Haswell model specs leaked

Discussion in 'CPUs and Overclocking' started by boxleitnerb, Dec 11, 2012.

  1. mikk

    mikk Golden Member

    Joined:
    May 15, 2012
    Messages:
    1,632
    Likes Received:
    10

    Depends on the game.

    Trine 2 3770k HD4000

    DDR3-1600= 34,1
    DDR3-2400= 39,7

    Or here some more: http://www.hardwarecanucks.com/foru...-intel-i7-3770k-ivy-bridge-cpu-review-21.html


    Expected was a 95W TDP, so I'm surprised that people are upset from a 84W TDP. The CPU core itself is ~15% bigger: AVX2, FMA, TSX, more instructions units, two new ports, integrated FIVR. Furthermore a bigger iGPU with new VQE in QS and dx11.1 support. In fact this 84W is lower than expected.
     
  2. TuxDave

    TuxDave Lifer

    Joined:
    Oct 8, 2002
    Messages:
    10,576
    Likes Received:
    0
    If you want to look at physiological barriers, you can first start by looking at the turbo frequency of chips to see what's possible without touching the design. Mind you the absolute physiological barrier is higher if you could go back into the chip and actually change the physical device sizing etc... (and the worst case is to go into the uArch). So there's no real physiological barrier on why the frequency is where it is.

    I mentioned it earlier, it's just because there's better places to spend your thermal budget (and power delivery budget!)
     
  3. Fjodor2001

    Fjodor2001 Diamond Member

    Joined:
    Feb 6, 2010
    Messages:
    3,395
    Likes Received:
    0
    Well, as mentioned in my previous post, the CPU frequency increased from 233 MHz to 2.0 GHz in 4 years from 1997 to 2001. I.e. it increased about 1000 %.

    Do you really think we could repeat the same journey in the next 4 years if the thermal budget would be spent differently? I.e. we'd be at 35 GHz in 2016. I don't think we'll be anywhere near that, regardless of thermal budget decisions or TDP increases.

    So clearly the CPU frequency has topped out compared to how things progressed during 1985--2005. And clearly there must be a physological reason for it, otherwise we would have seen much larger CPU frequency increases from one CPU generation to the next during the last 4-6 years.
     
    #78 Fjodor2001, Dec 11, 2012
    Last edited: Dec 11, 2012
  4. Concillian

    Concillian Diamond Member

    Joined:
    May 26, 2004
    Messages:
    3,713
    Likes Received:
    0
    And what happened to TDP? What's the increase if it's normalized to CPU only power budget?

    That was in an era where CPU only power budget exploded, and the architectures receded in terms of any measurable computations per MHz metrics. We're now in an era where CPU power budget is shrinking every generation to allow for higher GPU power budgets and increased integration of motherboard components and metrics for compute performance per MHz are improving. Of course frequency per time is huge in your example time-frame, but measurable compute efficiency (computational output per power input) is a significantly more meaningful metric, and I'm willing to bet my house that if you used that metric the growth from your example time-frame will look significantly less favorable compared with the last (or next) 4 years.
     
    #79 Concillian, Dec 11, 2012
    Last edited: Dec 11, 2012
  5. TuxDave

    TuxDave Lifer

    Joined:
    Oct 8, 2002
    Messages:
    10,576
    Likes Received:
    0
    You could repeat that journey to 20GHz if you want. 50ps cycle time? Don't know how much logic you'll get in between each stage but devices definitely work at that speed. So if you want clock speed (and had no thermal concerns) you do exactly the opposite of what I mentioned. You start removing performance features to gain performance via clock speed. I'm pretty sure at the 20GHz range you'd be far worse off than before.

    So as I mentioned, there is no physiological (as in it's impossible to physically build something at that speed) barrier up to 20GHz. Sure the dies have to shrink by a ton to get the clock grid churning at that speed but yes, you can physically build something really dumb at 20GHz. But instead when thermal budgets crashed the party, the whole idea of performance per watt came into play where scaling up frequency turned out to be one of the worse tradeoffs you can make.
     
    #80 TuxDave, Dec 11, 2012
    Last edited: Dec 11, 2012
  6. Lonyo

    Lonyo Lifer

    Joined:
    Aug 10, 2002
    Messages:
    21,939
    Likes Received:
    0
    There have been various articles on AT (the website) about cores vs speed.
    Search function is such that I can;t think of how I would hope to find any, but the basic gist from what I remember is... adding more cores is a much easier way of improving performance (albeit in multi threaded situations).

    Increasing clock speed increases power consumption a lot more than adding more cores does (e.g. 5GHz vs 2x2.5GHz, ideal world perfect scaling etc from a performance standpoint, 2x2.5GHz uses less power).

    Also while clock SPEED may not have increase, IPC/performance per clock has increased, so 4GHz now is faster at tasks than 4GHz 5 years ago.

    As to why GHz makes power use go up, not 100% on the specifics, IDC might be a good bet, or hunting through old AT articles.


    Found what might be the main ones (based on titles...)
    http://www.anandtech.com/show/1611
    http://www.anandtech.com/show/1645




    Another aspect of clockspeeds is pipeline length, which also impacts IPC:
    http://www.anandtech.com/show/495/2
    Longer pipeline = higher frequencies = more penalty for misses = lower IPC.
     
    #81 Lonyo, Dec 11, 2012
    Last edited: Dec 11, 2012
  7. TuxDave

    TuxDave Lifer

    Joined:
    Oct 8, 2002
    Messages:
    10,576
    Likes Received:
    0
    Assuming all else equal, the dumb (but accurate) way of answering that is looking at activity factor. Higher frequency means more electrons are travelling from power to ground per second and so without changing anything else except frequency, higher frequency --> more electrons moving around --> more energy used.

    There are (sadly) other effects that cause the power vs clock frequency to not be as linear as that.
     
    #82 TuxDave, Dec 11, 2012
    Last edited: Dec 11, 2012
  8. Idontcare

    Idontcare Elite Member

    Joined:
    Oct 10, 1999
    Messages:
    21,130
    Likes Received:
    0
    Pentium II 233MHz reportedly used 34.8W on 350nm.

    Pentium 4 2.0GHz reportedly used 71.8W on 180nm.

    So power basically doubled while clockspeeds increased 8.6x, all from a mere 2 nodes of shrinking.

    That is actually rather remarkable considering how little is gained from a process node nowadays.

    For comparison, on 22nm if you triple the power budget for the 3770k it buys an extra 1.3GHz of clockspeed (3.5GHz -> 4.8GHz).

    [​IMG]
     
    #83 Idontcare, Dec 11, 2012
    Last edited: Dec 11, 2012
  9. Lepton87

    Lepton87 Platinum Member

    Joined:
    Jul 28, 2009
    Messages:
    2,499
    Likes Received:
    0
    Xeon X5698 It was actually 4.4GHz I remembered it wrong, but it's still the fastest X86 stock clocked CPU that has ever been released. From what I remember it was OEM only so you couldn't just buy it.
     
  10. Torn Mind

    Torn Mind Platinum Member

    Joined:
    Nov 25, 2012
    Messages:
    2,723
    Likes Received:
    1
    No, Hz measures "pulses" not speed. Its units is 1/second. To get speed, you need an actual variable in the numerator, i.e distance or instructions.

    Yeah, that Pentium 4 2.0 Ghz gets ABSOLUTELY destroyed by the paltry-looking Celeron G4xx series.

    That Pentium 4 at about 3.0 GHz would also offer performance comparable to Dothan processors clocked at 1.86 Ghz. Pentium 4 actually is the reason the processor makers realized there is a physics wall in regards to increasing clockspeed.


    The point is, using clockspeed as a measure of performance ONLY works with processors that have the exact same microarchitecture. Start comparing processors of different microarchtectures, and you gain absolutely NO information as to which processor performs better.
    A Prescott 3.2 GHz or a Celeron G550 going at 2.6 Ghz? The answer is the Celeron because it offers superior Instructions per second, which is derived from multiplying IPC by clockspeed.

    http://en.wikipedia.org/wiki/Clock_rate
     
    #85 Torn Mind, Dec 11, 2012
    Last edited: Dec 11, 2012
  11. Concillian

    Concillian Diamond Member

    Joined:
    May 26, 2004
    Messages:
    3,713
    Likes Received:
    0
    You need to show the other version... the one with the arrows for +14% at the same power consumption and the -2x% power at the same clock speed.

    That's still not the whole story though...

    P4 < P3 in IPC.
    P4 at 2 GHz is more like a 1.5 GHz or so there's another % chunk less.
    By the same token, for any reasonable performance metric IB at the same speed is (slightly) better at the same clock speed as SB. Clock speed is a red herring. What matters is usable performance compared to power required.

    So in terms of actual performance you go from 266 --> 1500 ish (~5.7) while power is approximately double, so your metric of performance over TDP increase is really only ~2.8x.

    This is far less impressive than the iniital 266 vs. 2000 MHz comparison. It's world class marketing level manipulation in order to make current CPU advancements look bad compared to "the good ol' days."

    The same comparison of 4 years ago vs. IB at release:
    i7 940 vs. i5 2400 (Cinebench R10 single thread from bench) = 1.3x on newer core & very slightly higher speed...
    TDP 130 vs. 77 ==> 0.6
    1.3 / 0.6 ==> 2.2x
    That's even with very little MHz difference between the two CPUs... and the IB has close to 50% of the die dedicated to GPU... so 77 is probably too high to be using for the TDP in that comparison.... suddenly we're really close to the 2.8x that we saw from 1997 --> 2001... maybe even higher

    CPU advancements haven't slowed as much as people want to believe. The issue is that "we" (as a society) don't need faster CPUs for most tasks, so the "advancement budget" is going somewhere other than CPUs. If faster CPUs were as necessary as they were in 2000 for normal tasks, you can bet we'd still be seeing huge advancements in CPU capability.
     
    #86 Concillian, Dec 11, 2012
    Last edited: Dec 11, 2012
  12. Torn Mind

    Torn Mind Platinum Member

    Joined:
    Nov 25, 2012
    Messages:
    2,723
    Likes Received:
    1
    Hardcore clockspeed increases began in the Pentium 4 era, and Intel even was trying to get 7 Ghz with Tejas and Jayhawk. They stopped and came to their senses though. If they were released, Intel would have been seriously burned.

    http://en.wikipedia.org/wiki/Tejas_and_Jayhawk

     
  13. lopri

    lopri Elite Member

    Joined:
    Jul 27, 2002
    Messages:
    11,590
    Likes Received:
    1
    Correct. Northwood -> Hammer -> Conroe -> Sandy are the only meaningful jumps in IPC improvements for me. Admittedly there was a multi-core revolution during the Hammer/Conroe era so I can cut some (large) slacks for that. But there hasn't been anything interesting to me since the original Nehalems and Thubans. Sandy is fast, to be sure, but is kind of boring. Ivy is.. ugh. Bulldozer? LOL.
     
  14. hokies83

    hokies83 Senior member

    Joined:
    Oct 3, 2010
    Messages:
    835
    Likes Received:
    0
    No that sux XD i have 120hz Monitor XD

    I can resale my 1155 stuff and 3770k for top dollar why not go 4770k?
     
  15. IntelUser2000

    IntelUser2000 Elite Member

    Joined:
    Oct 14, 2003
    Messages:
    4,057
    Likes Received:
    1
    This isn't true. That's because the the 266MHz is a Pentium II. The Pentium 4 gets behind with the on-die 256KB L2 Pentium III's, but I doubt it'll be much behind a off-die L2 cache Pentium II. With a 66MHz FSB which is half the Pentium III as well!

    Here's what I think Intel did the past couple of years with frequency and power.

    Intel's tricks to clock speed increase at "same power"

    Penryn to Nehalem:

    Trick #1: Notebooks went from 25W SV chip + 10W MCH to 35W combined. Because the MCH has harder time reaching TDP than CPU cores do, it gained "free clock speed" due to thermal headroom of the extra few watts that the MCH doesn't use.

    #2: Turbo! While the peak power usage stayed the same, it increased average power use at higher(but not peak) loads by use of Turbo. Since it can clock down to Base when necessary, its "free" again.

    Nehalem to Sandy Bridge:

    -Power Tricks: Things like the Physical Register File and uop cache was used for decreasing power, rather than increasing performance. That allowed Base frequency to go up, while Turbo stayed nearly the same.

    Sandy Bridge/Ivy Bridge to Haswell:

    -My predictions are the increase in TDP is to increase Turbo frequency close to the max Turbo frequency. Right now with 3770K, it goes 3.5/3.6/3.7/3.8/3.9GHz. In 4770K, it might be 3.5/3.8/3.8/3.8/3.9GHz. That means in scenarios where lot of cores and threads are active, there may be frequency gains.
     
  16. IntelUser2000

    IntelUser2000 Elite Member

    Joined:
    Oct 14, 2003
    Messages:
    4,057
    Likes Received:
    1
    I think even HD 4000 can benefit from faster memory and things like the on-package DRAM. But early documents suggest the absolute performance improvement may be slightly less than you suggest. The pure FLOPs per EU doesn't change, meaning the changes elsewhere are to bring up its weakness rather than boost peak. For example, Intel states that in Haswell, the graphics unit texture performance in some cases may go up to 4x, which will help in texture bound games.

    Early leaks suggested 15-25% for GT2 and 2x for GT3. If that's for mobile desktops are probably better as mobile Ivy Bridge clocks relatively better against mobile Haswell than on the desktop comparisons do.

    Think 20-30% instead. Bottleneck isn't all or nothing though. Because it varies from application to application, scenario to scenario and even frame by frame! The fact that Ivy Bridge's HD 4000 scales going to DDR3-1866 means its somewhat bottlenecked, but not as much. Perhaps it'll become sensitive as fastest desktop Llano does.
     
    #91 IntelUser2000, Dec 11, 2012
    Last edited: Dec 12, 2012
  17. Khato

    Khato Golden Member

    Joined:
    Jul 15, 2001
    Messages:
    1,003
    Likes Received:
    2
    It is indeed a question of what the actual bottlenecks in HSW graphics are. I'm still hoping that it might fix whatever bottlenecks are causing IVB to go from being competitive against Trinity on some games to being approximately half as fast on others. 'Cause if that bottleneck is gone it's a lot more than just a 20-30% performance gain that could be seen... Sadly, I doubt such will be the case.
     
  18. boxleitnerb

    boxleitnerb Platinum Member

    Joined:
    Nov 1, 2011
    Messages:
    2,596
    Likes Received:
    1
    Nehalem improved IPC also considerably if you're looking at todays applications and games. About 30% or so.
     
  19. meloz

    meloz Senior member

    Joined:
    Jul 8, 2008
    Messages:
    320
    Likes Received:
    0
    What happened to this forum? Why are so many (erroneously) co-relating Haswell's 84 watt TDP with the 'stagnant' clock it shares with IVB, and then acting as if the sky is falling?

    By themselves, neither the TDP nor clockspeed says anything about performance unless we know about the IPC, and we do not know the IPC. This is all the more true since Haswell is a 'Tock', a new microarch and not a shrink+ like IVB.

    With a new arch, it is entirely possible to increase performance while staying 'stagnant' at a given frequency and core count.

    And some people are even complaining about a paltry 7 watt increase in TDP on a desktop part? What happened to performance-per-watt, if system performance has increased correspondingly, we have nothing to complain about.

    Besides, with Haswell Intel are moving more and more stuff to CPU from MB, so it is better to compare platform power consumption. Which will be at least 20% lower (for desktops) according to Intel's own presentation at IDF. So again, why the tears?

    I would be more than happy with a 'mere' 10% performance increase on CPU side, for us Linux users Haswell is all about iGPU anyway. That's where I have higher expectations. 10% better CPU, 25% better iGPU along with an overall 20% lower energy consumption is a solid upgrade (as long as Intel do not jack up the prices).

    I strongly urge members to read: Intel's Haswell Architecture Analyzed: Building a New PC and a New Intel


    [​IMG]


    The one thing that raises some doubt about accuracy of this chart is the presence of "4600" series iGPU on all chips. This is unlike Intel, they like to segment. Could they be using "4600" name as a placeholder until they are prepared to reveal more? :confused:

    But then again, I note that the 'K' SKUs have VT-D disabled in typical Intel fashion. So maybe this chart is authentic after all! (har har)
     
  20. ShintaiDK

    ShintaiDK Lifer

    Joined:
    Apr 22, 2012
    Messages:
    20,071
    Likes Received:
    13
    Same with the BGA socket rumour. Seems alot of people on this forum search for drama that aint there.
     
  21. Idontcare

    Idontcare Elite Member

    Joined:
    Oct 10, 1999
    Messages:
    21,130
    Likes Received:
    0
    Folks are bored, so the mundane and the absurd get equal billing as bonafide content.
     
  22. Fjodor2001

    Fjodor2001 Diamond Member

    Joined:
    Feb 6, 2010
    Messages:
    3,395
    Likes Received:
    0
    Turn it around. Why are so many here vigorously defending the low CPU performance increases we've seen during the last 4 years, compared to how things progressed all the way from 1985 to around 2003 (and a one time jump in 2006 when transitioning to new uArch with Conroe)?

    Do you have to create some hallelujah spirit to motivate purchasing Haswell? I don't think so. It has plenty of benefits such as better iGPU, integrated VRM, and lower power consumption for some use cases (mainly applicable to ultrabooks/laptops/etc).

    But when it comes to pure CPU performance increases... sorry, it's nowhere near what we saw going from one CPU generation to the next during the golden years. Is that so hard to admit? :confused:
     
  23. Torn Mind

    Torn Mind Platinum Member

    Joined:
    Nov 25, 2012
    Messages:
    2,723
    Likes Received:
    1
    So, you still clinging to the megahertz myth? There's a reason benchmarks are used to guage performance instead of clockspeed: benches are more accurate, even the inconsistent Passmark is more useful than seeing 1.6 Ghz except when talking about processors with identical microarchtectures.

    The gains of the Pentium 4's Netburst archtecture were not astronomically high compared to its Pentium III predecessors, and lower clocked Pentium Ms would provide the same performance as a highly clocked Pentium 4.

    Four years ago, Nehalem was just released in November and offered substantial performance increases over Conroe while also lower power consumption. Now we are at Ivy Bridge, where the i5s are unequivocally faster than the QXXXX series.
     
  24. Smartazz

    Smartazz Diamond Member

    Joined:
    Dec 29, 2005
    Messages:
    6,128
    Likes Received:
    0
    We didn't see the biggest jump with Sandy Bridge, yet it's considered a great architecture.
     
  25. Fjodor2001

    Fjodor2001 Diamond Member

    Joined:
    Feb 6, 2010
    Messages:
    3,395
    Likes Received:
    0
    You don't need to tell me Hz are not everything. But still, it matters. A lot. And going from PII 266 MHz -> P4 2.0 GHz we did see an IPC increase too. In addition to an 8.6x CPU frequency increase.

    Going from Yorkfield->IB we've perhaps seen more IPC increase. But only 2.83 GHz (Q9550) -> 3.5 GHz (3770K) => 1.23x frequency increase. So the IPC increase going from Yorkfield->IB is nowhere near enough to compensate for the higher frequency increase we saw going from PII->P4.

    To sum it up: Show me benchmarks where the relative CPU performance increase going from PII 266 MHz (1997) -> P4 2.0 GHz (2001) is lower compared to going from Q9550 (2007/2008) -> 3770K (2012), then I'll believe you.