Haswell model specs leaked

[Ben90]

Anybody have an idea how much energy the current VRMs on a mobo consume?

Unfortunately Intel botched it (for me and many others here) with exorbitant pricing on both the CPUs and the buggy platform. ($400 for motherboard effectively makes a $300 CPU a $500 one, no matter what reviewers say) The platform (x58) was also a furnace.

You could get a solid X58 board for $200 at launch, it just wasn't labeled "extremiest power overclockable board of ultra XXX"

 

[WhoBeDaPlaya]

You could get a solid X58 board for $200 at launch, it just wasn't labeled "extremiest power overclockable board of ultra XXX"

This. I purchased several Gigabyte EX58-UD3R + i7 920 D0 combos from Microcenter (drove my a*s all the way to the Minnesota store) for ~$350 each.

All clocked to 4GHz and have been running solidly 24/7 to this day. While I definitely lust after some of the niceties of "Ultra seXXXor" edition motherboards, they're hardly necessary.

 

[ShintaiDK]

Anybody have an idea how much energy the current VRMs on a mobo consume?

And more phases means more wasted. The answer is, quite alot.

 

[Lepton87]

Is Haswell supposed to be compatible with existing cooling solutions or is it not known yet? I'm thinking about buying Noctua NH-D14. Temps with my current cooler are terrible. I have Zalman CNPS 10X with IC DIAMOND TIM. Temps with noctua NT-H1 were comparable with fresh application then they shoot up over 20C over the course of 6 months. 60C at stock clocks. (27-28C ambient) Stock clocks with my mobo are 3.7GHz on every core even in IBT. 77C at 4.4GHz 1.35V and 70C at 4.4GHz 1.3V after 12h of IBT testing.Still I will never encounter such temps under normal usage but they are still high. Can I have a bad sample of the CPU or the cooler? Something like CPU die not soldered well to the IHS or heatpiepes making little contact with the base plate on the heatsink.

 

[inf64]

Mobile Haswell roadmap:
http://chinese.vr-zone.com/44251/ha...s-gt3-cpu-and-naming-rules-changing-12142012/

Looks like the roadmap also contains HD4600 GPU in it so original desktop roadmap seem legit. What's going on with the GPU clocks on the top mobile i7? 400Mhz vs 650Mhz on HD4000 for top Turbo clock? That is 62% lower while having 40EUs(or 20EUs?). It cannot be 20EUs since it would be massively slower than HD4000 so that means it's 400Mhz 40 improved EUs vs 16EUs @ 650Mhz equaling ~69% better GPU performance. Looks in-line with what intel claimed for top GT3 part.

 

[NTMBK]

Mobile Haswell roadmap:
Looks like the roadmap also contains HD4600 GPU in it so original desktop roadmap seem legit.

Given that they're both from the same source, it's hardly a resounding confirmation.

 

[Idontcare]

Given that they're both from the same source, it's hardly a resounding confirmation.

You know this is ALL fabrication because the model numbers are an overly simplistic extension of the existing numerology...something Intel has NEVER done in their branding iterations.

Even in the simple Sandy shrink to Ivy, Intel did not simply increment the family generation otherwise we'd be talking about i5-3500K and i7-3700K...but no, Intel had to also increment the third digit as well, it isn't the 3500K but rather it is the 3570K.

No way the product labels are real in these fictitious model specs, and that alone makes the purported model specs equally suspect.

I remember when the fake specs for Gulftown were being bandied about, everyone who wanted page hits pulled together "leaked roadmaps" detailing a never-ever-has-existed i9 CPU...and even then ALL the purported model specs were wrong - from TDP to core count to cache size to clockspeed.

It is always, always, complete bullocks with these leaks.

 

[NTMBK]

You know this is ALL fabrication because the model numbers are an overly simplistic extension of the existing numerology...something Intel has NEVER done in their branding iterations.

Even in the simple Sandy shrink to Ivy, Intel did not simply increment the family generation otherwise we'd be talking about i5-3500K and i7-3700K...but no, Intel had to also increment the third digit as well, it isn't the 3500K but rather it is the 3570K.

No way the product labels are real in these fictitious model specs, and that alone makes the purported model specs equally suspect.

I remember when the fake specs for Gulftown were being bandied about, everyone who wanted page hits pulled together "leaked roadmaps" detailing a never-ever-has-existed i9 CPU...and even then ALL the purported model specs were wrong - from TDP to core count to cache size to clockspeed.

It is always, always, complete bullocks with these leaks.

I entirely agree, as post #2 of this thread demonstrates.

 

[mikk]

What's going on with the GPU clocks on the top mobile i7? 400Mhz vs 650Mhz on HD4000 for top Turbo clock? That is 62% lower while having 40EUs(or 20EUs?).

Base clock is irrelevant for GPU performance. Dynamic frequency matters on a Intel iGPU. The Roadmap contains only GT2 models btw, according to vr-zone GT3 models will arrive in Q3.

 

[ShintaiDK]

Is Haswell supposed to be compatible with existing cooling solutions or is it not known yet? I'm thinking about buying Noctua NH-D14. Temps with my current cooler are terrible. I have Zalman CNPS 10X with IC DIAMOND TIM. Temps with noctua NT-H1 were comparable with fresh application then they shoot up over 20C over the course of 6 months. 60C at stock clocks. (27-28C ambient) Stock clocks with my mobo are 3.7GHz on every core even in IBT. 77C at 4.4GHz 1.35V and 70C at 4.4GHz 1.3V after 12h of IBT testing.Still I will never encounter such temps under normal usage but they are still high. Can I have a bad sample of the CPU or the cooler? Something like CPU die not soldered well to the IHS or heatpiepes making little contact with the base plate on the heatsink.

The mounting holes on LGA1150 are at the exact same place as LGA1155 and LGA1156 for that matter.

 

[hokies83]

You know this is ALL fabrication because the model numbers are an overly simplistic extension of the existing numerology...something Intel has NEVER done in their branding iterations.

Even in the simple Sandy shrink to Ivy, Intel did not simply increment the family generation otherwise we'd be talking about i5-3500K and i7-3700K...but no, Intel had to also increment the third digit as well, it isn't the 3500K but rather it is the 3570K.

No way the product labels are real in these fictitious model specs, and that alone makes the purported model specs equally suspect.

I remember when the fake specs for Gulftown were being bandied about, everyone who wanted page hits pulled together "leaked roadmaps" detailing a never-ever-has-existed i9 CPU...and even then ALL the purported model specs were wrong - from TDP to core count to cache size to clockspeed.

It is always, always, complete bullocks with these leaks.

So IDC what do you theorize Haswells per core performance per clock will be over IB / SB?

I have Also heard Haswell will use Tim under the IHS like IB due to the smaller core size and the solder cracking?

Tim can be used and work just as good if not better then solder if Intel does it right.

 

[Idontcare]

I have Also heard Haswell will use Tim under the IHS like IB due to the smaller core size and the solder cracking?

I have heard this as well (that the IHS will not be soldered).

As to the reasons, there are various plausible reasons but one would not expect the solder itself to crack.

The die itself cracking because of the rigidity of the solder causing efficient transference of the stress created by a mismatch in coefficients of thermal expansion between the silicon die and the metal IHS is entirely possible though.

(the very same mechanical forces that bring about the pump-out effect with the gell-like TIMs)

Tim can be used and work just as good if not better then solder if Intel does it right.

Unless the solder is just horribly applied, solder should always deliver superior results to any non-solder TIM owing to the difference in thermal conductivity.

That said, the lack of solder enables us enthusiasts to eliminate yet another set of interfaces by way of mounting our coolers to the bare-die, and in theory that should trump even a soldered-on IHS. But it would still be inferior to a soldered-on cooler

So IDC what do you theorize Haswells per core performance per clock will be over IB / SB?

My expectation is 12%.

 

[inf64]

I think those 7% from SB to IB are a bit too optimistic:
http://www.hardware.fr/articles/863-12/performances-frequence-egale-ddr3-2133-pci-express-3-0.html

Hardware.fr measured an average 3.7% increase in application performance with both CPUs @ 4.5Ghz with DDR3-2133. They saw outliers though,some workloads were as high as 7.7% faster but some were as low as 0% faster. Games saw similar gains (~3% average) with outliers such as SC2 with 5.9% jump in performance.

So I would take that chart above as not exact 1:1 ratio of projected to real performance gains. That being said I expect that Haswell will indeed have ~10%(or a bit better) IPC across the board. The changes in execution ports are just too large to not be felt in legacy workloads(FMA is another case where the jump will be massive).

 

[Smartazz]

Looking at history, how's this for a golden period BTW:

Apr-72: 8008, 200 KHz
Apr-74: 8080, 2 MHz
Jun-78: 8086, 10 MHz

That's a massive 5000% frequency increase in just 6 years.

http://www.intel.com/pressroom/kits/quickreffam.htm#i486

Could you imagine if that happened again how powerful even smart phones would be. I think that our modern transistors are toward the end of their life though.

 

[Idontcare]

I think those 7% from SB to IB are a bit too optimistic:
http://www.hardware.fr/articles/863-12/performances-frequence-egale-ddr3-2133-pci-express-3-0.html

Hardware.fr measured an average 3.7% increase in application performance with both CPUs @ 4.5Ghz with DDR3-2133. They saw outliers though,some workloads were as high as 7.7% faster but some were as low as 0% faster. Games saw similar gains (~3% average) with outliers such as SC2 with 5.9% jump in performance.

So I would take that chart above as not exact 1:1 ratio of projected to real performance gains. That being said I expect that Haswell will indeed have ~10%(or a bit better) IPC across the board. The changes in execution ports are just too large to not be felt in legacy workloads(FMA is another case where the jump will be massive).

The 7% gain is Sandy Bridge over Nehalem, not Ivy Bridge over Sandy Bridge.

The IB over SB gain is reported as being a mere 3.4% in the same graph that reports Haswell being 12% more than SB.

IB / SB = (305 + 116)/(305 + 102) = 421 / 407 = 1.034

3.4% or 3.7%, either way you must agree the results are congruent, yes?

 

[inf64]

The 7% gain is Sandy Bridge over Nehalem, not Ivy Bridge over Sandy Bridge.

The IB over SB gain is reported as being a mere 3.4% in the same graph that reports Haswell being 12% more than SB.

IB / SB = (305 + 116)/(305 + 102) = 421 / 407 = 1.034

3.4% or 3.7%, either way you must agree the results are congruent, yes?

Yes,you are absolutely correct. I don't know what I looked at when I saw the chart the first time . 3.4% looks spot on when compared to hardware.fr. I don't know the exact number for Nehalem to SB improvement but I suppose that it's also more or less spot on.

And yep,Haswell indeed looks like it's capable of up to 15% IPC jump,no doubt about that. It would be quite a feat since IvyB. is extremely well optimized core,especially from perf./watt angle and even at 32nm(SB variant).

 

[Puppies04]

3.4% or 3.7%, either way you must agree the results are congruent, yes?

 

[nsKb]

That said, the lack of solder enables us enthusiasts to eliminate yet another set of interfaces by way of mounting our coolers to the bare-die, and in theory that should trump even a soldered-on IHS. But it would still be inferior to a soldered-on cooler

A heatsink soldered directly onto the die should trump all as for heatsink+TIM on a bear die beating heatsink+TIM on a soldered IHS I would like to see what values you used for thermal conductivity.

 

[Concillian]

A heatsink soldered directly onto the die should trump all as for heatsink+TIM on a bear die beating heatsink+TIM on a soldered IHS.

He said exactly that in slightly different words.

 

[IntelUser2000]

Mobile Haswell roadmap:
http://chinese.vr-zone.com/44251/ha...s-gt3-cpu-and-naming-rules-changing-12142012/


Looks like the roadmap also contains HD4600 GPU in it so original desktop roadmap seem legit. What's going on with the GPU clocks on the top mobile i7? 400Mhz vs 650Mhz on HD4000 for top Turbo clock? That is 62% lower while having 40EUs(or 20EUs?). It cannot be 20EUs since it would be massively slower than HD4000 so that means it's 400Mhz 40 improved EUs vs 16EUs @ 650Mhz equaling ~69% better GPU performance. Looks in-line with what intel claimed for top GT3 part.

Mmm, try again.

400MHz versus 650MHz is 38.5% lower frequency. Oh, and that's Base clocks. Turbo clocks are the same.

Even IF the Base clocks were all there was...

400MHz x 20EU = 8,000 AUs(Arbitrary units)
600MHz x 16EU = 9,600 AUs

That's not massively slower. That's even assuming that there aren't any changes. I assume even if the AUs are lower, it would perform almost on par.

Intel's claims were also 15-25% for GT2 and 2x for GT3 over Ivy Bridge GT2. Also, HD 4600 is GT2 Haswell since its same as Desktop brand number.

 

[IntelUser2000]

A heatsink soldered directly onto the die should trump all as for heatsink+TIM on a bear die beating heatsink+TIM on a soldered IHS I would like to see what values you used for thermal conductivity.

That would be a stupid thing to do unless A) Intel is looking for future lawsuits B) The heatsink is flexible as rubber

If it contracts and expands the fragile die would break the first day you get your PC.

 

[IntelUser2000]

My expectation is 12%.

The numbers don't make sense.

Because Sandy Bridge is way greater than 10% faster compared to Nehalem/Westmere.

And single thread-wise, Nehalem/Westmere is not greater than what Sandy Bridge brought.

Lastly, Y being that high wouldn't make sense for the 2004 bar versus what looks like the next bar(Core 2).

I think the graph can only be summarized as following:

-Big gain from 2004 to 2006's Core 2 and small one for Penryn
-Another big gain from Nehalem and nothing for Westmere
-Big one from Sandy Bridge and small one for Ivy Bridge
-Big one from Haswell

Maybe its something that can't be quantified, because it wasn't meant to do that in the first place.

 

[jpiniero]

Because Sandy Bridge is way greater than 10% faster compared to Nehalem/Westmere.

I think that's about accurate. What I think is throwing you off is that the 2500/2600 replaced products at lower clock speeds. That's not happening with Haswell.

 

[Idontcare]

The numbers don't make sense.

Because Sandy Bridge is way greater than 10% faster compared to Nehalem/Westmere.

And single thread-wise, Nehalem/Westmere is not greater than what Sandy Bridge brought.

Lastly, Y being that high wouldn't make sense for the 2004 bar versus what looks like the next bar(Core 2).

I think the graph can only be summarized as following:

-Big gain from 2004 to 2006's Core 2 and small one for Penryn
-Another big gain from Nehalem and nothing for Westmere
-Big one from Sandy Bridge and small one for Ivy Bridge
-Big one from Haswell

Maybe its something that can't be quantified, because it wasn't meant to do that in the first place.

The numbers are from Intel and based on tested applications.

I suppose they could have just made up the results and graphed them, but I am inclined to take them at face value as being legitimate.

The only thing they obfuscated was the absolute value of the y-axis, its offset. The y-axis obviously does not start at zero, it is some value above zero. In terms of absolute pixels the best fit value there was 305.

I'm familiar with looking at graphs like this because I use to make them myself, sanitizing existing engineering data and obfuscating the axis when placing it in the public domain.

Once you know how the real thing is morphed into what you end up seeing, it is rather straightforward to undo the morphing. Which is one of the arguments we engineers are constantly telling the communication team when we are required to sanitize our charts for publication, but they never listen to us so we have to do it anyways. But anyone who wants to know what the chart originally looked like can figure it out pretty easily.

 

[RussianSensation]

My expectation is 12%.

The IPC increases between those generations look too low to me from actual reviews. Those numbers you posted might come from Intel but I doubt they cover as many programs as professional reviewers do. The general consensus on our forums for years has been about a 20% increase in IPC for Nehalem (although I see it's more now in the latest apps) and 14-16% for SB over Nehalem. Reviews back this up.

Sandy Bridge --> Nehalem is not 7%, but 15-16%
Core i5 760 2.8ghz = 100%
Core i5 2500K 2.8ghz = 115%
or
Core i7 860/930 2.8ghz + SMT = 100%
Core i7 2600K 2.8ghz + SMT = 116%
http://www.computerbase.de/artikel/prozessoren/2011/test-intel-sandy-bridge/46/

Secondary source confirms that it's around that range:
i5 2500K 2.8ghz is faster than i5 760 2.8ghz by 14%
http://ixbtlabs.com/articles3/cpu/sandybridge-core-vs-lynnfield-p2.html

Nehalem/Lynnfield was actually a much larger increase in IPC over Kentsfield/Yorkfield than SB is over Nehalem.

"5 - 10% increase in general application performance at the same clock speeds as Penryn. Where Nehalem really succeeds however is in anything involving video encoding or 3D rendering, the performance gains there are easily in the 20 - 40%"
http://www.anandtech.com/show/2658/20

Using the same source at Computerbase:

Core 2 Quad Q9550 2.8ghz = 100%
Core i5 760 2.8ghz = 122%
Core i7 930/870 2.8ghz = 127%

So about 22-27% without/with HT for Nehalem. This is somewhat skews since Kentsfield/Yorkfield perform very poorly in encoding and rendering and neither has HT, which can swing the average greatly in multi-threaded apps.

If Haswell IPC increases just 10%, it would be the worst increase in IPC from a new Intel architecture in 5 years. Although if they fix the solder issue, 10% IPC and 5.0ghz+ easy overclocks for Haswell on air without delidding, it wouldn't be that bad, but obviously as others have noted the pace of CPU performance at Intel has slowed down tremendously. If they don't fix the solder at all, it'll be a yawn as a 10% increase in IPC alone is nothing to write home about for existing SB/IVB users.

I'll even go as to say that for a lot of people people rocking i7 860/920 @ 3.9-4.0ghz+, there is still very little reason to upgrade the CPU for games outside of MMOs (WOW) that are insanely CPU dependent or if you happen to have 2 high-end GPUs where CPU bottleneck shows up. Otherwise, you are just wasting $. GPU bottlenecks will only be exacerbated as next generation consoles launch and games made for DX11 from the ground-up are made. Far Cry 3 is going to be nothing compared to next gen engines and by that point cards like GTX680 will be choking to death and it won't matter at all if you have a 4ghz Nehalem or a 10ghz Haswell.

http://www.guru3d.com/articles_pages/far_cry_3_graphics_performance_review_benchmark,7.html

The main draw for Haswell will probably be improvements in very specialized programs (Monte Carlo simulation due to Random Number Generator) or some other programs that heavily benefit from new instructions. For the general public gaming or using office apps, it will be a worthless speed upgrade given the cash outlay for a new mobo+CPU. I can see enthusiasts upgrade as usual as they enjoy playing with new parts and OCing, but like SB/IVB were over Lynnfield/Nehalem, I don't expect any earth shattering performance increases over Core i7 2600K OC, maybe not even over i7 920 @ 4.0ghz. Even the Z87 chipset isn't looking to bring any new cool features. Z77 also has Thunderbolt if you want that and PCI Express/DDR4 are still years away. LGA2011 also looks lame as IVB-E is the only thing it has going by Q3 2013. All the excitement of PC upgrades now rests in the GPU and SSD space.