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

[Memnarchon]

EVERYONE. ABOUT THE HEAT.

Found this link on the TR forums, links to an AT forum post:
http://forums.anandtech.com/showpost.php?p=34053183&postcount=570

The dude found that the TIM isn't the problem, but there's a small gap that Intel leaves between the IHS and the CPU. Eliminating that gap leads to huge reductions in CPU temp, keeping the TIM constant.

The funny thing is that Intel's TIM seems to be even better than Noctua's TIM. :O
But they still leaving a 0.06mm gap in order to keep away non-enthusiast people from o/cing...

 

[ojas]

Ian Cutress of AnandTech just posted this absolutely epic article.

Compares a huge, HUGE combination of motherboards, CPUs and multi-GPU configs, and Haswell and Piledriver (including Thuban) have shown pretty strong results.

He's gone into the details of how a CPU effects stuff like game performance, performed a lot of real-world tests, CPU bottlenecks, etc.

It's really interesting so far (only a few pages in, but there are a lot of pages!).

http://www.anandtech.com/show/6985/choosing-a-gaming-cpu-at-1440p-adding-in-haswell-/

 

[ojas]

EVERYONE. ABOUT THE HEAT.

Found this link on the TR forums, links to an AT forum post:
http://forums.anandtech.com/showpost.php?p=34053183&postcount=570

The dude found that the TIM isn't the problem, but there's a small gap that Intel leaves between the IHS and the CPU. Eliminating that gap leads to huge reductions in CPU temp, keeping the TIM constant.

The funny thing is that Intel's TIM seems to be even better than Noctua's TIM. :O
But they still leaving a 0.06mm gap in order to keep away non-enthusiast people from o/cing...

I don't know about that. It costs them zero to reduce the gap, and they're shooting themselves in the foot by allowing higher temps.

Don't know what's up with that.

 

[cangelini]

That power consumption graphs don't seem to be right... if I do Wh / AWP for 4700K it is more than 2, for 2700K it is less.. does it mean that 2700K was faster? I don't see it in other graphs...

Going back over the logs, I discovered one extra number run for Haswell than the other systems. Looks like Vegas should have been rem'ed from the batch, but wasn't when the suite was run. The new chart over time and average power don't look very different, but the Wh measurement is now equal with -3770K. Thanks for noticing this.

 

[InvalidError]

7-8% is a pretty small improvement in the historical perspective. Particularly when it's supposed to be the meaty tock in a tick-tock cadence of improvements.

Frankly, you should be impressed that they managed to improve IPC at all.

With Haswell, the CPU has to look 192 instructions ahead (up from 168 on SB/IB) to find enough stuff to do to produce that ~7% improvement. That requires a considerable amount of work and execution resources. The CPU can only do so much speculative execution before dependencies and mispredictions become ultimate limits for single-threaded IPC. The closer you get to that ultimate limit, the more costly getting any closer gets - this is true for just about everything in physics and technology too.

Architectures designed for high-performance computing (GPUs, UltraSparc T1/2/3/4, Power6/7/8, etc.) have for the most part given up on single-threaded performance and use multi-threading to keep their execution resources busy when individual instruction flows stall on dependencies instead of relying heavily on costly (power+area+wasted cycles) prediction and speculative execution.

 

[mkw34]

So it's more of tweak than a tock or tick?

 

[hixbot]

Cangelini, isn't the benchmark on the first page (Itunes IPC) proof that IPC only improved about 2%?
Since clock rate isn't fixed in the other benchmarks, couldn't the improvement measured be that they are operating at different clocks at the time of the bench (due to different turbo/thermal characteristics at that specific moment)?

 

[InvalidError]

So it's more of tweak than a tock or tick?

It is a significant overhaul architecturally-speaking (15% deeper re-order buffer, 40% wider execution pipeline, 100% faster caches, etc.) but Intel is already so close to the absolute maximum IPC/ILP that can be extracted from individual threads of typical x86 code that all those enhancements only delivered a modest performance improvement.

If Intel fails to produce major IPC improvements, it certainly isn't from lack of trying - they increased the size of some key structures by 20-100% to produce that ~7% improvement.

When you approach practical limits, things simply have a tendency to get a whole lot more difficult.

 

[mkw34]

The new architecture is geared for mobile users. Performance is not as changed as I thought it would be.

 

[bildo123]

Intel made the Sandy bridge because of the phenom X6, without competition from AMD Intel will make small improvements every generation to make more money 🙁

Tell me where this money will come from if the product is so marginally insignificant/different it produces little to no sales? See, this release wasn't meant to be a generational gap where everyone with a Sandy Bridge would upgrade. The sales in this product line is mostly if not all going to be from new systems unless you're the type that must have the latest CPU. Unless the consumer is stupid enough to always upgrade their CPU every single release and (deservingly) gets milked your whole "small improvements to getz da profitz" conspiracy is non-sense.

 

[InvalidError]

The new architecture is geared for mobile users. Performance is not as changed as I thought it would be.

The only part geared toward mobile is FIVR. Every other change is still a massive effort to push IPC harder than ever before without busting power and area budgets.

If Intel had designed Haswell entirely for Mobility's sake, they would have axed resources rather than significantly increase them - simplest way to significantly reduce power without sacrificing much performance when you are already this far into the diminishing-return zone.

 

[mad tech]

so glad i didn't w8 for haswell and bought ivybridge and got more out of my time instead of w8ing for the crap work of lazy intel

 

[slomo4sho]

so glad i didn't w8 for haswell and bought ivybridge and got more out of my time instead of w8ing for the crap work of lazy intel

Regardless of generation selection, Intel appreciates your support. Thank you for reinforcing Intel's business decisions with your $$$. You sure showed them!

 

[InvalidError]

Regardless of generation selection, Intel appreciates your support. Thank you for reinforcing Intel's business decisions with your $$$. You sure showed them!

lol.

With that train of thought, buying AMD chips supports AMD's "business decision" to be over two generations behind Intel CPU performance wise.

By buying into Ivy Bridge instead of Haswell, he is voting with his cash on Ivy Bridge being the better value at this point in time. If enough people choose to do the same (however unlikely that is), Intel would eventually have to admit people did not support Intel's business decisions - somewhat like how end-user backlash ended up forcing Microsoft to admit more people cared about the Start Menu than they thought they did.

Of course, avoiding Haswell will become increasingly difficult once Intel starts discontinuing Ivy Bridge models 6-8 months from now.

 

[hannibal]

And Intel will gets money from Ivy and get rid of old stock... Intel would thank you for that and increase the price, because they don't have to sell those Haswels too cheap because there would not be competition even from itself...
IMHO buying old products from Intel does not help the situation. Only competition from outside could do it.
We can only hope that AMD will sell so much APU:s that it will become profitable and then in time maybe get some upper class CPUs out...

 

[sarinaide]

If trinity is already better in the GPU right now over haswell....what's gonna happen when kaveri ships in Q4?

GT2 isn't intended to be a particularly serious IGP; it is mostly whatever fits in the extra die area Intel needs to add to the CPU cores anyway to fit all the uBGA balls under the die.

GT3/3e on the other hand does increase die area specifically to accommodate extra IGP resources and GT3e benchmarks do show much better IGP performance - at least in games/benchmarks that don't hit some (hopefully driver) quirks.

Maybe you should tell him of the GT3e fallacy, ie the one that a $650 non upgradeable BGA setup can get mercilessly destroyed by a fuddy duddy APU + 7790 at half the cost, then you say add the 7790 to the 4770R, but then you have a 4770K +7790 for cheaper and faster. Some experts have been really critical of IRIS and it lies in that Intel spent a massive fortune on Iris and its price point renders it not appealling at all. Then of course you have Kaveri in the midst so Iris may end up being the biggest of the failures not from a performance point of view but from a positioning point of view, basically has not application due to its excessive price mark.

On mobile where it is expected to do well again for $200 extra you can get yourself a i7 with discrete Nvidia or AMD higher end GPU's that will literally decimate Iris as a value part. Said this before anything came out, once the $600+ price tag got leaked 2-3 months ago, I said any performance will be mitigated.

Richland offers bigger x86 gains than iGPU gains and still remains the more appealing integrated solution.

 

[Sakkura]

Maybe you should tell him of the GT3e fallacy, ie the one that a $650 non upgradeable BGA setup can get mercilessly destroyed by a fuddy duddy APU + 7790 at half the cost, then you say add the 7790 to the 4770R, but then you have a 4770K +7790 for cheaper and faster. Some experts have been really critical of IRIS and it lies in that Intel spent a massive fortune on Iris and its price point renders it not appealling at all. Then of course you have Kaveri in the midst so Iris may end up being the biggest of the failures not from a performance point of view but from a positioning point of view, basically has not application due to its excessive price mark.

On mobile where it is expected to do well again for $200 extra you can get yourself a i7 with discrete Nvidia or AMD higher end GPU's that will literally decimate Iris as a value part. Said this before anything came out, once the $600+ price tag got leaked 2-3 months ago, I said any performance will be mitigated.

Richland offers bigger x86 gains than iGPU gains and still remains the more appealing integrated solution.

Pricing is always crucial. But at least Intel have now shown that they can produce competitive graphics. That's a first. Now it looks like they need to improve their drivers and pricing.

 

[ingtar33]

On mobile where it is expected to do well again for $200 extra you can get yourself a i7 with discrete Nvidia or AMD higher end GPU's that will literally decimate Iris as a value part. Said this before anything came out, once the $600+ price tag got leaked 2-3 months ago, I said any performance will be mitigated.

actually you're missing the point in an attempt to make it.

In mobile you can get an i7+650m for $100 less then the i7 w/ iris. The 650m will beat the iris in games/FPS, the iris will beat the i7+650m in power draw.

So for +$100 you get an inferior part with less power draw.

Pricing is always crucial. But at least Intel have now shown that they can produce competitive graphics. That's a first. Now it looks like they need to improve their drivers and pricing.

Actually i think what Iris proves is you can get a lot of GPU power by brute forcing a solution, by using super expensive non-upgrade-able ram like an L4 cache... and in the end, ending up with a part too expensive to sell outside of an Apple store.

Without apple, this is a part without a market... anywhere. And it never would have made it off the designing room floor. This is a half-baked technology, that needs a pile of refinement before we see it anywhere practical or affordable.

 

[sarinaide]

Iris is a beast of process but leaves not much more for Intel to squeeze without ballooning or sacrificing efficiency. Iris is marketed as a performance part to take on the 650M but the problem it ran into is that the 650M is still faster, since a average joe is not going to need iGPU gains they will settle on a cheaper HD4600 or lower option, GT3e is targeting gaming and thus fails at the first hurdle brute force could not beat a low end mobile part and so the sentiment remains for gaming note book i5 or i7 + discrete may cost around the same but be a lot faster. On desktop since IRIS is BGA and basically locked in the 4770R is slower than the 4770K and costs more than double, then you factor in a 4770K + Low end discrete still beating Iris comfortably at low cost and the picture is not a good one. Iris was not well thought out, it was just Intel butthurt at IGPU claims so they made the beast that will watch AMD A-Series out sell it by healthy margins.

Back onto stocks of 4770's received on pre order, already 3 of 25 were RMA'd, they exhibited odd behaviour and temps before dying completely, I have heard a lot of them are being RMA'd. There was rumour of Intel 14nm parts melting and then came the news that broadwell was pushed to 2015 before an article about it being scrapped and skylake pushed sooner amidst more rumours of intel thermal woes chasing legacy x86 design to minute size. the 4770K we tested did 1.275v achieved 4.5 max stable and a mind blowing 87 degrees centigrade which is just not appealing at all.

So now my question is, is Haswell to intel what Zambezi was to AMD. I think its different, Zambezi was a radical departure from traditional design and wasn't expected to be amazing from the get go, it was a teething process and in a Vortex review, 45 benches the 8120 beat the 1100T in 34 of them and operated at lower power, Piledriver is almost a complete sweep and shows AMD's process is improving. On the intel side this is a mature arch now and the 3770K wins in a lot of benches and is more power efficient, is failwell intels faildozer? or is the situation more critical as Intel in 3 generations have hardly moved IPC's or performance despite a die shrink, is AMD's gamble on future systems architecture yet to pay dividends while Intel drives traditional design into bedrock.

 

[Raghar]

I'm bit curious about how Haswell would work with my benchmark. Can you do a comparison with Ivy and Haswell at the same frequency, without boost, HT enabled then HT disabled?

Download Java here:
http://java.com/en/download/manual.jsp

Download benchmark from here:
https://mega.co.nz/#!QJgXSTjL!berfU1KkBqBqrtTSQA1-Uh762A4yIVWgMvZYOyghkoY

Run pack.cmd, pack4C.cmd, and pack8C.cmd with hyperthreading.
Run pack.cmd, and pack4C.cmd without hyperthreading.

It creates HTML files with: "times in ns, hash" lines. It prints a long column because I wanted to test OS spikes as well. Well, I guess posting last 5 lines from each tests should be enough for a CPU only synthetic benchmark.

I hope I didn't did some boo boo in this benchmark, but because I'm making a GUI version that creates a single score + deviations, rechecking old version is unnecessary. As long as the original version will not change, the results should be usable anyway.

 

[rrbronstein]

I think people are WAY underestimating this new 4770k. If I didn't already have a 3570k I'd definitely get a 4670k(I much rather a QUAD core without HT because of the extra heat and it overclocks less and takes more voltage, not because of the money). From the looks of things, Ivy increased about 3-4% IPC over sandy in single threaded workloads. Haswell looks like it increased about 5-7% over Ivy. In some benchmarks the increase from Ivy to Haswell is twice as much as from Sandy to Ivy. In one of these benchmarks the 4770k scores 1100 while the 3770k scores 1285(lower score is better), that's about a 15% increase! And in Cinebench 11.5 the 4770k's single core score was 1.75 while the 3770k score was 1.54, that's about an 8% increase, per core. And did you see those SiSoft Sandra benches? The 4770k looks like a beast right there. Granted, I don't like synthetics as they don't normally show real world performance. But imagine programs that take advantage of the faster cache and AVX2 instructions, the 4770k would be beast.
Overall, by looking at every score, I think the 4770k is about 7-8% faster than Ivy during most workloads while Ivy was only maybe 5% faster than Sandy. And whether you like it or not, 7-8% is a very good improvement. It's about 50% more performance than we got from Sandy to Ivy.
I mean what were you expecting to see? Some 15-20% improvement? When you already have a groundbreaking chip it's hard to make it even better, and they did that.
The bad part is, is that it overclocks around .2Ghz worse than Ivy from the looks of things. This should take away about 5-6% of the 8% that it gained. So overall, you should still see a 2-3% performance gain. And that's not taking into account the programs that may be 15% faster like the one I noted earlier.
So we are getting a better chip here. And I'm betting that because of that nice little jump in single threaded performance that games will run faster too. From Sandy to Ivy FPS increased by maybe 3%, I imagine that from Ivy to Haswell FPS will increase by 5%. If you look at it in terms of FPS, 5% can be a lot.
So as you can see, I'm fairy impressed with the new Haswell chip, taking everything into consideration. Some of you just read what Tom's has to say and don't even make your own opinion.

Nice positive thoughts man, im also in agreeance, i am not disappointed by haswell at all, and although i wouldve taken ivy at any moment, haswell is a just a further improvement i dont get why people sit here and bitch about intel always making improvements, and although this chip has mostly mobile improvements, its still amazing engineering feats that we should all be proud of. Who the hell needs more CPU performance anyways, ivy gives enough, haswell didnt need huge IPC gains.

 

[rrbronstein]

Iris is a beast of process but leaves not much more for Intel to squeeze without ballooning or sacrificing efficiency. Iris is marketed as a performance part to take on the 650M but the problem it ran into is that the 650M is still faster, since a average joe is not going to need iGPU gains they will settle on a cheaper HD4600 or lower option, GT3e is targeting gaming and thus fails at the first hurdle brute force could not beat a low end mobile part and so the sentiment remains for gaming note book i5 or i7 + discrete may cost around the same but be a lot faster. On desktop since IRIS is BGA and basically locked in the 4770R is slower than the 4770K and costs more than double, then you factor in a 4770K + Low end discrete still beating Iris comfortably at low cost and the picture is not a good one. Iris was not well thought out, it was just Intel butthurt at IGPU claims so they made the beast that will watch AMD A-Series out sell it by healthy margins.

Back onto stocks of 4770's received on pre order, already 3 of 25 were RMA'd, they exhibited odd behaviour and temps before dying completely, I have heard a lot of them are being RMA'd. There was rumour of Intel 14nm parts melting and then came the news that broadwell was pushed to 2015 before an article about it being scrapped and skylake pushed sooner amidst more rumours of intel thermal woes chasing legacy x86 design to minute size. the 4770K we tested did 1.275v achieved 4.5 max stable and a mind blowing 87 degrees centigrade which is just not appealing at all.

So now my question is, is Haswell to intel what Zambezi was to AMD. I think its different, Zambezi was a radical departure from traditional design and wasn't expected to be amazing from the get go, it was a teething process and in a Vortex review, 45 benches the 8120 beat the 1100T in 34 of them and operated at lower power, Piledriver is almost a complete sweep and shows AMD's process is improving. On the intel side this is a mature arch now and the 3770K wins in a lot of benches and is more power efficient, is failwell intels faildozer? or is the situation more critical as Intel in 3 generations have hardly moved IPC's or performance despite a die shrink, is AMD's gamble on future systems architecture yet to pay dividends while Intel drives traditional design into bedrock.

I highly doubt haswell is fail. Its right at launch, so there could be a few RMA's and initial problems with beginning batches but as always it will improve over time. Haswell is beast and you cant deny that. We dont need huge IPC gains as ivy is powerful enough, we know most of the gains are in the mobile sector and thats fine. You have to give intel credit for all the work they have done to try and compete with ARM(still a long way off though), as AMD is no longer a key high performance competitor until they really reveal something special. Your paragraph is just super negative and i dont understand why, I do hope AMD can come back to a performance segment to match intel, but we just have to wait and see.

 

[InvalidError]

On mobile where it is expected to do well again for $200 extra you can get yourself a i7 with discrete Nvidia or AMD higher end GPU's that will literally decimate Iris as a value part.

Except GT3e will destroy i5/i7 mobile + discrete GPU on battery life while still providing more than adequate 3D acceleration for most uses, which is exactly what most people buying ultrabooks are likely to care the most about.

GT3e is not a solution for every market/pocket but for premium devices where high performance is desired or required while power, space and TDP are heavily constrained (ex.: ultrabooks and tablet-PCs), GT3e may make the most sense out of all solutions available regardless of its crazy pricing.

I know if I was seriously considering buying into the ultrabook hype, I would much prefer to have a larger battery filling the space freed up by using GT3/GT3e to provide decent 3D performance instead of a discrete GPU with associated RAM, VRMs, heatsink(s), fan(s), heat and battery power draw - no point in wasting space and power on a discrete GPU in a machine I would end up using mainly for business / non-gaming purposes.

 

[sarinaide]

Iris is a beast of process but leaves not much more for Intel to squeeze without ballooning or sacrificing efficiency. Iris is marketed as a performance part to take on the 650M but the problem it ran into is that the 650M is still faster, since a average joe is not going to need iGPU gains they will settle on a cheaper HD4600 or lower option, GT3e is targeting gaming and thus fails at the first hurdle brute force could not beat a low end mobile part and so the sentiment remains for gaming note book i5 or i7 + discrete may cost around the same but be a lot faster. On desktop since IRIS is BGA and basically locked in the 4770R is slower than the 4770K and costs more than double, then you factor in a 4770K + Low end discrete still beating Iris comfortably at low cost and the picture is not a good one. Iris was not well thought out, it was just Intel butthurt at IGPU claims so they made the beast that will watch AMD A-Series out sell it by healthy margins.

Back onto stocks of 4770's received on pre order, already 3 of 25 were RMA'd, they exhibited odd behaviour and temps before dying completely, I have heard a lot of them are being RMA'd. There was rumour of Intel 14nm parts melting and then came the news that broadwell was pushed to 2015 before an article about it being scrapped and skylake pushed sooner amidst more rumours of intel thermal woes chasing legacy x86 design to minute size. the 4770K we tested did 1.275v achieved 4.5 max stable and a mind blowing 87 degrees centigrade which is just not appealing at all.

So now my question is, is Haswell to intel what Zambezi was to AMD. I think its different, Zambezi was a radical departure from traditional design and wasn't expected to be amazing from the get go, it was a teething process and in a Vortex review, 45 benches the 8120 beat the 1100T in 34 of them and operated at lower power, Piledriver is almost a complete sweep and shows AMD's process is improving. On the intel side this is a mature arch now and the 3770K wins in a lot of benches and is more power efficient, is failwell intels faildozer? or is the situation more critical as Intel in 3 generations have hardly moved IPC's or performance despite a die shrink, is AMD's gamble on future systems architecture yet to pay dividends while Intel drives traditional design into bedrock.

I highly doubt haswell is fail. Its right at launch, so there could be a few RMA's and initial problems with beginning batches but as always it will improve over time. Haswell is beast and you cant deny that. We dont need huge IPC gains as ivy is powerful enough, we know most of the gains are in the mobile sector and thats fine. You have to give intel credit for all the work they have done to try and compete with ARM(still a long way off though), as AMD is no longer a key high performance competitor until they really reveal something special. Your paragraph is just super negative and i dont understand why, I do hope AMD can come back to a performance segment to match intel, but we just have to wait and see.

AMD doing things differently to Intel, and I wouldn't say that they are not competing, its not like Intel just run away with every parameter, its just as above AMD got radical and changed their approach to system designs then you factor in that AMD do what they do on millions of dollars not tens of billions of dollars. Life is easy at Intel corp, throw billions into a massive process and after billions are wasted release something that people may deem adequate. In sunnyvale its a little more complicated.

I was just merely commenting on the parallels between Haswell and Zambezi, since the 3770K wins in a lot of benchmarks the only wins for Haswell seems to be on the IMC and iGPU fronts, and its faster in itunes which is important over here. Other than that its like pulling teeth to overclock and runs hot, some complained about Ivy running hot boy are they in for a surprise. Then there is the fact the 4770 runs on quite a bit more power than the 3770K I would term Haswell more akin to a side grade than a upgrade, to many parallels to Zambezi and since that was met with a lot of criticism I don't see how this is worthy of different because it bares a blue sticker.

On mobile where it is expected to do well again for $200 extra you can get yourself a i7 with discrete Nvidia or AMD higher end GPU's that will literally decimate Iris as a value part.

Except GT3e will destroy i5/i7 mobile + discrete GPU on battery life while still providing more than adequate 3D acceleration for most uses, which is exactly what most people buying ultrabooks are likely to care the most about.

GT3e is not a solution for every market/pocket but for premium devices where high performance is desired or required while power, space and TDP are heavily constrained (ex.: ultrabooks and tablet-PCs), GT3e may make the most sense out of all solutions available regardless of its crazy pricing.

I know if I was seriously considering buying into the ultrabook hype, I would much prefer to have a larger battery filling the space freed up by using GT3/GT3e to provide decent 3D performance instead of a discrete GPU with associated RAM, VRMs, heatsink(s), fan(s), heat and battery power draw - no point in wasting space and power on a discrete GPU in a machine I would end up using mainly for business / non-gaming purposes.

I covered that point, a laptop to me is business and recreational only so I don't need grunt but then I can look at a Kabini based A6 5200 or 5000 I don't need monstrous levels of IPC to type a document and from those that got to test Kabini based systems say they are cool even after load, as we have seen its power is really low. Best yet I can pick up a Entry Kabini for little over $250 and does all I need.

 

[Marcus52]

Haswell is not a "yawn" to me. Sandy Bridge was, since I was already on Nehalem. Ivy Bridge was a slight improvement again, but still not enough to cause me to want to upgrade. Now, I get why Haswell would be a yawn to those that already have Ivy Bridge, even Sandy Bridge, but there is finally enough performance difference for me to justify replacing my i7 920.
I'm going to wait awhile and see what Ivy Bridge-E brings to the table before I replace my main rig with Haswell, but as of now, Haswell looks to be it.