Intel Discloses Detailed Skylake Architecture Enhancements

MojoKid writes: Intel is still keeping a number of details regarding its complete Skylake microarchitecture and product line-up under wraps for a few more weeks, but at a public session at IDF, some of the design updates introduced with Skylake were detailed. Virtually every aspect of Skylake has been improved versus the previous-gen Haswell microarchitecture. I/O, Ring Bus, and LLC throughput has been increased, the graphics architecture has been updated to support DX12 and new eDRAM configurations, it has an integrated camera ISP, support for faster DDR4 memory, and more flexible overclocking features. All of these things culminate in a processor that offers higher IPC performance and improved power efficiency. There are also new security technologies dubbed Intel Software Guard Extensions (Intel SGX) onboard Skylake, which support new instructions to create and isolate enclaves from malware and privileged software attack, along with Memory Protection Extensions (Intel MPX) to help protect stack and heap buffer boundaries as well. A new technology, dubbed Intel Speed Shift, also allows Skylake to switch power states faster than previous-gen products, controlling P states fully in hardware, whereas previous-gen products required OS control. The end result is that Skylake can switch P states in 1ms, whereas it takes roughly 30ms with older processors.

Comments

Re:Motherboard compatibility?

[Rockoon]

I wonder if this Skylake can be considered an upgrade path..

That would violate Intels strategy, which is never allow a realistic upgrade path.

Re: Motherboard compatibility?

[myrdos2]

You'll need to wait for Skylake-E or whatever they will call it.

Sheesh. Skynet. They're going to call it Skynet.

Sorry Intel, not enough

[JoeyRox]

Sounds like the last few generations - lots of incremental improvements and excellent technology but wont amount to much of a difference in general performance.

Re:Sorry Intel, not enough

[Kjella]

Sounds like the last few generations - lots of incremental improvements and excellent technology but wont amount to much of a difference in general performance.

Actually they've made quite substantial improvements, but Intel is using that to deliver 105% of the last generation's performance with far less resources. A 32nm Sandy Bridge (i7-2700k) is 216 mm^2, a 14nm Skylake (i7-6700k) is 122.4 mm^2. So on the same wafer Intel can produce 75% more processors. By letting AMD pave the way with APUs they've force-bundled integrated graphics killing low-end discrete chips without an antitrust whimper, almost 20% on Steam now game on Intel and Skylake adds 25% more shaders with 20 vs 16 EUs on regular desktop chips.

Perhaps the most important part of the Skylake announcement has gone largely unnoticed because we haven't seen it in any actual product yet is that Skylake will go up to 72 EUs, as opposed to 40/48 on Haswell/Broadwell. Since Broadwell quad core chips only launched a few months ago, the real comparison is Haswell which means a 72/40 = 80% increase in shaders, clearly Intel is planning to take the midrange laptop graphics too. A fully stacked Skywell seems to be a nVidia 950m-class competitor, both around 1100 GFLOPS and 26 times more powerful than "Intel HD Graphics" from five years ago.

Basically, we're being used. I think Intel knows as well as we do that no matter how fast they released processors people aren't going to throw away their three year old computer anymore. They'll sell new processors when the old go out of service or the market expands, not because it's outdated. So they're using their strengths for profit and market gain, because what's your high end alternative? It's either Intel or an extremely old FX processor or a severely under-powered laptop chip. They know your business is not going anywhere, it's only a matter of how long they need to wait.

Re:BMI/BMI2

[Rockoon]

Sure looks like it

From the errata:

Executing CPUID with EAX = 7 and ECX = 0 may return EBX with bits [3] and [8] set, incorrectly indicating the presence of BMI1 and BMI2 instruction set extensions.

Attempting to use instructions from the BMI1 or BMI2 instruction set extensions will result in a #UD exception.

and in the errata summary, its currently labeled NO FIX so they dont even have a fix that will trap the exception and emulate the instructions (which would perform terribly anyways... but hey, working is better than not working.)

Re:Slower in games, faster in vector maths

[0123456]

Intel's pricing (and refusal to offer 6-core mainstream parts) is a consequence of Intel's effective MONOPOLY in the x86 space. AMD's current CPU offerings are a BAD JOKE, offering around 50% per core of Intel's core performance. No serious PC gamer would opt for anything less than a true 4-core i5. AMD isn't even in the picture.

So why did I pay less for my i7 a couple of years ago than I did when I bought my Pentium-4 back in the days when AMD was actually competitive?

Intel's current competition is ARM, not AMD.

Re:Slower in games, faster in vector maths

[FreonTrip]

If there was ever a true multi-threaded application AMD would take the prize. As such Intel dominates because of single threaded applications.

There are embarrassingly well-threaded applications where AMD does well. The x264 encoder does a fantastic job and hammers all 8 of the cores in my FX-8320 at >90% utilization, and it was cheerfully faster at that than the i5 3570K I used to keep around. But IPC does ultimately win out, and Haswell's AVX2 support is sufficient to let an i5 4690K generally pull out ahead of my FX. That's especially true on interlaced media, where the deinterlacer's essentially single-threaded and the rest of the chip's basically waiting for that single core to finish before tackling the rest of the workload. For most other uses it's somewhere around a Nehalem quad core: certainly fast enough for what I do, but the overall performance outside of niche applications isn't impressive in absolute terms. At least it took to undervolting well, and it's a friggin' behemoth for virtualization.

LLC? ISP? WTF?

[Lost Race]

"Limited Liability Corporation" and "Internet Service Provider" don't make much sense, but then again I'm pretty far behind the times on CPU architecture. Who knows what coprocessors they're spending their insane transistor budgets on these days.

OK, "ISP" appears to mean "Image Signal Processor". "LLC" could mean "Last Level Cache" or "Logical Link Control". "Last Level Cache" makes more sense in context, though this is the first time I've seen that phrase. Usually cache levels are explicitly numbered (first, second, third, etc).

It looks like they spelled out everything else except "IPC" which is obviously(?) "Instructions Per Cycle".

Good job there, author, submitter, and editor!