kaby lake Intel 7th Gen 7700K + others leaked in Sisoft Sandra database

[79wjd]

6 minutes ago, JoeyDM said:

Devils canyon wasn't just binning, it was swapping the TIM and adding more capacitors. Idk about kaby lake, but I doubt it's binning. We'll see though. There isn't a ton of concrete info on Kaby lake.

I didn't realize they added capacitors. But those were added due to haswells inconsistent overclocking and the TIM (like I said before) was upgraded because haswell had temp issues. Neither of those are problems with skylake, so I can't imagine what else they would do other than bump clocks/improve binning. Afaik there are no major issues with skylake (other than whatever bullshit happened with too much mounting pressure from coolers causing damage).

[Unhelpful]

1 minute ago, djdwosk97 said:

 

Yeah. But I don't want to make speculation about Kaby Lake, since so little info is out on it. But I'm with you, I can't think of what any tweaks will be.

[CommandMan7]

I'm still on Ivy Bridge, but I don't think that it'll last me much longer than the next 2 years. My I7-3770 is struggling in Ashes of the Singularity, paired with my 1500 Mhz G1 970. On the benchmark's heavy batches, the game is something like 40% CPU bottlenecked, and it's holding my framerate below 60 FPS no matter how low I put the graphics. I think that most people will be looking to upgrade to either Zen Plus or Cannonlake, as 14nm is not going to offer enough performance increase.

[Tam3n]

Looks like Sandy Bridge is still going to last a long time...

[NumLock21]

3 hours ago, djdwosk97 said:

I didn't realize they added capacitors. But those were added due to haswells inconsistent overclocking and the TIM (like I said before) was upgraded because haswell had temp issues. Neither of those are problems with skylake, so I can't imagine what else they would do other than bump clocks/improve binning. Afaik there are no major issues with skylake (other than whatever bullshit happened with too much mounting pressure from coolers causing damage).

Skylake still use tim. Devil's Canyon was a different cpu to that of haswell refresh.

[79wjd]

4 minutes ago, NumLock21 said:

Skylake still use tim. Devil's Canyon was a different cpu to that of haswell refresh.

Yes, but the TIM isn't an issue in Skylake like it was in Haswell. And the only thing differentiating DC and Haswell Refresh is the extra capacitors to make overclocking more consistent between chips. 

[patrickjp93]

15 hours ago, djdwosk97 said:

Considering Kaby Lake is supposed to just be a Skylake Refresh (like Devils Canyon was for Haswell), it seems very likely that it will have to have a base clock higher than 3.6ghz (for obvious reasons). 

It's an optimization step, so there are architectural tweaks. It's not the exact same architecture. Further, totally new GPU architecture.

 

10 hours ago, deviant88 said:

Intel is going down at this rate, at least in the desktop pc market...

Anyway, 7700k= 6700k= 4790k = 3770k= 2700k

We havent had a new intel arhitecture since 2011...

Oh good grief. If that was true the HPC crowd would still be on Sandy or Ivy Bridge. The problem is software developers, not Intel. Instruction latencies are, for the most part, as low as they are going to get. SISD performance has reached its asymptotic ceiling. When Zen catches up, you'll find AMD can do no better and innovate at no faster a rate. Even ARM, IBM, and Oracle can't do better by any significant degree. We must move on to SIMD and multithreading and heterogeneous acceleration in consumer workloads or we stagnate. We have OpenMP. We have CilkPlus. We have compiler intrinsics to make SSE and AVX programming pretty much a snap. Heck we even have metaprogramming for getting pre-computed results at compile time which are applicable to a lot of game physics if you know a few tricks with hash tables and bloom filters. When you go from manipulating 1 integer or float at a time to 2 (MMX and SSE 1), 4(AVX 1), 8 (AVX 256), or now 16 (AVX 512, which is confirmed for Cannonlake and speculated on Kaby Lake), your performance can in fact multiply by 2, 4, 8, or 16.

 

Fewer instructions needed means smaller loops which means better use of small loop detectors that can avoid the instruction fetch stage which decreases latency and increases throughput. It also means having more room in instruction cache and denser cache lines which affects performance more than any other factor in a modern system. And then as we crunch data more quickly, RAM speed will begin to matter again, as it should.

 

I can make a 6700K do twice as much work as you could possibly get out of a 2600K at the same clock speed. It's not Intel's fault that consumer software developers are too low-skill to learn about these systems.

[patrickjp93]

4 hours ago, CTR640 said:

I truely forgot when 6700K has been released and now they are already talking about 7700K.

Kinda too fast for almost zero improvements?

 

My problem with Intel is they keep asking sacks of money for very few small performance and AMD just won't hurry.

If consumer software developers kept up with multithreading and heterogeneous acceleration, you'd change your tune. the age of SISD computing has grown long in the tooth. Intel has provided both the hardware and software tools to provide an escape. The horses merely need drink (OpenMP, CilkPlus, SIMD instructions). The 6700K can do more than twice as much work per clock compared to the 2600K. One merely needs know more than the basics in programming to see it.

[patrickjp93]

7 hours ago, Energycore said:

Cool. Not worth upgrading from sandy bridge yet. If anything I'd upgrade to something with more cores for compute stuff.

7700K:   4 cores * 4.2*10^9 cycles/second/core * 2 ops/cycle  * 256bits of width/32 bits per datem = 268.8GFlops

               24 EUs * 8SPs/EU * 2ops/SP/cycle* 1.150*10^9 cycles/second                                         = 441.6GFlops  (710.4GFlops total)

2600K:   4 cores * 3.4*10^9 cycles/second/core * 2 ops/cycle  * 128bits of width/32 bits per datem = 108.8GFlops

               Non-programmable GPU with no compute capability                                                          = 0 GFlops       (108.8GFlops total).

 

That's not even accounting for the better branch prediction; larger, faster caches, better memory controllers, and all else that affects performance. And even though the units are flops, this applies to integer operations as well.

[Commander Llama]

36 minutes ago, patrickjp93 said:

I can make a 6700K do twice as much work as you could possibly get out of a 2600K at the same clock speed. It's not Intel's fault that consumer software developers are too low-skill to learn about these systems.

Game programmers in particular are soooooo lazy. And to appropriate your comment on SISD in one of those posts up there, most game engines are "long in the tooth" too. They're old ass engines from single/dual core era that just keep getting little flashy bits and pieces slapped onto them, and looking at pretty much all modern games and their performance, it shows.

 

From a gaming point of view, though it applies to very few people here since most of us use discrete GPUs, Intel has stepped up their GPU game pretty amazingly. I used to try running games on the iGPU of a 2600K and it's a fucking potato (D2 and HOMM3 worked though... mission accomplished). The iGPU on my Surface Pro can run freaking modern games, aside from the most demanding titles. Really huge jumps in performance and that's, as you say, with terrible programming standards.

[porina]

39 minutes ago, patrickjp93 said:

7700K:   4 cores * 4.2*10^9 cycles/second/core * 2 ops/cycle  * 256bits of width/32 bits per datem = 268.8GFlops

2600K:   4 cores * 3.4*10^9 cycles/second/core * 2 ops/cycle  * 128bits of width/32 bits per datem = 108.8GFlops

Assuming the above for 7700k applies to 6700k (and even Haswell), it is something I hadn't considered before. Does Sandy Bridge "only" work on 128-bits at a time? I see ~50% throughput per clock boost (where not otherwise limited) from Haswell in Prime95-like tasks, which is FPU intensive. I had attributed it to the presence of FMA which is sued. Then again, presumably FMA and the 256-bit data are not really separable so it isn't too meaningful to attempt to do so.

[patrickjp93]

2 minutes ago, porina said:

Assuming the above for 7700k applies to 6700k (and even Haswell), it is something I hadn't considered before. Does Sandy Bridge "only" work on 128-bits at a time? I see ~50% throughput per clock boost (where not otherwise limited) from Haswell in Prime95-like tasks, which is FPU intensive. I had attributed it to the presence of FMA which is sued. Then again, presumably FMA and the 256-bit data are not really separable so it isn't too meaningful to attempt to do so.

Sandy Bridge only covers SSE 4 and AVX 1 (128-bit). You can work with 2/4 ints/floats or 1-2longs/doubles at a time.

 

Prime95 uses a specific algorithm to get both the dedicated FPU and the Vector processor hot simultaneously. It's not actually a useful computational algorithm of any sort. It's just designed to make things hot.

[LAwLz]

It's really sad that the most exciting new thing about this launch is support for 10bit HEVC in hardware... Something we can already do on our phones.

 

I really hope Intel pulls their head out of their ass and gives us a 6 core on the mainstream socket. They could just move everything down one tier and all of a sudden everyone would get super excited for the next generation.

i3 = quad core

i5 = quad core with HT

i7 = 6 core with HT

 

Then the enthusiast platform could be 8, 10 and 12 cores.

 

Right now I am sitting with a 5 year old CPU and can't even be bothered to upgrade to their current mainstream platform. I will most likely be getting the i7-6800K.

[porina]

32 minutes ago, patrickjp93 said:

Prime95 uses a specific algorithm to get both the dedicated FPU and the Vector processor hot simultaneously. It's not actually a useful computational algorithm of any sort. It's just designed to make things hot.

Prime95 was born to do a job. The clue is in the name, to find prime numbers (of a specific type). Its other function as a stress test is a by product of that. The core routines are designed to multiply large numbers together as efficiently as possible, which is the most computationally intensive part of proving number is prime or not. The main part of that involves FFTs, which are commonplace in engineering applications, although not normally of the size involved with prime finding. The gwnum library that contains this code is also used by others in LLR and PFGW for example, which can test for prime numbers of other formats. The high level proof methods vary, but fundamentally it requires multiplying massive numbers together and that's what it has been optimised to do well.

[patrickjp93]

10 minutes ago, porina said:

Prime95 was born to do a job. The clue is in the name, to find prime numbers (of a specific type). Its other function as a stress test is a by product of that. The core routines are designed to multiply large numbers together as efficiently as possible, which is the most computationally intensive part of proving number is prime or not. The main part of that involves FFTs, which are commonplace in engineering applications, although not normally of the size involved with prime finding. The gwnum library that contains this code is also used by others in LLR and PFGW for example, which can test for prime numbers of other formats. The high level proof methods vary, but fundamentally it requires multiplying massive numbers together and that's what it has been optimised to do well.

That was the original purpose, but now all it is is a CPU torture test. There are vastly superior to what it uses for actually finding prime numbers. Bit pattern aggregates give clues far more helpful in tracking down leads and eliminating possibilities of factors.

[DildorTheDecent]

All about the 5GHz dream though. 

[patrickjp93]

23 minutes ago, DildorTheDecent said:

All about the 5GHz dream though. 

Maybe when we change materials at 7nm, but I think for silicon it's over. The high-leakage bulk processes are gone, and FinFET, even when combined with FDSOI design principles, is still going to have problems stretching above 4.8GHz for the most part.

[porina]

54 minutes ago, patrickjp93 said:

That was the original purpose, but now all it is is a CPU torture test. There are vastly superior to what it uses for actually finding prime numbers. Bit pattern aggregates give clues far more helpful in tracking down leads and eliminating possibilities of factors.

There's a lot of people participating in the search for prime numbers (not just with Prime95), and if there is any better way, you can be sure it will be used! Prime95 and related software is the last step in the process. There is also a lot of other work to eliminate candidates in a more time efficient manner, but those elimination methods don't find primes, they just eliminate non-primes relatively quickly. At some point, you still need to test if a number is prime or not. CPU wise, Prime95 and similar implementations are still the fastest way to do that final test. 

[Clyne]

TFW when using some CPU that has the IPC of a Q9550 or so...

But MULTI-THREADED beats the 6600K!

I'm just going to wait to Vega(GPU) and see what CPUs are around that time...

[DildorTheDecent]

53 minutes ago, patrickjp93 said:

Maybe when we change materials at 7nm, but I think for silicon it's over. The high-leakage bulk processes are gone, and FinFET, even when combined with FDSOI design principles, is still going to have problems stretching above 4.8GHz for the most part.

Not with that attitude. 

[handymanshandle]

When I build my first PC this summer, I'll be using a 6600K. The only thing that might make me want to upgrade would be Zen. Though I'm not going to hold my breath for Zen or Kaby Lake. Though I would consider upgrading to a Kaby Lake or Cannonlake i7 if I wanted the extra cores in the future but for the time being certainly wouldn't wait for Kaby Lake for an incremental upgrade unless it meant I would get an i7 instead of an i5 and needed the extra performance gained by the i7 in non-gaming stuff.

 

[Energycore]

3 hours ago, patrickjp93 said:

7700K:   4 cores * 4.2*10^9 cycles/second/core * 2 ops/cycle  * 256bits of width/32 bits per datem = 268.8GFlops

               24 EUs * 8SPs/EU * 2ops/SP/cycle* 1.150*10^9 cycles/second                                         = 441.6GFlops  (710.4GFlops total)

2600K:   4 cores * 3.4*10^9 cycles/second/core * 2 ops/cycle  * 128bits of width/32 bits per datem = 108.8GFlops

               Non-programmable GPU with no compute capability                                                          = 0 GFlops       (108.8GFlops total).

 

That's not even accounting for the better branch prediction; larger, faster caches, better memory controllers, and all else that affects performance. And even though the units are flops, this applies to integer operations as well.

It won't give me a better framerate upgrade than getting a new graphics card.

[79wjd]

9 minutes ago, Energycore said:

It won't give me a better framerate upgrade than getting a new graphics card.

Not everything is about gaming. 

 

Plus it's not Intel's fault that that CPUs are significantly ahead of GPUs in terms of gaming. 

[Energycore]

1 minute ago, djdwosk97 said:

Not everything is about gaming. 

This is my personal opinion. It doesn't interest me because I'm not using it for the kind of stuff that benefits from the additional CPU power. If games start using it more, i'm onboard. But on a limited budget upgrading the graphics makes more sense.

[Thony]

3.6-4.2 that's so dissapointing unless it eats half the energy as previous gen.

 

My Haswell refresh i7-4799 is 3.6-4.0

 

Cannonlake better launch i7 6core mainstream line or I'm gonna flip tables ! We have been stuck with 4 core i7 for far too long now.  

 

I mean there would be no i7 4core at all  just 6 for consumer and 8+ for prosumer.