The Core i7-4770K Review: Haswell Is Faster; Desktop Enthusiasts Yawn

[InvalidError]

I wouldn't have even released the Haswell desktop processors, but oh well.

With all the R&D already paid for, there is no reason to shy away from updating the whole product lineup if the new parts are superior in almost every way... except overclocking and heavy/full-load power.

I'm not surprised to see FIVR hurt Haswell's power-efficiency and overclocking. With the bulk of FIVR's losses being due to magnetics, there may not be much that Intel can do about it. One thing Intel can do something about though is work on increasing FIVR's input voltage to 5V which would reduce I2R losses on the motherboard VRM output and FIVR input by ~75%.

 

[icemunk]

HaZzzwell

 

[ubercake]

Again, stock boost speeds magically stuck at 3.9GHz. Why not 3.99? C'mon?!

 

[InvalidError]

This is what happens when monopoly rules. It wouldn't be a surprise if this chip is internally crippled on purpose.

Haswell has multiple legit reasons to behave the way it does and it essentially boils down to Intel planning their exit strategy from traditional PCs.

The number of people interested in enthusiast-class PCs is shrinking, PC sales in general are dropping because most people do not require anywhere near as much processing power as recent mainstream PCs offer (many people are still perfectly happy with their Core2Duo/Quad - I still was happy with my C2D until 8GB RAM wasn't enough anymore; now I have an i5 with 32GB RAM) so Intel has to look for non-traditionally-PC and new market segments to expand into. Most of those markets and form factors require low power and tight integration, which is exactly where Intel is going with Haswell and Broadwell.

For enthusiasts, things are likely to get worse before they get better.

 

[ericjohn004]

I gotta say, I'm impressed with the IPC improvements. Just look at Cinebench. I forgot the exact numbers but it's like 1.54 per core for Ivy and 1.75 per core for Haswell. That is a nice little increase right there. It looks as if the performance increase is greater from Ivy to Haswell as it was from Sandy to Ivy. Not to mention the theoretical performance in SiSoft Sandra. Granted, I always hated synthetics, but it's impressive nonetheless.

It seems as if from Sandy to Ivy we saw about a 3% per core IPC improvement, whereas from Ivy to Haswell there's about a 5% improvement. So overall, there's probably a good 6-10% improvement here. From Sandy to Ivy you only had about a 4-8% increase.

So I've gave my praises now lets get to the bottom line. Anyone who buys these parts are going to want to overclock. So far with this Intel design, as IPC increased, overclocking decreased. It went from an average MAX overclock on Sandy Bridge of 5.0Ghz, to 4.8Ghz on Ivy, and now 4.6Ghz with Haswell. So basically, if you have a Sandy Bridge, DO NOT upgrade as you won't notice any difference and may even regress. All of this is great for non overclockers though.

I know people who own Sandy Bridge have been sitting on their chip for two years and may want an upgrade. DON'T DO IT. It's not worth it at all. I have a 3570k @ 4.6Ghz gettings temps of 70c MAX on the hottest core with an H100i. I'd rather have a 2500k@4.8Ghz with MAX temps in the 60's personally. It's all the same.

Is it just me or were all of you expecting to see what these temps were when Tom's overclocked this processor? I was expecting a temp graph. Hopefully we get another review showing how well Haswell overclocks. As of now, it doesn't look too good.

 

[ubercake]

Great article. Saves me some cash. I'll be using my trusty 3930K for another year or two.

 

[Sakkura]

It seems as if from Sandy to Ivy we saw about a 3% per core IPC improvement, whereas from Ivy to Haswell there's about a 5% improvement. So overall, there's probably a good 6-10% improvement here. From Sandy to Ivy you only had about a 4-8% increase.

But Ivy Bridge was just a shrink of Sandy Bridge. Haswell is an entirely new architecture. It's with a new architecture that you expect the biggest performance improvements; instead, Haswell is only barely a bigger improvement than Ivy Bridge was.

 

[ericjohn004]

While Ivy Bridge was far from a huge upgrade from Sandy, it added:
-23% less power use
-10% clock-for-clock gain in performance
-USB 3.0
-PCI 3.0
-Intel 4000

That was a great deal better feature wise, and with good cooling my IB i5 reaches 4.7 GHz which easily beats any SB i5 up to 5.2 GHz (Good luck getting their!).


However Haswell has:

-9% MORE power used
-10% greater clock-for-clock performance
-Even LOWER overclocking headroom

FAIL!

I see this all the freaking time and I just don't understand it whatsoever. The per core IPC improvement from Sandy to Ivy was about 3%. The per core improvement from Ivy to Haswell seems to be about 4-5%. Which is a little more than what Ivy increased over Sandy. So I am somewhat impressed by this. 10% is such an exaggeration. That's twice as much as it actually is. But if you actually read and look at the benchmarks you'll see that in single core performance, the gain is usually no more than 5%. You MAY see 10-12% at the absolute MAX in some multithreading programs but that's just because it's using all four cores. If you gain 3% per core and have 4 cores, then you may see a gain up to 12% because 3x4=12. But 12% is close to the highest you'll ever see, this won't be the norm. NORMALLY, you going to see 5% gains from Sandy to Ivy or from Ivy to Haswell. The only exeption is synthetics like SiSoft Sandra that use the AVX2. But that performance gain is because of instruction sets, NOT IPC.

So the correct thing to say here would be "Haswell has a 4-5% IPC improvement over Ivy Bridge". So, now we know that an Ivy clocked 2.5Ghz over a Haswell, will be equal to a Haswell. And a Sandy that's clocked 2.5Ghz over an Ivy will be roughly equal to an Ivy. We also know that Sandy usually overclocks .2-.3Ghz higher than an Ivy and Ivy will overclock about .2-.3Ghz over a Haswell. So the bottom line is that from Ivy to Sandy to Haswell, we have seen exactly 0% performance increase.

But like someone else mentioned. With Ivy we got PCIe 3.0, HD4000, different instructions, lower power draw, and most everything by now had USB 3.0 and SATA III. So we got quite an upgrade overall. But with Haswell, everythings the same except we get a 4-5% IPC improvement, worse overclocking, HD4600, new instructions, and worse power draw. Definitely not a good upgrade. In fact, I don't know if I would upgrade if it was for FREE.

I'm just curious to see what the temps are on a 4670k. I'm surprised Tom's didn't include this.

 

[InvalidError]

But Ivy Bridge was just a shrink of Sandy Bridge. Haswell is an entirely new architecture. It's with a new architecture that you expect the biggest performance improvements; instead, Haswell is only barely a bigger improvement than Ivy Bridge was.

And IMO, this shows how difficult it is becoming to improve IPC and single-thread performance. Haswell has a bunch of major x86 core architecture enhancements (100% wider busses between caches, doubled associativity, 33% wider execution pipeline, 15% deeper reorder buffer, etc.) but all of them combined yield only a ~7% improvement on code that isn't optimized for new instructions.

The complexity of extracting more IPC is doubling several times faster than performance does. Unfortunately for people still hoping to see major IPC/single-thread performance breakthroughs in the future, that ship has already sailed deep into diminishing-return waters.

It isn't Intel being lazy; Intel is simply already extracting just about all the IPC that can possibly be squeezed out of typical x86 code within reasonable silicon, power and engineering budgets. AMD may be able to catch up with Intel's IPC but they won't be able to magically leapfrog them if/when they get there.

 

[InvalidError]

If you gain 3% per core and have 4 cores, then you may see a gain up to 12% because 3x4=12.

Flawed math.

If each individual core is 3% faster than the previous model's cores but you compare quad vs quad, the improvement is still only 3%:
- old core: 4x 100% = 400% theoretical throughput of a single core
- new core: 4x (100% + 3%) = 412% theoretical max
- (412 - 400) / 400 = 3%

 

[Lord Fartinghard of Eton-beens]

[citation][nom]TheMadFapper[/nom]How is this disappointing? You people act like IB/SB is slow. Did we really need a significantly faster processor? No. It's absolutely the way Intel should be going as they expand into the mobile area. I wouldn't have even released the Haswell desktop processors, but oh well.[/citation]

Here's what I've found in the forums.
http://www.neogaf.com/forum/showthread.php?t=573649

 

[Lord Fartinghard of Eton-beens]

[citation][nom]swordrage[/nom]I was thinking the same thing. I thought I somehow missed it, but no.. and how about cooling with water or blood?I'm interested in performance of a high end closed loop cooler like H100i (if they are compatible) or the likes. Any sources?[/citation]

Here's what I've found in the forums.
http://www.neogaf.com/forum/showthread.php?t=573649

 

[Lord Fartinghard of Eton-beens]

[citation][nom]TheMadFapper[/nom]How is this disappointing? You people act like IB/SB is slow. Did we really need a significantly faster processor? No. It's absolutely the way Intel should be going as they expand into the mobile area. I wouldn't have even released the Haswell desktop processors, but oh well.[/citation]
Sorry that reply is for someone else.

 

[ojas]

Well, on reevaluating the entire situation...here's what i've concluded so far:

1) I think we should chill the f*** out as far as performance is concerend. You're getting more performance than IVB at stock.

2) It can OC the about the same as IVB and has a better IPC. So even if it OCs a bit less, performance is the same, which is slightly disappointing.

3) The disappointment that everyone's feeling is probably because, as Sakkura said, it's a new microarchitecture, everyone probably hoped for a Nehalem to Sandy Bridge type jump in performance and efficiency, but instead this looked more like a tick in terms of performance.

4) The only other disappointments on the desktop have been higher load power consumption, and a degradation in QuickSync's performance and quality ( http://www.anandtech.com/show/7007/intels-haswell-an-htpc-perspective/8 )

5) All in all, not mind blowing, but was a fairly successful launch. Intel's focus this time has been on power consumption, and as InvalidError pointed out, it's commendable that they've managed to tweak it like they have, it's just unfortunate that they've not got too much out of the redesigns. It's possible software still has to catch up, though.

6) As someone else pointed out, probably the only way Intel can increase mainstream desktop performance is to stay stagnant on the GPU while increasing the resources thrown at the x86 side of things. Maybe if graphics gets good enough for Intel by the time Broadwell arrives, or Steamroller puts enough pressure on their i5s, they'll make i7s 6 or 8 core parts with Skylake.

7) I've read on Tech Report that future Pentiums/Celerons may have Silvermont cores instead of full Haswell cores, and if Airmont bridges the Core and Atom arch completely, this may well be true (BTW Intel sort of confirmed this http://techreport.com/news/24891/report-pentium-celeron-branded-atoms-to-power-android-devices ). In which case, it may be likely that Intel separates its Core branding from Atom/Celeron/Pentium completely in terms of performance, by limiting Pentiums to 4/8 Silvermont/Airmont cores (more likely 4), while starting Core i3s with 4 Skylake Cores, moving up to 8 for Mainstream i7s.

Tech Report has a separate article on Haswell OCing:
http://techreport.com/review/24889/haswell-overclocked-the-core-i7-4770k-at-4-7ghz

 

[ojas]

All indications point to overclocked Haswell processors requiring more aggressive cooling than their Ivy predecessors. The Core i7-4770K does have a higher TDP than the 3770K, but the associated heat is also spread over a larger die area. The 4770K's TDP per area works out to 0.47W/mm², while the 3770K's is 0.48W/mm². Haswell and Ivy seem to be on even footing in that regard. The die layouts follow the same basic blueprint, as well.

Haswell and Ivy Bridge also use a similar interface material between their dies and external heat spreaders. Intel used to employ a fluxless solder between those two pieces, but it switched to thermal paste with Ivy.

We don't have a definitive explanation for Haswell's apparent need for most robust cooling, but the chip's integrated VRM may play a role. Voltage regulation was handled off-chip in Ivy Bridge, but Haswell brings it—and the associated heat—onboard the die. Integrated voltage regulation is a big part of Haswell's appeal for mobile platforms. Unfortunately, it may also limit the processor's overclocking potential on the desktop.

-Tech Report

 

[mapesdhs]

10tacle writes:
> o/c SB 2500k. Sure, the 4770/1150 platform is a worthy upgrade to my SB,
> especially as one who wants to move up to i7 for video editing, ...

If video editing is your main task, a better upgrade would be SB-E, or
IB-E *if* it hasn't been hamstrung with the same silly cap material as IB.


Grandmastersexsay writes:
> ... Why didn't you try with an NH-D14? ...

One shouldn't have to use something as over the top as a D14 to oc a chip.
That's complete overkill. Even an old used TRUE is enough to run a 2600K
at 5GHz.


> ... All these things might need is some better thermal paste like Ivy Bridge.

*All* they need?? Presumably you're ok with trashing your warranty and the
much higher chance of damaging the chip compared to the comparatively easier
'enthusiast' task of lapping a chip.


> Haswell will be a great upgrade over a 775. ...

Of course it will, but so would a used 2500K, cheap TRUE, couple of fans,
4.7GHz in 3 minutes with a typical Z68 board. No way on earth haswell offers
anything really useful beyond this for the vast majority of users. And just
spend the amount saved on a better GPU or whatever.


> ... The $330 I7 4770K and $230 I5 4670K will be the fastest single threaded
> processor available, ...

The differences are not statistically significant most of the time. Merely
using an SSD or not could negate any visible difference. And since older CPUs
oc better, it's entirely possible overall performance with a Haswell oc'd using
the same parts would be worse.


> ... and at around
> half the cost of the $590 I7 3930K Sandy Bridge-E. ...

SB-E is an excellent chip *if* one can make use of the extra cores. And it
oc's very nicely too.


> ... As hyper threading really doesn't help with games, ...

That depends on the game/task; some games do benefit. For those that don't,
then again SB is a winner with the 2500K.


> ... Now is a great time to upgrade, ...

Sure, but I wouldn't bother with Haswell. I still think the sweetspot is
a used 2500K. Or even new if one can find them for a sensible price, but
of course Intel doesn't want its older chips being sold at all, or if they
are then the prices are really high despite being old. Are you suggesting
that the 2500K wouldn't absolutely fly off the shelves if it was still
available new for a sensibly low price? And imagine if one could by an
adjusted 2500K so it could fit into S1150? It's blatantly obvious that
Intel is dragging its heels now; reading your comments if feels like you're
praising Intel for winning a race in which they were the only competitor.


GIJoeMyGoodness writes:
> ... Ivy get's more done with a lower frequency and my i7-3770K is running
> 4.5Ghz on air. ...

To which I'm sure those with 2600K/2700K setups would say, *only* 4.5?

And that's the point. Without a very risky cap mod and no warranty,
SB oc's far better, and no need for crazy cooling.


tnagy writes:
> I have i7 920 OC to 3.8 should I upgrade?

That depends on what you're doing. What are your main tasks? Is it games?
Some are more CPU-dependent, others less so. Are are you running pro apps
or other non-gaming tasks? Do you benefit from lots of cores? Are you on
a restricted budget? Would you consider buying 2nd-hand for better value?
One can't answer a question like that without a lot more info.

Plenty of people will slag off your system for being 'old' in all sorts
of ways, but don't believe the hype. Unless your main task matches the
kind of upgrade you're considering, you may see a far lower performance
improvement than others will imply. If you don't already have one, the
best change to make in general is to fit an SSD (250GB Samsung 840 is good,
doesn't cost much).


ko300zx writes:
> ... but now I think I'll just grab a 3570k, and save the $100. ...

Or buy a used 2500K, it will oc much better, faster in overall terms;
spend the cash saved on a better GPU or SSD, etc.


TheMadFapper writes:
> How is this disappointing? ...

Because for a chip that requires a new mbd, new socket and uprated PSU,
it offers little or nothing over previous generation products.


> ... You people act
> like IB/SB is slow. ...

It isn't usefully *better*, that's the point, despite all the hype.


> ... Did we really need a significantly faster processor? ...

For a new socket/mbd/PSU, YES!


> ... I wouldn't have even released the Haswell desktop processors, ...

That indeed would have been a better approach, but Intel didn't do that;
instead they released a product which is expensive to adopt, offers
little in the way of gains over SB/IB, and with higher base pricing than
we've ever seen before (prices here are crazy). Very disappointing indeed.

I'm half-tempted to bag a newer AMD one of these days just to gain a feel
for something different as regards oc'ing, platform behaviour, etc., even
though I know it won't show any kind of comparative performance edge over
anything I already have (4.7 i3 550, 4.2 i7 870, 4.3 i7 875K, 5.0 2500K,
5.0 2700K, 4.7 3930K, 4.4 i5 760, some not yet setup including i5 670,
990X, another 870 and several more i3 550s). Just tired of waiting for
Intel to make something worthy of an enthusiast spend.

Ian.

 

[InvalidError]

7) I've read on Tech Report that future Pentiums/Celerons may have Silvermont cores instead of full Haswell cores, and if Airmont bridges the Core and Atom arch completely, this may well be true

Replacing the low-end by Atoms makes sense to me: much cheaper to produce than massively crippled i3 and it makes the performance slope between Pentiums and i3 much steeper.

Converging Atom with the high-end on the other hand, I don't think so.: the markets Atom and Core are addressing are far too different; Atom which is intended for applications calling for sub-watt TDP wherever possible (ex.: iWatch / Google Watch) cannot afford the substantial power and transistor budget a 10+W Core CPU spends on IPC.

I am expecting Atom and performance chips to remain largely separate with Atom picking up only the most power/footprint-efficient performance tweaks from the performance line - can't afford sacrificing 10% extra power unless it yields at least 10% more performance when on battery power since you likely want to achieve gains on battery life in the process.

 

[Sakkura]

All indications point to overclocked Haswell processors requiring more aggressive cooling than their Ivy predecessors. The Core i7-4770K does have a higher TDP than the 3770K, but the associated heat is also spread over a larger die area. The 4770K's TDP per area works out to 0.47W/mm², while the 3770K's is 0.48W/mm². Haswell and Ivy seem to be on even footing in that regard. The die layouts follow the same basic blueprint, as well.

Haswell and Ivy Bridge also use a similar interface material between their dies and external heat spreaders. Intel used to employ a fluxless solder between those two pieces, but it switched to thermal paste with Ivy.

We don't have a definitive explanation for Haswell's apparent need for most robust cooling, but the chip's integrated VRM may play a role. Voltage regulation was handled off-chip in Ivy Bridge, but Haswell brings it—and the associated heat—onboard the die. Integrated voltage regulation is a big part of Haswell's appeal for mobile platforms. Unfortunately, it may also limit the processor's overclocking potential on the desktop.

-Tech Report

Interesting. Guess that means the small process node can't really be blamed for any of this. It's about the VRMs and IGP.
Too bad Intel don't make a K SKU with the IGP lasered off; they only make them into locked SKUs (like the Core i5-3350P).

 

[InvalidError]

GIJoeMyGoodness writes:
> ... Ivy get's more done with a lower frequency and my i7-3770K is running
> 4.5Ghz on air. ...

To which I'm sure those with 2600K/2700K setups would say, *only* 4.5?

Ivy Bridge has 5-7% better IPC so you need to run SB 200-300MHz faster to match it, which negates most of the clock boasting.

Personally though, I can't get excited about anything significantly short of 100% so a 10-20% gain from overclocking is just not worth the trouble for me - ain't going to make or break anything I do enough to really notice.

 

[InvalidError]

Too bad Intel don't make a K SKU with the IGP lasered off

I doubt that would make any difference: with the IGP's power cut off straight from the FIVR, the IGP is already as good as not there TDP-wise.

And "lasering" it off does not physically remove the IGP, it merely blows programmable fuses to prevent people from enabling it. The IGP cannot be sawed off either since part of the reason why it is there in the first place is to provide extra die area to fan out CPU power and IO pads so they are far enough apart from each other to accommodate the pad pitch on the LGA carrier substrate and associated uBGA balls.

If the IGP could be sawed off, you would be throwing away a local heatsink/heat-spreader and that would likely make things at least somewhat worse.

 

[ahhyeah]

sdfsdfsdf

 

[ahhyeah]

Do people really upgrade/build their machines every 2 years? At the the 3 year point (when I generally upgrade/build a new machine), I wanted to get a Sandy, and then I said hell, I can wait for the Ivy. Right before pulling the trigger on the Ivy, I decided to hold out yet again! WTF?! After 6 years with my current machine (q6600), it's definitely time. I am not a gamer, so I was able to do so. I can't wait for a fresh new Haswell system and an SSD. I feel like waiting this long it will have been totally worth it!

 

[dragonsqrrl]

[citation][nom]ericjohn004[/nom].If you gain 3% per core and have 4 cores, then you may see a gain up to 12% because 3x4=12.[/citation]
...
[citation][nom]InvalidError[/nom]Flawed math.If each individual core is 3% faster than the previous model's cores but you compare quad vs quad, the improvement is still only 3%:- old core: 4x 100% = 400% theoretical throughput of a single core- new core: 4x (100% + 3%) = 412% theoretical max- (412 - 400) / 400 = 3%[/citation]
On behalf of the community, thank you.

 

[selfmade_exe]

TO ALL THOSE RESPONSIBLE FOR THIS TEST: I DARE YOU "CLEVER" TESTERS TO TEST AMD TRINITY WITH 1866 RAM OR ABOVE, AND ALSO COMPARE WITH RICHLANT FLAGSHIP. ALSO, PLZ RUN GPU_CAPS_VIEWER -> THE JULIA SET WITH GPU OPENCL AND COMPARE WITH TRINITY! PERIOD!

 

[soldier44]

Like others are saying disappointed its not the boost I was looking for over my 2600k at 4.5. Will just be happy with my recent GTX 780 buy and upgrade to 16gb of ram and see what next year brings to the table in CPUs.