Share This article

At IDF this week, Intel shared some of the architectural details for Skylake that were missing from the initial launch. While not a complete disclosure, we now have a better picture of what the CPU is capable of, and where Intel focused to improve performance and power consumption. The briefings at IDF covered both the CPU and GPU; we’ll cover CPU details in a future story. For now, here’s what we’ve learned about Intel’s new on-die GPU.

When AMD launched its Kaveri back in early 2014, it focused on the fact that the GCN GPU accounted for more than half the die. The argument was straightforward — the GPU was going to be ever more important to hitting high performance targets and deserved a larger slice of the pie. Intel, it seems, heard this line of reasoning and agreed. Skylake’s GPU slices can be trimmed or expanded depending on which market segments the company wants to hit, but the full Skylake GPU takes up more die space than any previous chip.

Skylake is also Intel’s first GPU with explicit DirectX 12 support. It’s not clear how the company is handling DX12 on previous chips, we’ve seen early demos of games running DX12 on Haswell and Broadwell, but Intel continues to officially list DirectX 11.2 as the API for its earlier chips.

Intel hasn’t released any hard data on what the part numbers on future Skylake GPUs will be, or disclosed exactly which chips will receive which cores, but we can infer some basic relationships. Typically, Intel has launched high-end desktop SKUs with what it calls a GT2 configuration. GT3 and GT3e (with EDRAM) configurations tend to be reserved for mobile chips or special desktop processors, while GT1 ships on lower-end chips in the Pentium and Celeron families.

Skylake will shift this somewhat by adopting a new GT4 design (more on that in a moment), but the basic structure remains the same.. Like Intel’s previous GPUs, Skylake’s graphics core is divided into slices, sub-slices, and unslices. Each slice contains three sub-slices, and each sub-slice contains eight execution units (EUs). There are 24 EUs per slice, and Intel can stack up to three slices in a GPU configuration, for a total of 72 EUs.

Skylake’s new GPU has more back-end fill rate capabilities, better performance with MSAA enabled, and larger caches to feed the GPU. Like AMD and Nvidia, Intel has increased its own lossless color compression, saving bandwidth and power in the process. The gains here are highly dependent on the title, but Intel claims that Arkham Origins is 11% faster, BioShock Infinite picked up 4%, and Tomb Raider gained 3%. While 3% performance isn’t much in and of itself, it still matters when rolled into a group of more extensive improvements.

GT4, multimedia, and EDRAM

In the past, Intel offered one or two slices in its GT1/GT2 and GT3 architectures. With Skylake, certain SKUs will also include a GT4 GPU that adds a third slice. This will further improve EU throughput, Z/Stencil and pixel operations, and total GPU performance. While the benefits of adding another slice typically aren’t linear, we should still see significant improvements for GT4 over and above GT2 and GT3.

Skylake’s video decoder will support HEVC and VP8 encode and decode in hardware, and the new EDRAM equipped SKUs should significantly improve integrated performance. Intel is also adding support for hardware-accelerated encode and decode for additional formats, as shown in the table below:

We don’t know which chips will carry eDRAM, but Skylake will deploy it in both 64MB and 128MB configurations. Past benchmarks have shown that EDRAM equipped Broadwell processors are significantly faster than AMD’s Kaveri or any other integrated graphics solution. That’s not really surprising, since Kaveri is bandwidth constrained by its dual-channel DDR3, but it could make life even more difficult for AMD if Intel pushes a 64MB Skylake processor into APU territory.

Intel is serious about its graphics performance

From 2000 – 2010, Intel’s integrated graphics performance was barely worthy of being called a joke. Even first generation Atoms, which weren’t exactly known for stellar performance, improved noticeably when paired with an Nvidia Ion GPU. And by “improved,” we mean that even browsing and desktop navigation was noticeably faster. Intel was a CPU company, not a GPU company, and it seemed as though this would never change, especially after Larrabee was canceled.

Sandy Bridge was arguably the first Intel GPU that could reasonably be described as such without provoking a round of snickering. Each generation of Intel processor that’s followed it has improved performance, boosted game compatibility, and delivered a better all-in-one solution. Today, Skylake is capable of holding its own with a $70 GPU, and that’s in a GT2 configuration without the benefit of EDRAM. Intel has signaled that it intends to bring EDRAM to more SKUs with Skylake, and that’s only going to improve low-end performance further. The fact that so many OEMs have ditched AMD and Nvidia GPUs in high-end mobile hardware is just another example of the trend.

If you’re a gamer, this is a unilaterally good thing. Developers have to consider the hardware configurations of everyone who might play their game, which means they’ve had to assume that at least some players would be using Intel integrated solutions. As Intel raises its own bar for acceptable GPU performance, it lifts all players who want a game that doesn’t look awful. Integrated graphics will never kill the enthusiast high-end, because the laws of physics dictate that a separate GPU with dedicated cooling and power will always be able to hit higher performance targets than an integrated solution.

Chipzilla won’t win any points for sprinting across the finish line in this regard — the company is only starting to roll out its EDRAM solution to a wider market now, over two years after it first shipped. Slowly but surely, however, the “good enough” bar is rising. Today, desktop Skylake is capable of driving modern games at 720-1080p with low to medium detail at >30 FPS. Higher end Skylake will likely offer 1080p as a given, with medium or even high detail depending on the game.

AMD will answer with new APUs and a new GPU architecture on 14/16nm, but not for some time. Zen CPUs are expected to ship next year, but Zen APUs may not launch until 2017. HBM will solve the bandwidth constraints that plagued Kaveri, but I suspect AMD will face a tough fight by the time it’s ready to ship its next generation of integrated hardware. Intel may have started off as a joke, but it’s evolved its GPUs into genuinely decent solutions — even if they’d never be the first choice of any hardcore gamer.

Thanks to Matthew Murray, who is on location at IDF this week and contributed to this report.