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

[guskline]

Overall, Haswell will improve over SB/IB. Is it dramatic? No, but Intel's push to improve in performance per watt is continuing. I've used AMD since the 386-40 and rarely strayed from them (486DX33, P60) until I made the jump to the SB 2500k and now the IB 3770k. Comparing power usage of my 3770k stock/OC to my FX8350 stock/OC is dramatic. Needless to say the 3770k is more powerful for almost any benchmark I can run. Haswell appears to continue the power saving march while continuing or slightly improving performance.

 

[inf64]

I would be surprised if we would to see >4 commercial desktop apps that can use AVX2 in next year or two.

 

[ShintaiDK]

I would be surprised if we would to see >4 commercial desktop apps that can use AVX2 in next year or two.

The first AVX2 commercial applications popped up 13 months ago. Consider that

 

[inf64]

Where can I DL these applications and can you post a list of them.

PS I'm interested in something useful and not a "proof of concept applications".

 

[strata8]

Everything. AVX2 will (over time) be one of the greatest leaps for many years.

It will be another increment, just like all the others. The issue you run into is you need 16 non-dependant 32bit vectors, you know how everyone loves to quote Amdahl's law around here . Manderbolt ( what the sandra multimedia benchmark is) gets perfect scaling it is in no way representative of common workloads, if it was CPU's should pack up and go home now, GPU's are king.

No.

What a convincing argument.

 

[lamedude]

software word is a total mess....i won't be surprised if there is programs beeing coded in x87 even today

Skyrim did until SkyBoost hacked in some SSE2 and someone at Bethesda was like maybe we should add /arch:SSE2 to our makefile. Its the default setting in VS2012 so hopefully in a few years it should be mostly dead (still no replacement for x87's 80bit precision).

 

[grimpr]

Where can I DL these applications and can you post a list of them.

PS I'm interested in something useful and not a "proof of concept applications".

I can make a wild guess. First apps that come through with AVX2 will be Intel optimized 3d renderers like Cinebench and others and video encoding engines like x264 and tmpgenc.

 

[grimpr]

Skyrim did until SkyBoost hacked in some SSE2 and someone at Bethesda was like maybe we should add /arch:SSE2 to our makefile. Its the default setting in VS2012 so hopefully in a few years it should be mostly dead (still no replacement for x87's 80bit precision).

I still can't believe they did this, releasing an epic PC rpg title with no optimizations at all and sticking with x87 compiles and if somebody from the community didnt come up with a hacked solution they wouldnt even cared, what the hell they where thinking?

 

[kn51]

what the hell they where thinking?

$$$ signs.

Enjoy reading these discussions, I learn a lot. Kind of clueless on the 80bit precision stuff though so made me research. Interesting stuff on why and what choices are made.

 

[ShintaiDK]

Where can I DL these applications and can you post a list of them.

PS I'm interested in something useful and not a "proof of concept applications".

One example, AVX2 ready from february 2012:

http://www.inartis.com/default.aspx

We have already included a new path for Haswell targets (using FMA and AVX2 instructions).
Thanks to the early support in the Intel compiler and the SDE I was able to port and validate very quickly the codeusing FMA and the 256-bit packed int instructions. A cool feature of the Intel C++ compiler is that legacy code using MUL + ADD intrinsics (such as _mm256_mul_ps / _mm256_add_ps) use FMA instructions wherever possible when compiledwith the "/QxCORE-AVX2" flag, it's a great time saver and we can continue to have exactly the samesourcecodefor all (legacy SSE &AVX andnew FMA+AVX2) paths. Also since we use wrapper classes around intrinsics, the source code is still very readable, for example
res = a*x + b*y +c;
is far more readable than if wehad to introduce FMA functions such as
res = madd(a,x,madd(b,y,c));
More optimization opportunities are still there using any to any permute and gather for example, I suppose that I'll wait for the real chipsfor these.

Seems its easy to forget how long the option to code for AVX2 already existed.

 

[inf64]

Ok so what is that to desktop user? Can you provide some links to AVX2 optimized common desktop (useful) apps? Kribi is more of a proof of concept to me. How good are the speedups in it, what does the developer say?

edit: 4 more to go to prove me wrong in my prediction above .

 

[ShintaiDK]

I still can't believe they did this, releasing an epic PC rpg title with no optimizations at all and sticking with x87 compiles and if somebody from the community didnt come up with a hacked solution they wouldnt even cared, what the hell they where thinking?

I remember reading about a game company with such issue. That stated their game was compiled by the publisher rather than developer due to DRM addition.

 

[mikk]

Ok so what is that to desktop user? Can you provide some links to AVX2 optimized common desktop (useful) apps?

Odd question. There is no AVX2 CPU officially available on the market.

 

[Acanthus]

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.

Yup. This is what i'm waiting to see.

I will need a 40% single thread increase to get me to bite. for ~$500. Right now, i'll only get about 20% peak at the same clockspeed. If it does indeed overclock better like some Intel guys have hinted at, I will be much more inclined to buy.

 

[cytg111]

I can make a wild guess. First apps that come through with AVX2 will be Intel optimized 3d renderers like Cinebench and others and video encoding engines like x264 and tmpgenc.

cant find any info about inclusion in the JVM or CLR for that matter .. But those would be prime candidates to get legacy code AVX2/TSX'ed as well

 

[IntelUser2000]

Obviously you never tested. HD3000 gained 10% after a couple of months with a new driver series. HD4000 similar, depending on the game there are huge increases from launch to now. Some games suffered from the famous low clock issue. In this games like UT2004, Half-Life 2, Gothic 2, Far Cry, recently Far Cry 2 you see a doubling or more with the new series for Ivy Bridge and Haswell. Also games like Civilization 5 or Dirt series gained a couple of fps on a HD4000 since launch a year ago.

That after launch driver for HD 3000 is the only one I can remember. Of course there's a gain sometimes, which is why I said almost. But barely. I've seen ATI and Nvidia claim performance gains and you can actually see a gain in many scenarios and in actual reviews. For Intel the gains are much less pronounced than claimed or in very few scenarios.

You can find a guy in Intel software forums doing testing with more than a dozen games and you see one big gain, and subsequent versions slowly reducing that gain to eventually about half after several versions.

 

[meloz]

That after launch driver for HD 3000 is the only one I can remember. Of course there's a gain sometimes, which is why I said almost. But barely. I've seen ATI and Nvidia claim performance gains and you can actually see a gain in many scenarios and in actual reviews. For Intel the gains are much less pronounced than claimed or in very few scenarios.

You can find a guy in Intel software forums doing testing with more than a dozen games and you see one big gain, and subsequent versions slowly reducing that gain to eventually about half after several versions.

Intel's GPU drivers are not anywhere near AMD or nvidia in terms of maturity.

It's a shame, they make billions of $ in profit, they can easily afford to throw more people at the problem. Re-organize their GPU drivers team so that they can catch up the other two quickly. But they don't seem to care because they already have 90%+ of the GPU market. That's all Otellini sees: marketshare and profits.

 

[mikk]

You can find a guy in Intel software forums doing testing with more than a dozen games and you see one big gain, and subsequent versions slowly reducing that gain to eventually about half after several versions.

This test is outdated, the 15.31 series is the driver which improved performance in various games and fixed the low clock issue. You don't see much differences for Nvidia and AMD as well in a bigger range of games. Only in recently released blockbusters due to game specific tweaks. The test you have mentioned has many older games included. In older games you won't see improvements on AMD/Nvidia more or less. Furthermore, I haven't seen driver tests on AMD iGPUs yet.

 

[frozentundra123456]

Yup. This is what i'm waiting to see.

I will need a 40% single thread increase to get me to bite. for ~$500. Right now, i'll only get about 20% peak at the same clockspeed. If it does indeed overclock better like some Intel guys have hinted at, I will be much more inclined to buy.

You may never see a 40% increase in single threaded performance over a 5ghz SB, until there is some major breakthrough in cpu design.

 

[inf64]

40% is possible if your starting level is very low(like it was with Prescott; this is what we got when intel gave us Conroe).
40% is highly unlikely if you have heavily optimized and overly complicated OoO uarchitecture like Core( 3rd gen. SB/IB). The bar is set so high that no matter how much HW you throw at it you have very diminishing returns. This problem is here to stay until heterogeneous computing and multicore programming becomes more common. These two areas are capable of giving much more than just 40% speedups ,but to get there is much harder on the software side.

 

[Blandge]

40% is possible if your starting level is very low(like it was with Prescott; this is what we got when intel gave us Conroe).
40% is highly unlikely if you have heavily optimized and overly complicated OoO uarchitecture like Core( 3rd gen. SB/IB). The bar is set so high that no matter how much HW you throw at it you have very diminishing returns. This problem is here to stay until heterogeneous computing and multicore programming becomes more common. These two areas are capable of giving much more than just 40% speedups ,but to get there is much harder on the software side.

40% increase over his current hardware. Not 40% in a single generation.

 

[Smoblikat]

HAH, FINALLY!!

So there hasnt been any performance increases since SB. There fore I can now justify myself buying the ASRock Z77 extreme 11 that I so desperatley want.......

 

[IntelUser2000]

This problem is here to stay until heterogeneous computing and multicore programming becomes more common. These two areas are capable of giving much more than just 40% speedups ,but to get there is much harder on the software side.

That's actually a big issue that this industry faces, and especially the PC one.

Relying on software improvements for performance gain is a bad idea. It never works. Plus, we shouldn't have software developers spend too much resources on optimizing for performance when other things should be done instead. Like making the software more useful.

I always like to give an example of Valve and Blizzard. They don't have the best graphics or use the latest graphics engines, but consistently make good games.

We like to see our developers do everything good, but its not going to work when there's finite resources.

 

[IntelUser2000]

This test is outdated, the 15.31 series is the driver which improved performance in various games and fixed the low clock issue.

I read it. It sounds very good. I was hoping if they'll really boost graphics with Haswell, the driver should follow as well.

 

[bronxzv]

Kribi is more of a proof of concept to me.

I'll agree it's a niche product affecting a small set of users, but note that it's used worldwilde, dayly, since several years, by 1000s of users in big companies like Procter&Gamble and Logitech just for this application : http://www.planogrambuilder.com/
so it's arguably more than a proof of concept

anyway, the take home message of the excerpt of a post of mine above (by ShintaiDK), is that the port was very easy, even that I use low level constructs (intrinsics in wrapper classes), for developers relying on the vectorizer it will be just a matter of recompilation, for the ones using the Intel libraries such as MKL/IPP merelly linking with new static libraries or shipping new DLLs