[THG]Core i7-4770K: Haswell's Performance-Previewed

[inf64]

Source
Thanks to mayankleoboy1 at SA forums for finding this. I quote his summary since it sums up what THG found out.

Some takeaways :

1. The unlocked chips have similar base clocks compared to their IB counterparts.

2. The unlocked chips dont have the new TSX

3. The desktop LGA1150 chips wont have GT3/Kick-ass graphics.

4. The CPU performance improvement is about 3%-8%.

5. GT2 perf has improved, but nowhere close to Trinity.

Per clock performance is here. The advantage over IB varies (naturally), it goes from 3%(lame-ST) to 7-13%(blender and visual studio-MT/ EDIT: these 2 results in blender and VS are not per clock but are with Turbo and power savings on ) which is not bad IMO.

 

[itsmydamnation]

i'm still an oblivion man, which means i need the fastest single thread perf i can get. Praying for better then IB high end air clocks.

 

[aaksheytalwar]

In synthetic benchmarks the difference in many of them is like 30-50% of clock for clock. Check carefully bro. It is these synthetic differences that actually result in a better and stutter free experience in the world world, even when the difference apparent in real world games is just 10-15%.

 

[inf64]

In synthetic benchmarks the difference in many of them is like 30-50% of clock for clock. Check carefully bro. It is these synthetic differences that actually result in a better and stutter free experience in the world world, even when the difference apparent in real world games is just 10-15%.

In which benchmarks do you see 30-50% advantage for Haswell over IB? Can you point them out in THG review? Are your referring to iGPU parts in these CPUs ?

 

[StrangerGuy]

Oh my dear 2500K, you are going to be with me for another 3 years at least.

 

[MisterMac]

Oh lord - i hope IPC is better than this in final silicon

Bad intel, bad.


Stupid iphone tablet hugging hippies ruining my desktop world.

 

[MisterMac]

Also - why is Tom's previewing this but not Anandtech or other more... heavy technical sites?

Why does Unlocked SKU's not have TSX?


In general Toms doesn't have the greatest rep - but is this real info or crap from everywhere as they usually post?

Did they litterally go get some chips from a source?

 

[NTMBK]

8% improvements, with a 9% increase in TDP. Meh. Should be nice for certain workloads once AVX2 becomes common, but that won't be for a while.

 

[Pilum]

Also - why is Tom's previewing this but not Anandtech or other more... heavy technical sites?

Because it's not a final system, at least the memory controller is underperforming (~17% lower throughput than IVB/SNB). This will have an impact on several benchmarks. I doubt the impact will be great, but it might sway some scores a few percent here and there. So let's have some patience until a reliable site benches Haswell.

 

[Roadrunners]

I am now thinking about grabbing a 3930k and not waiting for haswell.

 

[itsmydamnation]

Because it's not a final system, at least the memory controller is underperforming (~17% lower throughput than IVB/SNB). This will have an impact on several benchmarks. I doubt the impact will be great, but it might sway some scores a few percent here and there. So let's have some patience until a reliable site benches Haswell.

Remember that the cache system has some big changes, width doesn't come for free. We might be seeing a throughput for absolute latency trade off. im interested in AVX benchmarks that will tell us how usefull 256bit is over 128bit as the whole CPU is now geared for it, not just execution resources.

 

[ShintaiDK]

This is crazy and shows the potential for the future:

And Haswells L1 cache can almost do 1TByte/sec.

 

[grimpr]

The IPC was more or less expected as the integer throughput in AVX2 which is impressive & the new FMA.I'd like to know about the thermals & power consumtion compared to SB because as a whole platform it will be more power friendly, thats where Intel primarily focused besides AVX2/TSX.

 

[grimpr]

This is crazy and shows the potential for the future:

AVX2/FMA, almost 2x exactly like Intel promised.

 

[ShintaiDK]

The IPC was more or less expected as the integer throughput in AVX2 which is impressive & the new FMA.I'd like to know about the thermals & power consumtion compared to SB because as a whole platform it will be more power friendly, thats where Intel primarily focused besides AVX2/TSX.

Yep, a shame they didnt add a power consumption chart. Even tho its prerelease hardware.

 

[tweakboy]

I am now thinking about grabbing a 3930k and not waiting for haswell.

Wait my friend for Ivy Bridge E 8 core desktop enthusiast. Q3 2013....

gl

 

[meloz]

2. The unlocked chips dont have the new TSX

Ugh. Typical Intel. They continue to cripple the unlocked chips in some way or other.

Although the logic of disabling TSX is beyond me, unless TSX shares the same silicon area with the virtualization. Intel have traditionally crippled virtualization in unlocked chips, so TSX might be collateral damage.

 

[frozentundra123456]

Overall pretty much what was expected, but disappointing. And it is stupid of Toms to keep comparing it to a six core. A very efficient quad core can compete with a less efficient cpu with more cores, but not against a six core with basically the same efficient architecture.

Come on Intel, give us a clockspeed bump and six cores in the mainstream.

And as far as the new instruction sets, who knows whether they will ever be widely useful. "Our cpu is great, the software just has not caught up yet" ---we have heard that already. Just make it better for the current software.

Maybe final silicon will be slightly better, but not holding out much hope.

 

[grimpr]

Ugh. Typical Intel. They continue to cripple the unlocked chips in some way or other.

Although the logic of disabling TSX is beyond me, unless TSX shares the same silicon area with the virtualization. Intel have traditionally crippled virtualization in unlocked chips, so TSX might be collateral damage.

They "featurize" through blown fuses. All i can make from Intels decision disabling TSX on K models is simple greed, they dont want TSX users to buy cheap desktop unlocked chips to get the extra performance for free but want them to move to 6cores up in the 500-1000$ price range, they keep AVX2 in the entire Core lineup since its essential for the new extension to be supported and further distance them from AMD but they segment on TSX/VT-d, seems they think that TSX is a more business/server oriented feature rathen than something useful and essential for the home desktop.

 

[meloz]

And as far as the new instruction sets, who knows whether they will ever be widely useful. "Our cpu is great, the software just has not caught up yet" ---we have heard that already.

They have to make a CPU with new instruction first. Then the software follows. This is how it has always been. You do not put the cart before horse.

And initial impression is that it will be very useful.

 

[frozentundra123456]

They have to make a CPU with new instruction first. Then the software follows. This is how it has always been. You do not put the cart before horse.

And initial impression is that it will be very useful.

I understand that, but what is disappointing is that there is very little improvement in current applications, and they stubbornly refuse to increase clockspeed or bring hex core to the mainstream. A possible future benefit is not sufficient to offset the fact that over two generations the perfomance improvement is only around 10 percent, except in the igp, which I dont really care about on the desktop.

 

[JAG87]

It looks promising, but in the current state of software it's totally not worth the upgrade from a SB or IB system.

If it was a drop-in upgrade, sure, but mobo too... not enough on the plate to warrant that hassle.

 

[ShintaiDK]

I understand that, but what is disappointing is that there is very little improvement in current applications, and they stubbornly refuse to increase clockspeed or bring hex core to the mainstream. A possible future benefit is not sufficient to offset the fact that over two generations the perfomance improvement is only around 10 percent, except in the igp, which I dont really care about on the desktop.

Ask yourself what >99% of users demand. Then look on what you ask for. Then consider why the <1% should dictate the products. If you cant come up with anything rational to that. Then the product fits perfectly into demand.

 

[Torn Mind]

Shame he couldn't overclock...at least see if there are settings for BCLK overclocking...but I understand it wasn't his CPU.

 

[grimpr]

The ring bus in Haswell is decoupled from the core clock, i'd like to know at what speeds it runs and how much it affects performance and overclocking compared to Sandy Bridge.