Haswell new instructions

[Stickmansam]

Hi all,

I am wondering how useful AVX2 and FMA3 will be in general computing and gaming. I know stuff like Handbrake will likely pick up AVX2 and FMA3 quickly but how about things like games and more mundane programs? How long will it take for them to use these new instruction sets? I'm already running Ivy so I won't be getting Haswell but I want a fell as to when upgrading to the cpu arch of the day will see me some sizable gains. I can afford to be slower in Handbrake since I don't use it too much but when games actually need that extra CPU omph, I'll probably need to upgrade then.

Will these instructions sets be game changers in way? Like allow games to run like twice as fast on the cpu e.g. 80%load vs 40%load just by uisng these instruction sets.

 

[Burpo]

Read Test report HERE

 

[Stickmansam]

Read Test report HERE

I see the synthetic benchmarks with the new instructions are somewhere from 20-78% faster but what does this mean for games? Can games in the future use these instructions and how long will it take before games will use them and how long until I will need a cpu that has theses instructions just to let the game be playable?

 

[ShintaiDK]

Some new games already use them. The benefit varies on how much of the code is SSE/AVX.

 

[Durp]

Some new games already use them. The benefit varies on how much of the code is SSE/AVX.

Can you please provide an example?

 

[ShintaiDK]

Can you please provide an example?

Rome 2 I would guess?

Compilers supported AVX2 code since last year. Its not exactly something brand new. And autovectorization is not making it hard either.

 

[Stickmansam]

Rome 2 I would guess?

Compilers supported AVX2 code since last year. Its not exactly something brand new. And autovectorization is not making it hard either.

So does Rome 2 run exceptionally faster on Haswell? I guess to maintain market, most games will still be relatively playable on high clocked 1156/1155 cpus for quite a while but those with the newer instructions will get much better perf, at least cpu wise?

 

[NTMBK]

Rome 2 I would guess?

Compilers supported AVX2 code since last year. Its not exactly something brand new. And autovectorization is not making it hard either.

Given the shambolic state of any Creative Assembly game, I wouldn't put too much faith in them moving to the latest compiler technology very quickly! Their QA is bad enough as it is, without switching to an entirely new compiler and set of standard libraries a few months before release.

 

[ShintaiDK]

Well I play Rome 2. And I tried both on a 3570K and a 4670. And there is a noticeable difference.

You make it sound like compiler support was just around the corner. We talk 1 year+. Not to mention its an Intel sponsored game. Haswell IGP is also supported with direct memory access option.

 

[Durp]

The only CPU performance review of Rome 2 that I can find is from GameGPU and they don't have a Haswell system...

 

[Magic Carpet]

Well I play Rome 2. And I tried both on a 3570K and a 4670. And there is a noticeable difference.

You didn't bench them side by side? You could one of the first to bench Rome 2 on Haswell.

Haswell IGP is also supported with direct memory access option.

What does it do, it's so exclusive to Haswell IGP. D-GPUs don't need it?

 

[ShintaiDK]

You didn't bench them side by side? You could one of the first to bench Rome 2 on Haswell.


What does it do, it's so exclusive to Haswell IGP. D-GPUs don't need it?

I can try see if I can run the benchmarks later today with same settings on both systems.

For Haswell option.

 

[NTMBK]

Well I play Rome 2. And I tried both on a 3570K and a 4670. And there is a noticeable difference.

You make it sound like compiler support was just around the corner. We talk 1 year+. Not to mention its an Intel sponsored game. Haswell IGP is also supported with direct memory access option.

Interesting, I didn't know it was an Intel sponsored game! Definitely increases the odds of it having an up to date Intel compiler. Haswell benchmarks would definitely be appreciated, if you get the chance at all.

 

[Magic Carpet]

I can try see if I can run the benchmarks later today with same settings on both systems.

For Haswell option.

Thanks, also if you get the chance to bench it with that DRA setting :thumbsup:

 

[MisterMac]

Shintai - bench now.

 

[ShintaiDK]

Shintai - bench now.

Well, both gives 66FPS in the ingame forest benchmark. And the benchmark is capped at...66FPS

 

[Nothingness]

If we are to pick Rome 2 as to what AVX2 brings to gaming here you go: http://www.bit-tech.net/hardware/2013/06/12/intel-core-i5-4670k-haswell-cpu-review/5

Intel Core i7-3770K (3.5GHz) min 25 ave 32
Intel Core i7-4770K (3.5GHz) min 28 ave 35

Note this is a built-in CPU test. Good but not impressive.

That being said for my number theory stuff, Haswell brought me a huge speedup over my i7-920, but that's extremely specific

 

[ShintaiDK]

If we are to pick Rome 2 as to what AVX2 brings to gaming here you go: http://www.bit-tech.net/hardware/2013/06/12/intel-core-i5-4670k-haswell-cpu-review/5

Intel Core i7-3770K (3.5GHz) min 25 ave 32
Intel Core i7-4770K (3.5GHz) min 28 ave 35

Note this is a built-in CPU test. Good but not impressive.

That being said for my number theory stuff, Haswell brought me a huge speedup over my i7-920, but that's extremely specific

First of all its Shogun 2, not Rome 2.

Secondly, its a prescripted battle. That means very low CPU utilization. We already seen how big a difference there is on prescripted and actual gameplay in other games.

Example with Tomb Raider:
40% difference in prescripted benchmark. 400% difference ingame between 800Mhz and 4900Mhz.
http://forums.anandtech.com/showpost.php?p=35371206&postcount=7

 

[Nothingness]

First of all its Shogun 2, not Rome 2.

Oops my bad

Secondly, its a prescripted battle. That means very low CPU utilization.

Then why does the company call it a CPU test? Odd.

EDIT: Some results here: http://www.twcenter.net/forums/show...erformance&p=13102376&viewfull=1#post13102376

Hopefully this is not BS again

 

[ShintaiDK]

Oops my bad

Then why does the company call it a CPU test? Odd.

EDIT: Some results here: http://www.twcenter.net/forums/show...erformance&p=13102376&viewfull=1#post13102376

Hopefully this is not BS again

Its the same prescripted benchmark with different GPU settings.

I actually assume the only somewhat way is to use a campaign CPU movement turn length to measure with.

 

[NTMBK]

Well, both gives 66FPS in the ingame forest benchmark. And the benchmark is capped at...66FPS

Well that sucks!

Perhaps you could try artificially crippling your performance? Is it possible to underclock a K-series processor by setting the multiplier extremely low (e.g. 1GHz)? If you clocked both of them down to the same speed, you might be able to see how their IPC compares.

 

[Magic Carpet]

Its the same prescripted benchmark with different GPU settings.

I actually assume the only somewhat way is to use a campaign CPU movement turn length to measure with.

Well, it would be nice to measure it, can't you use Fraps? That way you can pick a specific cpu-heavy part and run it on two machines. If the difference was noticeable, the improvement should be in the two figure area, which Fraps should capture and display in the logs.

 

[ShintaiDK]

Well, it would be nice to measure it, can't you use Fraps? That way you can pick a specific cpu-heavy part and run it on two machines. If the difference was noticeable, the improvement should be in the two figure area, which Fraps should capture and display in the logs.

Problem is they wouldnt be identical.

Gonna try a run now with 1600Mhz, since the IB cant go lower.

 

[Magic Carpet]

Problem is they wouldnt be identical.

Gonna try a run now with 1600Mhz, since the IB cant go lower.

I know, but it will still give some idea, +/- 10% is acceptable. If there is a 25%+ performance difference, it will be easy to spot. Far more useful than your prev. run with identical 66 fps. Fraps will log everything in its dir, so we can see min / avr / max fps.

 

[ShintaiDK]

Ingame benchmark at 1600Mhz. 58FPS for IB, 62FPS for HW. But large time of the bench both are capped.