaby Lake G-SKUs, neatly split versions with a 100 W TDP and those with a 65 W TDP. The 100 W SKUs use Vega M "GH" graphics, a fully unlocked version of the GPU with 24 processing units and a base clock of 1063 MHz, which is increased to 1190 MHz. The 65 W versions use Vega M "GL" graphics, which reduce the GPU to 20 computing units and reduce the clock speeds on a 931 MHz basis and 1011 MHz boost. This is the main difference between the 100 W and 65 W chips. The additional headroom offers more CUs and higher clocks for better GPU performance.
The CPU is pretty much the same across all five SKUs. We look at four cores and eight threads with slight variations in clock speeds between models. The Core i7-8705G that we're looking at today is clocked on a 3.1 GHz base, with a maximum turbo clock of 4.1 GHz on a single core. This drops to 3.9, 3.8, and 3.7 GHz for two, three, and four core workloads.
The 8709G uses the exact same CPU, but this chip has the top-end 24 CU Vega variant. The top end 8809G slightly increases the CPU clock rate from there. And at the bottom, we have the Core i5-8305G, which has lower CPU clock speeds and less L3 cache, but the same 20 CU Vega GPU as the 8705G we're looking at today.
The Core i7-8706G is odd in that it is exactly the same as the 8705G from the specification perspective, except that Intel corporate functions such as vPro and Trusted Execution Technology are enabled. The 8809G is the only overclockable chip.
In HBM2 configurations, all Kaby Lake G-Chips are equipped with 4 GB on a 1024-bit bus. However, the memory is clocked at 800 MHz for the 100 W SKUs and 700 MHz for the 65 W models. All Kaby Lake G-SKUs are also equipped with Intel graphics cards that are integrated into the CPU chip. This is the standard HD 630 GPU with a clock speed of up to 1100 MHz. The idea here is that the laptop can switch off the power-hungry AMD GPU completely on battery power and can only work with integrated graphics. When the "discrete" AMD GPU is needed, it can be turned on for far better graphics performance.
The last thing we want to discuss before we start with the performance is the power distribution between CPU and GPU. The 8705G has a total TDP of 65 W, but the PL1 power limit for the CPU – in other words, the permanent power consumption limit for the CPU – is limited to 47 W in the Dell XPS 15 2-in-1. This also makes sense, since the quad-core CPUs of the Kaby Lake H series from Intel had a TDP of 45 W and Kaby Lake G increased the clock rates across the board by around 300 MHz.
This gives the GPU a full 18W to work with, at least when the workload is on. In practice, the GPU runs at full load of the GPU and the CPU depending on requirements over 18 W and the CPU under 47 W. As with every SoC, algorithms determine how much power is allocated to each part of the chip depending on the workload. However, you can already see at the PL1 limit that the GPU still has some headroom even when the CPU is fully loaded.