Could Skylake present us a regression in single-threaded performance?

[Grooveriding]

Poll choices are off. Major improvement is not over 5%, maybe something on the order of 25% would qualify as a major improvement. We've been getting minor improvements for years now.

 

[NTMBK]

Tejas was cancelled over 10 years ago. I expect any useful concepts from it were already integrated long ago.

 

[Alatar]

Intels ES are mostly run at production clock speeds, at least the ones that show up only a few months vefore launch like these ones.

It's not a few months from launch, these ES chips are literally the first round.

Early ES chips from intel are usually at low frequencies. Hell SB-E ES chips came in at 1.6ghz. The final set of ES chips that reviewers etc. get are the ones that are clocked the same as normal production CPUs.

 

[MisterMac]

Isn't that what happened over at AMD after the Phenom II's?

In some ways - but it was also a design bet on MT being king.

This is pretty much saying FX 8150 should have a been 6 core, sameclocked and lower IPC Product.

Even AMD would not let that happen, hence why 8 "cores" were devised.

 

[MisterMac]

If the choice was to accept a small hit in absolute performance to get much higher performance/watt, I suspect most proud engineer\manager\VP\etc would happily swallow their pride. I guess it also depends which aspect of CPU you are more proud of: its outright performance or peerless performance/watt.

Anyway, looks like I was freaking out over nothing. Y'all guys are confident Skylake should be a decent improvement over Broadwell and that's good enough for me. I was just beginnning to fret if Intel had extracted all the performance from their core design and now their focus would shift to increasing performance/watt at all costs.

It's extremely unlogical that overall Intel would sell a product with worse IPC across the line.

We have low power variants - for better perf\watt.
Your basicly saying intel would release a 2500k>3500k with worse IPC.

That's what i'm saying is unthinkable and stupid.
I don't have high hopes for skylake - but atleast 10% ipc.

seems to be a simple goal they can achieve\control - and incrementally improve each generation with that mostly satisfies all buyers all around all product groups.

 

[NTMBK]

It's extremely unlogical that overall Intel would sell a product with worse IPC across the line.

To be fair, that's precisely what Intel did with Pentium 4. I guess *in theory* they might drop IPC if it lets them hit their high clocks at a lower power consumption- a 3GHz processor in the Surface Pro 4 would be pretty awesome, and increasing clock speeds on their multicore server chips might be nice. Doesn't seem very likely though.

I wonder how they will handle AVX-512. My personal guess is that they will still have 256-bit execution units, and break it into 2 ops. It feels more like a compatibility play, making it easier to develop software for the Xeon Phi by providing binary compatibility on the Xeon. Still, the improved instruction set will be nice, even if the total throughput doesn't increase.

 

[mikk]

AVX-512 is only for Skylake EP/EX and KNL most likely.

 

[III-V]

AVX-512 is only for Skylake EP/EX and KNL most likely.

That would be an interesting departure from the norm.

 

[Nothingness]

That would be an interesting departure from the norm.

What norm? Intel has become dumber and dumber with utterly silly segmentation. I wouldn't be surprised to not see AVX-512 on all Skylake variants. Even AVX isn't available on all Haswell And TSX is not on all K variants.

 

[NTMBK]

What norm? Intel has become dumber and dumber with utterly silly segmentation. I wouldn't be surprised to not see AVX-512 on all Skylake variants. Even AVX isn't available on all Haswell And TSX is not on all K variants.

Actually TSX isn't on anything right now... there was a bug in it, Intel disabled it through microcode for every SKU.

 

[III-V]

What norm? Intel has become dumber and dumber with utterly silly segmentation. I wouldn't be surprised to not see AVX-512 on all Skylake variants. Even AVX isn't available on all Haswell And TSX is not on all K variants.

Assuming AVX-512 support includes the implementation of 512-bit vector units, mikk's claim, if true, would undoubtedly mean that Skylake-E would have a different core layout compared to 1151 variants. The units would be too large to sit there unutilized.

 

[mikk]

https://gcc.gnu.org/wiki/cauldron20...AVX-512_Vector_ISA_Kirill_Yukhin_20140711.pdf


Intel refers to Skylake Xeon, theoretically the entry Skylake Xeon could support AVX 512. On page 3 complex & versatile big cores makes EP/EX Xeon more likely. AVX 512 is not important for the client platform anyway.

 

[Nothingness]

Assuming AVX-512 support includes the implementation of 512-bit vector units, mikk's claim, if true, would undoubtedly mean that Skylake-E would have a different core layout compared to 1151 variants. The units would be too large to sit there unutilized.

Agreed, but do you really think having 512-bit SIMD on ultra low power mobile chips makes any sense? We are already seeing Intel decreasing frequency when AVX is getting used in higher-end Haswell chips, and I can see voltage being increased on my Haswell when AVX is used. I doubt 14nm will bring enough magic dust to overcome the power cost of 512-bit wide SIMD units So either Intel will have different masks, or will simply accept to have a significant part of the die unused (though note AVX{2,-512} probably isn't that large compared to the rest of CPU and the GPU). And if they do that for mobile parts, why not do that for other parts too?

That being said, Core M seems to have AVX. But the first results of benchmarks and throttling aren't encouraging, so I don't think this could prove anything regarding AVX-512 support in the future.

Of course, this is pure speculation on my side and disappointment by Intel attitude wrt segmentation

@NTMBK: Indeed :biggrin:

 

[dahorns]

I wouldn't be surprised to not see AVX-512 on all Skylake variants.

Holy double negative Batman!

 

[Nothingness]

Holy double negative Batman!

Heh, and it correctly said what I meant

 

[witeken]

Skylake is the architectural successor to Tejas.

P6M Generation;
- Banias
- Dothan
- Yonah
- Core
- Penryn
- Nehalem
- Westmere
- Sandy Bridge
- Ivy Bridge
- Haswell
- Broadwell

Skylake is the evolution of P68(Netburst) and P7(Itanium). Technically, it would be P8 if Intel was still using that scheme.

So, expect big things from an architecture, so efficient they dropped FIVRs.

What do you base all this hype ("expect big things") on, exactly?

 

[witeken]

To be fair, that's precisely what Intel did with Pentium 4. I guess *in theory* they might drop IPC if it lets them hit their high clocks at a lower power consumption- a 3GHz processor in the Surface Pro 4 would be pretty awesome, and increasing clock speeds on their multicore server chips might be nice. Doesn't seem very likely though.

Do you notice the trend?

What they said about Broadwell didn't make me think they had have intention to do worse with the next architecture.

 

[Ken g6]

I predict that IPC will increase, but clock speeds won't. In fact, I predict that desktop Skylake won't come out for a long time after mobile Skylake. But even if it did come out sooner:

Maximum non-overclocking Intel CPU frequency: 3.9GHz. (Excluding maybe one Xeon.)
Maximum overclocking Intel CPU stock frequency: 4.4GHz.

Difference: 12.8%
Maximum likely IPC improvement from Skylake: 10%

So OC-able Broadwell will almost certainly be worth it.

 

[NTMBK]

Do you notice the trend?

What they said about Broadwell didn't make me think they had have intention to do worse with the next architecture.

As I said, it doesn't seem terribly likely

 

[NostaSeronx]

I expect any useful concepts from it were already integrated long ago.

Tejas/Nehalem (Netburst) would have used Enhanced Hyperthreading. Which has only been used in Poulson (Itanium).

Skylake is architecturally the successor to Tejas (Netburst).

The performance and efficiency improvement is >2x. Between that of Broadwell and Skylake.

 

[ShintaiDK]

I predict that desktop Skylake won't come out for a long time after mobile Skylake.

Desktop Skylake is the first Skylake to launch.

 

[Lepton87]

Tejas/Nehalem (Netburst) would have used Enhanced Hyperthreading. Which has only been used in Poulson (Itanium).

Skylake is architecturally the successor to Tejas (Netburst).

The performance and efficiency improvement is >2x. Between that of Broadwell and Skylake.

Dude, you always have the most strange theories here, especially about SOI.

 

[e-ricesucks]

interesting.
need 24 more to pm.

 

[Fjodor2001]

If Skylake-S will be faster than Broadwell-K, then Broadwell-K is pretty pointless given that they will be released so close in time.

Sure, Broadwell-K can be overclocked. But the share of users actually overclocking their CPU should be very small. So for most people buying Broadwell-K will not make sense.

Previously, the K version was always released on the latest uArch and process node. So then it made sense to buy it even if you did not intend to overclock (it was the top end SKU regardless). But with Broadwell-K vs Skylake-S, this is no longer the case.

 

[NostaSeronx]

Dude, you always have the most strange theories here, especially about SOI.

It's not a theory.

SOI is the only way AMD will ever be competitive. If they go bulk they lose market share, if they stay SOI they maintain market share.

----
Hyperthreading = SoEMT (1 Front-end)
Enhanced Hyperthreading = SMT (2 Front-ends)