First complete review of Haswell i7-4770K

[biostud]

So you dont know, just making random assumptions.

This survey says 77% overclock their cpu

http://forums.anandtech.com/showthread.php?t=2234279&highlight=overclock

 

[A5]

What's the use for AVX2? For gaming, graphics, physics and such...can I get a dumbed down version on how this is a good thing for performance? I mean, does it even impact gaming? (If it gets supported)

And how or why is it different than other "technologies", why should people care? We got OpenCL, CUDA, OpenMP, C++AMP, DirectCompute, HSA etc..

AVX2 is a set of instruction extensions for Single Instruction, Multiple Data operations. From the most reductive possible viewpoint, it is just some more useful instructions added to SSE (AVX is also double the "width" of SSE).

In general, SIMD lets you perform operations on groups of data quickly. Lets say you are dealing with a group of 32-bit floating point numbers, like the color values of pixels in a video clip.

In an environment with no SIMD instructions, you have to iterate through every R,G, and B value on each pixel, and each operation takes as long as it takes to get through your FP stage in the pipeline.

In a SIMD environment, you can bundle these values together in a "vector" and operate on all of them at once. So in the case of AVX2, you would be able to bundle 8 32-bit values together and get through them in one FP pipeline time.

I'm not up on exactly what operations are added in AVX2 vs AVX that are getting people more excited, but you can obviously see that the more things that are covered by this the better.

Edit: OK, people are excited about AVX2 because it extends a lot of this to integer math (AVX is almost all FP) and the new data-gather support enables non-contiguous data to be used.

 

[tviceman]

Iris/Iris Pro hasn't been benchmarked anywhere yet. You're looking at benchmarks of HD4600

I know it hasn't been benchmarked yet, I was refering to Intel's own PR slides for the basis of what Iris will perform at on anand's overview/preview.

 

[aigomorla]

very very disappointing...

if i dont see a 8core variant, im waiting til at least refresh.

 

[lagokc]

very very disappointing...

if i dont see a 8core variant, im waiting til at least refresh.

Expect Intel to save the 8core variant for the Xeon line and expect it to cost a lot more.

 

[Carfax83]

Edit: OK, people are excited about AVX2 because it extends a lot of this to integer math (AVX is almost all FP) and the new data-gather support enables non-contiguous data to be used.

So can that help games?

 

[aigomorla]

Expect Intel to save the 8core variant for the Xeon line and expect it to cost a lot more.

problem tho, unless intel keeps with the enthusiest = enterprise motif they have been doing, that Xeon wont be unlocked.

Sigh....

I think i may just end up going dual lga2011 and running ivy's.

 

[Cerb]

So can that help games?

Depends on game, and what it has that might benefit. It's very generic, and situational. In cases where games could do more with Altivec or VMX on the consoles, like Havok or Physics, and probably the latest one getting popular, bullet, it should fly. It should offer anywhere from a 2x to 8x speedup, compared to SSE2, and possibly 20x compared to having to use scalar processing instead of vector. But, those big speedups only apply to certain code loops, so even if those loops get 20x faster, the whole application will only get so much faster, based on how much time those loops had been taking up. FI, if that was 20% of its time, and it got sped up by 10x, it will now need 82% of the original time for some work unit, and so can run 22% faster.

Because of it being so situational, and clock speed improvements not being, there's good reason they didn't bother improving the ISA in that area, until clock speeds and IPC started topping out (it generally takes them 2-3 years to get emergency product changes out, and more like 5-8 years for major changes, like this). Until then, they could instead put work into making the whole CPU run faster, at everything.

 

[galego]

What would be the TDP of an 8-core variant? 190W?

Old 6-core models such as the 3960k and the 3970X have TDPs of 130W and 150W respectively.

 

[aigomorla]

What would be the TDP of an 8-core variant? 190W?

Old 6-core models such as the 3960k and the 3970X have TDPs of 130W and 150W respectively.

im guessing it wont exceed 150.
If we super impose haswell those values would be more like:
95W on the first 115W on the second.

Adding 2 more cores on the cpu would probably bring it back up to the TDP of its predecessors while being able to do more work in the same package.

 

[Carfax83]

Depends on game, and what it has that might benefit. It's very generic, and situational. In cases where games could do more with Altivec or VMX on the consoles, like Havok or Physics, and probably the latest one getting popular, bullet, it should fly. It should offer anywhere from a 2x to 8x speedup, compared to SSE2, and possibly 20x compared to having to use scalar processing instead of vector. But, those big speedups only apply to certain code loops, so even if those loops get 20x faster, the whole application will only get so much faster, based on how much time those loops had been taking up. FI, if that was 20% of its time, and it got sped up by 10x, it will now need 82% of the original time for some work unit, and so can run 22% faster.

Sounds like game physics will get a big boost from AVX2 then. A lot of the GPU accelerated PhysX effects we see in now that don't run very well in software mode, will get a massive boost in performance once the code has been optimized.

That's good!

 

[A5]

Sounds like game physics will get a big boost from AVX2 then. A lot of the GPU accelerated PhysX effects we see in now that don't run very well in software mode, will get a massive boost in performance once the code has been optimized.

That's good!

It'll help, but it will probably be awhile. Especially since the new consoles won't have it.

You'll be glad it exists when you buy your Skymont CPU in 2016, but it won't be a huge thing for games right away since AVX2 requires the application to be compiled for it.

 

[blackened23]

So can that help games?

Yeah? I'm pretty sure intel didn't create it specifically for that purpose. It will also take time.

The adoption for gaming always lags behind general use, especially since games are designed for multiple platforms. As an example, Quake was the first video game to make extensive use of floating point years ago, while other applications had used floating point for some time. Now the use of FP in games is fairly ubiquitous - id software led the way. It's too bad that id software isn't at the forefront of pushing boundaries anymore, although it is very understandable. The gaming landscape is far different now due to multi platform development costs.

 

[Phynaz]

Here's the slides about Haswell's voltage regulator

http://www.expreview.com/25414.html

 

[colonelciller]

Majority of desktop users still cares about performance/watt and request smaller form factors.

false, though if you keep saying it over and over and over and over and over like a mantra then maybe someone here will be convinced by the make-believe.

 

[colonelciller]

Good point. If perf/watts was the most important we would see a lot of advertising about it on desktop PC's

+10

 

[ThePeasant]

It should offer anywhere from a 2x to 8x speedup, compared to SSE2, and possibly 20x compared to having to use scalar processing instead of vector.

In what world are these multipliers plausible.

 

[Phynaz]

false, though if you keep saying it over and over and over and over and over like a mantra then maybe someone here will be convinced by the make-believe.

He's right. He said majority, do you know the definition of that word?

 

[galego]

He's right. He said majority

Statistics?

 

[IntelUser2000]

Here's the slides about Haswell's voltage regulator

http://www.expreview.com/25414.html

We've already established few months ago its a research prototype not a real thing.

 

[Cerb]

In what world are these multipliers plausible.

The kind where 256/32 = 8, one was stuck with scalar x87 or scalar SSE2, and is lucky enough to be able to process a whole cache line without interdependence, and arrange the data to be read in one cache line, and use adds and/or multiplies a lot. Hence why it's a range. You're not going to flip a compiler switch and get an 8x speedup, but that's about where a potential speedup would top out, for already-tight loops, though there is more potential for int cases, due to fewer instructions, registers, and operands needed for the front-end to keep track. Theoretically, Haswell can also perform twice as fast as Ivy Bridge, when it comes to vectors, so its not all just from AVX2 itself, though AVX2 will be needed to take full advantage of it.

 

[CakeMonster]

This is far from iron-clad, but Xbit has a pretty good track record IME.

This is a relief. It also makes for a good excuse to upgrade now to move on to LGA1150. (Oh well, a good enough excuse I guess :whiste: ).

Is there any reliable info whatsoever about what Broadwell is expected to bring CPU-wise and chipset-wise? Will the 1150 chipset refresh bring something more meaningful than than Z77 over P67/Z68?

 

[NTMBK]

This is far from iron-clad, but Xbit has a pretty good track record IME.

That just says that they will keep LGA for desktop- it doesn't specify that Broadwell is coming to desktop. Broadwell may be a mobile only release. Makes sense- they want to release 14nm Atom next year, and they can use the 14nm volume that would have gone to desktop Broadwell for Atom instead, and keep using 22nm for the desktop parts. Not really that big a deal, either- remember how disappointed us desktop enthusiasts were with IB?

 

[ShintaiDK]

That just says that they will keep LGA for desktop- it doesn't specify that Broadwell is coming to desktop. Broadwell may be a mobile only release. Makes sense- they want to release 14nm Atom next year, and they can use the 14nm volume that would have gone to desktop Broadwell for Atom instead, and keep using 22nm for the desktop parts. Not really that big a deal, either- remember how disappointed us desktop enthusiasts were with IB?

Broadwell is not mobile only. We also got H97 and Z97 chipsets confirmed. So lets not keep spinning the fear^H^H^H^H rumour mill.

 

[NTMBK]

Broadwell is not mobile only. We also got H97 and Z97 chipsets confirmed. So lets not keep spinning the fear^H^H^H^H rumour mill.

Chipsets != processors.

But yeah, I may well be wrong and we may see desktop Broadwell. I just don't think we've had proper confirmation either way yet.

EDIT: Also, I don't think it's something to be afraid of. Full towers, the home of LGA, doesn't need the lower TDPs that much, or improved integrated graphics. How many people here upgraded from SB to IB? But compact machines, whether they're nettops, all in ones or laptops, will benefit a lot from it. And Intel is helping make these smaller form factors more palatable- their thin mini-ITX platform is looking pretty nice. http://techreport.com/review/24785/intel-thin-mini-itx-platform-nine-months-later