First complete review of Haswell i7-4770K

[frozentundra123456]

It's probably not a stretch to say that Haswell is Intel's Piledriver. This is what Ivy Bridge should have been a year ago.

Buldozer was a step back and Piledriver was a large improvement over that.

IB and Haswell were both incremental improvements. I dont see any parallel at all.


Edit: the only bad thing I see about IB and Haswell are that the improvement is more like two ticks instead of the tic/tock that we are accustomed to. And since Broadwell is another tic, I dont expect much improvement there either for the desktop.

 

[SiliconWars]

Bulldozer wasn't a step back and I'm sure this has been done to death a million times already on this forum.

 

[WhoBeDaPlaya]

I hope this is nothing like SB -> IVB. In my mind, it was a ~5% IPC improvement at best.

 

[mikk]

if we wanted to compare IPC not CPU performance we should compare 3770k to 2700K, I think it's a safe bet that it is at least 2-3% faster then 2600k so we arrive right around 7% IPC speed-up.

Some people including you are not aware that Ivy Bridge has a higher multithreading Turbo. So even if you compare 2700k vs 3770k IVB clocks higher, means you cannot filter out the IPC. IPC increase from Sandy to Ivy was in the 3-4% range.

 

[Lepton87]

Some people including you are not aware that Ivy Bridge has a higher multithreading Turbo. So even if you compare 2700k vs 3770k IVB clocks higher, means you cannot filter out the IPC. IPC increase from Sandy to Ivy was in the 3-4% range.

Most boards use Multi-Core turbo so they are constantly running at their top turbo multipliers so it's a moot point. Both have 3.9GHz max frequency so their would be running at exactly the same frequency. I just assumed multi core turbo to be enabled as that's what happens on my board when I use stock settings. Maybe I shouldn't have assumed that. My point was that 10-15% IPC gain in IVY over SB looked severely inflated.

 

[mikk]

Most boards use Multi-Core turbo so they are constantly running at their top turbo multipliers so it's a moot point.

Yes and the top multiplier for 3/4 turbo core usage on Sandy Bridge was lower. In multithreading applications according to spec 3770k clocks higher than 2700k. Any IPC comparison based on that is nonsense. IVB improved its IPC by 3-4% and another 3% (depends on the test suite of course) due to higher MT turbo frequency.


edit: to be exact 100 Mhz higher from 2 core usage onwards.

 

[Lepton87]

Yes and the top multiplier for 3/4 turbo core usage on Sandy Bridge was lower. In multithreading applications according to spec 3770k clocks higher than 2700k. Any IPC comparison based on that is nonsense. IVB improved its IPC by 3-4% and another 3% (depends on the test suite of course) due to higher MT turbo frequency.

That's not how it works with multi-core turbo enabled, with this feature enabled both CPUs would be constantly running at 3.9GHz under load. This essentially locks the CPU in the 1 core turbo frequency and turns turbo off. It looks like overclocking but that's what high-end boards do.

 

[mikk]

That's not how it works with multi-core turbo enabled, with this feature enabled both CPUs would be constantly running at 3.9GHz under load. This essentially locks the CPU in the 1 core turbo frequency and turn turbo off. It looks like overclocking but that's what high-end boards do.

Intel specification:

i7-2700k
Base= 3.5 Ghz
with Turbo
1 core= 3.9 Ghz
2 core= 3.8 Ghz
3 core= 3.7 Ghz
4 core= 3.6 Ghz

i7-3770k
Base= 3.5 Ghz
with Turbo
1 core= 3.9 Ghz
2 core= 3.9 Ghz +2.6%
3 core= 3.8 Ghz +2.7%
4 core= 3.7 Ghz +2.8%

Every serious reviewer I'm aware of tested according to Intels specification. IPC increased in the 3-4% range. Check out reviews if you didn't know.

 

[Idontcare]

Bulldozer wasn't a step back and I'm sure this has been done to death a million times already on this forum.

Neither was Willamette or Prescott. We all know this story, its boring now.

 

[Lepton87]

Intel specification:

i7-2700k
Base= 3.5 Ghz
with Turbo
1 core= 3.9 Ghz
2 core= 3.8 Ghz
3 core= 3.7 Ghz
4 core= 3.6 Ghz

i7-3770k
Base= 3.5 Ghz
with Turbo
1 core= 3.9 Ghz
2 core= 3.9 Ghz +2.6%
3 core= 3.8 Ghz +2.7%
4 core= 3.7 Ghz +2.8%

Every serious reviewer I'm aware of tested according to Intels specification. IPC increased in the 3-4% range. Check out reviews if you didn't know.

I know Intel specification, I just said that most high-end boards ignore it and just lock the CPU at 1 core turbo frequency. Of course you can't compare IPC between 2700k and 3770k if you don't have this feature enabled. I admit 7% IPC increase was too kind for IVY but 3% seems too low. 5% is fair
ps.
Intel representative here pegged IPC improvement at 6%

Haha Thanks I could use a good laugh today.

I do agree with your numbers 10% to 15% for the 2nd generation and about 6% for the 3rd generation processor at the same clock speed.

 

[guskline]

Neither was Willamette or Prescott. We all know this story, its boring now.

Amen! Thank you!
I think haswell will turn out to be a solid improvement that will grow as new software is released.

 

[mikk]

I know Intel specification, I just said that most high-end boards ignore it and just lock the CPU at 1 core turbo frequency. Of course you can't compare IPC between 2700k and 3770k if you don't have this feature enabled. I admit 7% IPC increase was too kind for IVY but 3% seems too low. 5% is fair
ps.
Intel representative here pegged IPC improvement at 6%

Why? Here you have some facts:

http://www.computerbase.de/artikel/prozessoren/2012/test-intel-ivy-bridge/14/

I see 3-4% comparing 2700k and 3770k without Turbo.

 

[moonbogg]

So nothing new yet on OC-ability? That's all ANY of us should care about since this is the same CPU you bought 5 years ago.

 

[Lepton87]

Why? Here you have some facts:

http://www.computerbase.de/artikel/prozessoren/2012/test-intel-ivy-bridge/14/

I see 3-4% comparing 2700k and 3770k without Turbo.

3% and 5% is basically splitting hairs, it depends on the test suite. What's clear is that the gains in IPC are higher without HT enabled.

http://ixbtlabs.com/articles3/cpu/intel-ci7-123gen-p3.html

They tested IPC gains from Nehalem to Ivy. They arrived at 3.5% with HT and 4.5% without HT, but they tested at a very low frequency(2.4GHz) which may skew the results a bit, especially underestimate the improvement SB had over nehalem as its cache works at full core speed.

ps. there is at least one corner case where IVY is over 40% faster then SB.(I guess there are more) There is a game that behaves like that, but I can't remember its name. Anyone remembers what that game is?

 

[BallaTheFeared]

I was worried, but the game benchmarks eased my mind a lot considering this doesn't have a clock speed advantage over Ivy Bridge.

My concern is still overclocking though, I really hope it delivers.

 

[zebrax2]

@Idontcare

I guess i learn something today. The side jabs about my knowledge is not cool though as the explanation part would have sufficed

 

[SPBHM]

3% and 5% is basically splitting hairs, it depends on the test suite. What's clear is that the gains in IPC are higher without HT enabled.

higher IPC gains from SB to IB with HT off is because of what is explained here on the last part I think
http://www.realworldtech.com/haswell-cpu/2/


anyway, you can easily exclude clock (turbo) as a factor for sb vs ib comparing the 3.3GHz 3220 to the 2120(or the xxx5) or the 3.4GHz 3240 vs 2130 for sandy vs ivy (with HT)

http://www.xbitlabs.com/articles/cpu/display/core-i3-ivy-bridge_6.html

 

[Hulk]

Some people including you are not aware that Ivy Bridge has a higher multithreading Turbo. So even if you compare 2700k vs 3770k IVB clocks higher, means you cannot filter out the IPC. IPC increase from Sandy to Ivy was in the 3-4% range.

Ticks were originally not supposed to be performance enhancing from an architectural point of view. There were (and still are) meant to eliminate one variable (architecture) in a process shrink. But Penryn incorporated some performance enhancing features, as did the other shrinks so now we've suddenly come to expect performance increases on ticks and tocks.

 

[Wall Street]

Intel probably needs a completely new architecture to keep its CPUs IPC improving.

I think that:

1. It is much harder to improve an already good architecture than improve upon Pentium 4 which was a mistake.

2. To be fair to Haswell, much of the 'tock' innovation was in the power delivery, AVX2 and transactional memory. The power delivery doesn't get IPC but does improve Perf/Watt, the other two will only matter with code compiled for the new instructions. IF transactional memory and AVX2 take off, Haswell could look like a vastly better investment than Ivy Bridge in the future (although the k series lacks the transactional memory).

 

[Zucker2k]

So in this particular review, Haswell is 7% FASTER than IvyBridge, while consuming 12% LESS power than IvyBridge - on the same 22nm process. That's impressive!

 

[Durp]

So in this particular review, Haswell is 7% FASTER than IvyBridge, while consuming 12% LESS power than IvyBridge - on the same 22nm process. That's impressive!

When you look at it that way it looks decent. I just wish they would have focused on improving performance while staying at the same power consumption as IVY. I think IVY is already plenty efficient for desktop.

 

[cytg111]

So in this particular review, Haswell is 7% FASTER than IvyBridge, while consuming 12% LESS power than IvyBridge - on the same 22nm process. That's impressive!

That IS kind of depressing over a tock, but more a sign on the state of things than anything else. One threaded Moore does indeed seem one step closer to the grave. (but again, in the land of the blind the one threa... eyed man is king).

 

[mrmt]

Bulldozer wasn't a step back and I'm sure this has been done to death a million times already on this forum.

No? When the former marketing team touted Bulldozer as *THE* chip for the server market, and three years ago the new marketing team comes in front of an investor audience and tells world + dog that ARM is better than x86 for servers, how do you interpret that? That Bulldozer was a huge success?

 

[frozentundra123456]

When you look at it that way it looks decent. I just wish they would have focused on improving performance while staying at the same power consumption as IVY. I think IVY is already plenty efficient for desktop.

Totally agree, or even go to 100 watts as long as performance keeps increasing.

 

[MisterMac]

I had a thought that haswell and 22 nm it itself had all the potential to a killer absolute performance increase ala SB \ Conroe.

...but it seems like intel chose not to do this because of efficiency performance and... saving revenues for tougher times?

a 6 core mainstream chip would (even with haswell relatively weak ipc increases) would give crazy performance in the current "almost N threaded world".

And would probably not cost much more than Nehalem wafer wise.


Kind of like "Lets get all the margins\profit we can for the great mobile wars - because we know it'll take X time to achieve volume to sustain us" type scenario.