The launch of Skylake has been a bit of a change from Intel; previously, Intel typically has talked about the underlying architecture prior to the chips actually ending up in the reviewers’ hands. This time, all we got was a chip with some performance expectations, with no real knowledge of what has changed. This is particularly problematic as Skylake is a “tock”: a new architecture on a mature process. Broadwell is mostly the same underlying architecture as Haswell, only moved to 14nm, but Skylake is 14nm with new design elements.
That all changes today, as Intel has taken off the wraps and shed some light on what exactly Skylake is doing under the hood that makes it a better processor. We still don’t have all of the details we’d like, but we do know at least some of the core changes.
There’s a separate topic to also quickly mention: our initial Skylake numbers were somewhat lower than expected, either due to our initial choice of motherboard, RAM, and/or early firmware. We’ve retested with a different motherboard and memory, and performance on average is up 10 percent compared to our earlier results. We’re reworking the Skylake review and will post a followup later, but the short summary is that Skylake is quite a bit better than our initial experience led us to believe. Welcome to life on the bleeding edge.
Moving on to the Skylake architecture, you can download the full PDFs via Intel’s IDF 2015 SF site, but we’ll summarize some of the major highlights.
The above three slides summarize the major changes. Skylake has a larger Out-of-order window, more in-flight stores, more scheduler entries, and a slightly larger allocation queue. The branch prediction has seen further improvements, there’s a wider front-end, there are more execution units, and other additions. The details are a bit lacking in some areas, but one item we specifically asked about was the wider front end.
From the original Core 2 through Haswell/Broadwell, Intel has used a four-wide front-end for fetching instructions. Skylake is the first change to this aspect in roughly a decade, with the ability to now dispatch up to six micro-ops per cycle. Intel doesn’t indicate how many execution units are available in Skylake’s back-end, but we know everything from Core 2 through Sandy Bridge had six execution units while Haswell has eight execution ports. We can assume Skylake is now more than eight, and likely the ability to dispatch more micro-ops as well, but Intel didn’t provide any specifics.
There are plenty of other changes as well, including some new instructions, e.g., Intel Software Guard Extensions (SGX) can be used to improve software security, but those don’t generally affect performance. On the other side of the processor equation, there have also been a lot of power and efficiency improvements. That might seem a bit odd, considering Skylake is sporting a 91W TDP compared to Haswell’’s 84W TDP, but TDP isn’t the same as typical power use.
Intel pointed out that the process of designing Skylake began around five years ago, and at the time their goals were pretty traditional. They were looking at a 3X TDP scale (e.g., 40-120W), a 2X range of form factors, and a classic IO set for PCs. Between the start and now, the market experienced a shift and tablets and smartphones became a huge market. The result is that Skylake will have a 20X TDP scale (so around 4.5W-91W), a 4X form factor scale, and reductions in power and changes to IO that make Skylake a viable product for both PCs as well as tablets. This is the sort of responsiveness that has been enabled by Intel's tick-tock strategy, and the results have been impressive.
The other area where Skylake has the potential to dramatically improve performance is in processor graphics. Going back to 2010 and the original HD Graphics (Lynnfield/Arrandale, aka Iron Lake), Intel has been putting increasingly heavy emphasis on their graphics solutions. Their first solutions had up to 10 Executions Units (EUs) supporting the DX10 feature set, with Skylake pushing as far as 72 EU configurations on their top SKUs. Note that the currently available i5-6600K and i7-6700K are using the GT2 (24 EUs) chips, but we expect 48 EU GT3 chips to show up in the near future.
More importantly, there’s now a GT4 configuration coming, which has the potential to be up to 50% faster than the existing Iris Pro 6200 Graphics. Compared to Broadwell’s Iris Pro (GT3) solution, the Skylake GT4 graphics will be 1.5X larger, and based on what we’ve seen from Broadwell-DT we would expect performance to be in the 30+ FPS range at 1080p medium to high quality. Of course we probably won’t see GT4 outside of mobile chips, at least initially, and discrete GPUs will still offer superior performance.
From maximumpc
from http://bit.ly/1JgufG8
