Benchmarks of the Skylake i7 6700K have emerged... Only marginally faster than 4790K...

[patrickjp93]

no competition leads to what you see

skylake is a rip off i wouldnt even think about it unless iwas building a new PC

Hold your breath and pitchforks until we get independent reviews, and even then, get on software developers' and studios' asses about this before you jump on Intel. The amount of forced legacy support, use of Microsoft's junk compiler (by comparison to GCC/Clang/ICC), and lack of multiple code paths to let your newer instructions soar are the biggest reasons for performance improvement slowdowns in recent years. AVX 256 vs 128 is vastly higher performing if you have Haswell and you have the software which uses those instructions. Put a good optimizing optimizing compiler like GCC/Clang/ICC up to the task of compiling your code with -O3 and watch what happens. Even code that doesn't look parallel to you can be reduced to parallel SIMD operations. The more parallelism you get, the better performance, for free. Now, when you're stuck using SISD instructions which go all the way back to the Pentium III, yeah, we already optimized the living daylights out of those instructions years ago. See page 186 and look at cycle latencies for Haswell in the following. http://www.agner.org/optimize/instruction_tables.pdf

 

Most of the mathematical instructions are down to 1-3 cycles per instruction, which is the best you can do for addition/subtraction/multiplication/division anyway due to the hardware algorithm complexity. Anything that involves memory or register shuffling will have higher latency that is determined by the memory subsystem instead of the CPU. So in effect, stop blaming Intel who does all the work.

[patrickjp93]

You are correct at the most basic level, however you're forgetting that software development is not this agile.  There are very few times where a piece of software in it's normal state would change compilers or languages.  Such monumental changes would require a rewrite in most cases and at which point it would be a new version completely. 

 

What I'm comparing is the GPU vs CPU ecosystem.  GPU's get both game updates and driver updates that fundamentally give you better performance with the same basic experience.  CPUs on the other hand rarely are given the same treatment and instead of performance increases from a low level functionality, they are given patches, or codebase changes that change the way the code performs rather than the way the code interacts with the hardware. 

Oh you'd never change languages. But changing compilers and seeing a performance difference just to re-release the whole package? There are games with day 1 updates as big as the executable. If you think that can't happen, I'd say you're deluded honestly. It's a free upgrade! Literally download the newest iteration of GCC/Clang and get 5% better performance? Why not?! And it's not like you have to edit your code. The language does not change, and in fact the compilers' only changes are in the optimization engine or in supporting new language standards or markups, such as C++11 vs. C++14 or adding additional support for CilkPlus or fixing an OpenMP parsing bug. You don't have to touch your code. If the new compilation = better performance, just push out the whole new thing as an update.

[Techhog]

AMD overpromises and underdelivers: "Lol AMD u dun fucked up. What a terrible company."

 

Intel overpromises and underdelivers: "Well, this was expected, but it's ok. Intel FTW."

 

Ugh.

[patrickjp93]

AMD overpromises and underdelivers: "Lol AMD u dun fucked up. What a terrible company."

 

Intel overpromises and underdelivers: "Well, this was expected, but it's ok. Intel FTW."

 

Ugh.

There's no proof Intel's underdelivered. There's plenty of reason to dismiss this benchmark which suggests worse IPC than Broadwell if we follow the 5775C numbers with eDRAM disabled. And this wouldn't just be Intel underdelivering. No one really expects such a huge transformation as Prescott->Conroe, but to actually lose ground would incite riots in the communities. 

 

The double standard doesn't exist, at least not yet.

[Techhog]

There's no proof Intel's underdelivered. There's plenty of reason to dismiss this benchmark which suggests worse IPC than Broadwell if we follow the 5775C numbers with eDRAM disabled. And this wouldn't just be Intel underdelivering. No one really expects such a huge transformation as Prescott->Conroe, but to actually lose ground would incite riots in the communities. 

 

The double standard doesn't exist, at least not yet.

Based on the reactions from people who didn't notice the fishiness in the numbers, the double standard definitely exists. If This were Zen vs. Steamroller, even a 30% IPC increase would be met with a huge negative reaction.

[patrickjp93]

Based on the reactions from people who didn't notice the fishiness in the numbers, the double standard definitely exists. If This were Zen vs. Steamroller, even a 30% IPC increase would be met with a huge negative reaction.

No it wouldn't, We're all pretty sure AMD's exaggerating at least where consumer computing is concerned, but even if it is, an increase that big is to be hailed, even if we're disappointed that it still can't stand up to Intel. 

[Curufinwe_wins]

Based on the reactions from people who didn't notice the fishiness in the numbers, the double standard definitely exists. If This were Zen vs. Steamroller, even a 30% IPC increase would be met with a huge negative reaction.

Just saying, that for Zen to compete at anything other than budget it actually needs huge improvements.

And this is totally a bs article. Industry experts are expecting about 20-40% improvement haswell to skylake.

[Amrat]

Looks like my wait is not worth it. This makes me sad. 

[patrickjp93]

Just saying, that for Zen to compete at anything other than budget it actually needs huge improvements.

And this is totally a bs article. Industry experts are expecting about 20-40% improvement haswell to skylake.

More than 10-15% is insane to expect in my humble opinion. There's barely theoretical room left on the raw latency side for a 25% IPC increase, and I don't think deeper pipelines are the answer, 99% accurate branch predictors or not.

[patrickjp93]

Looks like my wait is not worth it. This makes me sad. 

Wait the 4-5 weeks for objective independent reviews. This suggests worse IPC than Broadwell. It's 99% not likely this benchmark is legitimate.

[Curufinwe_wins]

More than 10-15% is insane to expect in my humble opinion. There's barely theoretical room left on the raw latency side for a 25% IPC increase, and I don't think deeper pipelines are the answer, 99% accurate branch predictors or not.

Well broadwell is somewhere between 10-15% already over haswell. So 5-10 on top of broadwell would be 16-27% boost.

Note I didn't say over broadwell, just over haswell.

[Curufinwe_wins]

Well duh. Me want moar cores. What's your single core score? I assume it's around 140?

[RagnarokDel]

Anddddd am I glad I didnt upgrade.

[Homicidium]

Please AMD, get your shit together! These small improvements are getting pretty annoying.

[patrickjp93]

Well broadwell is somewhere between 10-15% already over haswell. So 5-10 on top of broadwell would be 16-27% boost.

Note I didn't say over broadwell, just over haswell.

No it isn't. In terms of IPC it's about the promised 5% better. In cases where eDRAM helps, that gain in performance is a reduction in memory seek time, and not all software is built so well as to take advantage of a bigger cache, and the L4 only catches what falls out of L3 and holds it for up to 300 microseconds. Anything taking advantage of this has memory or code path reuse that is proximal enough for this space to be useful. Otherwise the L4 cache is a hindrance as it adds to memory seek time in the event data sought is not in any of the caches (a cache miss gets worse the bigger and more levels of cache you have).

[patrickjp93]

Please AMD, get your shit together! These small improvements are getting pretty annoying.

These benches are in all likelihood fake. This projects worse IPC than broadwell. Hold your damn horses for 5 weeks until objective reviews come out. Seriously people, learn to think for yourselves and verify before trusting!

 

In fact, OP, could you please put a big fact [RUMOR] tag on the post title and a disclaimer saying these figures are most likely BS? Because the truth is these projections give worse IPC than Broadwell. When was the last time Intel did something like that? Pentium 4 vs. Pentium 3 actually. In all likelihood these numbers are faked and wrong. Intel would not be dumb enough to claim a Prescott  to Conroe jump and simultaneously lose IPC over the previous generation. Just do us a favor and put a disclaimer to this effect in your posting. It's desperately needed.

[Castdeath97]

Why hasn't anyone consider their name, who calls his website tech bang?

[MageTank]

Well broadwell is somewhere between 10-15% already over haswell. So 5-10 on top of broadwell would be 16-27% boost.

Note I didn't say over broadwell, just over haswell.

Where are you getting the 10-15% from? I've been following broadwell from the beginning, and i've not seen anywhere near that projected IPC improvement. Even using the same sources that you use, i just don't see it.

[patrickjp93]

Where are you getting the 10-15% from? I've been following broadwell from the beginning, and i've not seen anywhere near that projected IPC improvement. Even using the same sources that you use, i just don't see it.

He's confusing the performance delta and clock speed delta between the 5775C and 4790K for IPC gains without thinking about causality. He's forgetting reducing memory seek times does not increase IPC. It only means the CPU spends less time waiting before putting its crunching instructions back to work.

[escape093]

I'm guessing most of the work went into the iGPU.

[MageTank]

He's confusing the performance delta and clock speed delta between the 5775C and 4790K for IPC gains without thinking about causality. He's forgetting reducing memory seek times does not increase IPC. It only means the CPU spends less time waiting before putting its crunching instructions back to work.

Yeah, was gonna say something seemed off. Was just looking at https://www.overclockers.com/intel-i7-5775c-broadwell-cpu-review/ and it only confirms what we already know of broadwell. It also adds more skepticism to this skylake report as you can see the haswell numbers do not align with what we already see. 

 

Until i see official benchmarks from multiple sources, i just cannot give these tests any merit.

[Curufinwe_wins]

He's confusing the performance delta and clock speed delta between the 5775C and 4790K for IPC gains without thinking about causality. He's forgetting reducing memory seek times does not increase IPC. It only means the CPU spends less time waiting before putting its crunching instructions back to work.

That is true. Outside of when I mentioned it specifically I was always talking about performance delta not ipc. Obviously there are numerous contributors. (In fact, I'm less concerned about ipc at the moment than overall delta for the very reasons you have mentioned).

Still generally speaking the architecture change is a bigger ipc jump than the die shrink (although you are completely right on the hard limit on the type of instructions being used) and if broadwell is 5% then I would hope to see another 5-10% from skylake on its own.

Either way we all agree there is no way in the world this is legit.

[Curufinwe_wins]

Yeah, was gonna say something seemed off. Was just looking at https://www.overclockers.com/intel-i7-5775c-broadwell-cpu-review/ and it only confirms what we already know of broadwell. It also adds more skepticism to this skylake report as you can see the haswell numbers do not align with what we already see.

Until i see official benchmarks from multiple sources, i just cannot give these tests any merit.

on page 2 I showed cinebench for haswell and broadwell (overclocked to the 4.2 that skylake runs at) for Apples to apples comparison. This bench is 99.999999% likely bs.

[patrickjp93]

I'm guessing most of the work went into the iGPU.

given it's the same graphics architecture as Broadwell (Gen 8/8.5), no. Furthermore, it's the Israel team that gave us Sandy Bridge. Hell, to, the, no. these benches are fake, and in 5 weeks we'll have our real answers. Seriously OP, put up a damn disclaimer. 

[DildorTheDecent]

I7 5775c @ 4.2 (1.36V)

Hmmmm. Smells of shit all around.

Something sure does.