Intel Skylake Platform

when they mention flops are they including the gpu ?

 

I don't think Intel plans to integrate GPUs into Xeons.

 

Woa, Intel has invent Dual Precsion! That should read double precision, and that mistake makes the origin of the slide dubious.

 

Might as well just be a marketroid powerpoint user typo. In any case, other people (like Nick) seem to believe in a doubling of FP computing power in skylark. It's within the realm of reason methinks. They have to use all those extra transistors in the .14u node for something...

 

AVX-512 confirmed for Skylake?

Code:

1. Feature Overview:

      a) Name of feature:
Intel AVX-512 ISA 
      b) Feature Description:
    Intel AVX-512 consisting of AVX512F - Intel® AVX-512 Foundation
Instructions, AVX512PF - Intel® AVX-512 Prefetch Instructions, AVX512CD -
Intel® AVX-512 Conflict Detection Instructions, AVX512ER - Intel® AVX-512
Exponential and Reciprocal Instructions are an extension of AVX2. It introduces
the following architectural enhancements:
• Support for 512-bit wide vectors and SIMD register set. 512-bit register
state is managed by the operating system using XSAVE/XRSTOR instructions.
• Support for 16 new, 512-bit SIMD registers (for a total of 32 SIMD registers,
ZMM0 through ZMM31) in 64-bit mode. The extra 16 registers state is managed by
the operating system using XSAVE/XRSTOR/XSAVEOPT.
• Support for 8 new opmask registers (k0 through k7) used for conditional
execution and efficient merging of destination operands. Again, the opmask
register state is managed by the operating system using XSAVE/XRSTOR/XSAVEOPT
instructions
• A new encoding prefix (referred to as EVEX) to support additional vector
length encoding up to 512 bits. The EVEX prefix builds upon the foundations of
VEX prefix, to provide compact, efficient encoding for functionality
available to VEX encoding plus the following enhanced vector capabilities:
• opmasks
• embedded broadcast
• instruction prefix-embedded rounding control
• compressed address displacements
Extension consists of a set of new instructions enabling the usage of the new
architectural enchancements.

    2. Feature Details:

      a) Architectures:
      64-bit Intel EM64T/AMD64

      b) Bugzilla Dependencies:
      None.

      c) Drivers or hardware dependencies:
      Skylake architecture

      d) Upstream acceptance information:
      In the process of upstreaming.

      e) External links:

      f) Severity (H,M,L):
      High (required for Hardware Enablement)

      g) Feature Needed by:
      OSVs

    3. Business Justification:

      a) Why is this feature needed?
      Enabling full usage of the new Skylake processor for customers.

      b) What hardware does this enable?
      Skylake processor

      c) Business impact?
      Big. AVX-512 will be available both in SKL and KNL architectures target
whole spectrum of compute continuum.

 

Well, if it's a hoax, then it's a well concocted one. Areg Melik-Adamyan works for Intel on the Linux OS team, and his boss Keve A Gabbert leads it -- and is on the "copy to" email distribution for that bug. Areg has a history of submitting code and bug / feature requests for other *nix operating systems related to Intel low-level features.

 

Skylake-E should be pretty insane. With AVX-512, DDR4 and a minimum of 8 cores it's going to be quite a beast. It's safe to assume it'll be packing close to 2 TFLOPS and over 68GB/s. There's also every chance it'll contain more than 8 cores and support DDR4 2400Mhz.

Best case scenario - 12 cores and DDR4 2400Mhz @3.5Ghz gives us 2.8 TFLOPs and 76.8GB/s bandwidth. Here's hoping.

 

Come on, only 12 cores?

Ivy Bridge-EX will scale up to 15 cores (likely has 16 cores, one disabled for yield). Haswell-EX has been rumored to scale to even larger core counts (Wikipedia says 16-20 cores). 12 core maximum for Skylake would be a disappointment, unless of course you are talking solely about the consumer (i7) models. 14 nm process should allow them to cram in twice as many cores (in the same die area)...

 

I'm pretty sure he was talking about the consumer models.

Man, it'd be really nice to see core counts taking off.

 

Yep I was thinking about consumer models only. I'd love for standard desktop models to sport 6 or ideally 8 cores with the E going with 12 (more seems too much to hope for at this stage for an i7). Given the disappointment of 4 core Haswells though I not overly hopeful for more at this stage.

 

Intel has already announced that there will be a Haswell-E based 8-core consumer model (i7) launching next year (with DDR4). That's a slight (+33%) improvement in core counts over the Sandy-E/Ivy-E 6-core consumer models. If we assume similar similar improvements for Broadwell and Skylake, the top end consumer models should have 10+ and 12+ cores. 14 nm would easily allow that, but it all comes down to economics. Intel definitely would need some competition. And we definitely need more consumer software (and games) that show big gains for increased core counts. As long as most consumer software doesn't show any noticeable improvements (over 4 cores), there's no point in bringing CPUs with more cores to the consumer market. Software needs to evolve as well.

 

Even Broadwell only has 4 cores, as indicated by the pictures we've seen.

The consoles might help with that (and this may even have influenced AMD when co-designing the chips for the Xbone and PS4).

But Intel will always be reluctant to produce high-throughput consumer chips that might cannibalize Xeon sales.

 

Do you think the pervasiveness of 8 thread consoles games may drive Intel to 8 core mainstream CPU's?

I'm not sure how the 6/7 thread games of the current generation factor into that though.

 

That doesn't bode well but the possible saving grace there is that Broadwell is only for mobile so far isn't it? They may release desktop parts but I'd assume any leaked die shots to date would be of mobile parts.

I'm not sure this is something we need to worry too much about. Each generation brings much higher core counts in server parts as sebbbi says so if the EX is sporting at least 16 cores in the Haswell generation I wouldn't be surprised to see 24 cores by the time Skylake launches. That makes an 8 core mainstream CPU and even 12 core E version "low end" enough to not cannibalize the server chips too much I'd have thought.

 

True, but the desktop will be stuck with a minor refresh of Haswell, so no extra cores expected here either. Both AMD and Intel seem to agree that 4 cores are enough for a mainstream chip, and they're more interested in spending transistors on higher integration, cache or a bigger GPU. Basically things that make computers cheaper or lower-power.

To be honest, they're probably right. Vishera, with its 8 cores is often a match for the i7-4770 in well threaded applications, but the latter is still much more successful because of its higher per-thread performance, lower power, and integrated GPU. More than 4 cores would be wasted on 95% of the target market (if not more) so why bother?

I would expect Skylake to move towards even higher integration, perhaps some kind of fancy memory stacking à la Crystalwell, but available across all SKUs, a much bigger GPU, and all-around improvements in IPC, power and clocks.

 

I don't see this happening, even going beyond 4 cores for the non-E parts. Haswell-E, that's coming in late 2014, has only 8 cores, and actually has less Base frequency than the Ivy Bridge E part(3.0 vs 3.6GHz). The PC segment is still very per-thread performance dominated for it to be worth it, and a new trend is emerging, which is to be at absolute low power.

 

I hadn't realised clock speeds had been announced for Haswell-E, do you have a link please? 3Ghz is crazy low and very disappointing if so. Obviously Turbo should still be able to keep single threaded performance up there but even so, dropping so much frequency over the IB makes the extra cores seem hardly worth it.

The stupid thing is though that as much as Intel has no competition and so doesn't need to push boundries on the desktop, they must still be hurting their own sales through lack of competition with their older models. I have a 3.3Ghz quad SB at the moment and even a jump to the highest end, 2 generations newer Haswell today would result in a virtually unnoticable difference. At their current rate of progression it'll probably be another half decade before I'd be even tempted to upgrade. A 4Ghz, 8 core broadwell at mainstream prices in 6 months time though (which should be well within Intels capabilities) would certainly tempt me.

I know the software isn't there yet but maybe the new consoles with their heavy focus on GPGPU will drive up CPU requirements in games.

 

A 3.0 GHz number for Haswell-E is shown in a footnote of this slide (VR-Zone).

It's interesting that no Haswell so far (at least on Wikipedia) even goes up to 3.7 GHz, unless cTDP or Turbo are used. EDIT: I did some more research (on Wikipedia) and 3.7 GHz is the maximum 4-core Turbo Boost that the 4770K/4771 (the current Haswell-DTs with the highest base clock, 3.5 GHz) have. Given that the slide appears to use this Turbo speed, I speculate it may be possible that the 3.0 GHz for the Haswell-E represents the 8-core Turbo speed.

 

Wow, that's crap. We need another Athlon64.