Intel: Chips To Become Slower But More Energy Efficient

[vanished]

10 hours ago, patrickjp93 said:

We'll see. Samsung could pull a Houdini between now and the 5nm transition in 5-6 years.

I hope they do!  I hope they do... Or someone does

[patrickjp93]

51 minutes ago, LAwLz said:

It is open source, so you can try it out if you want.

However, the problem is that you don't just get double the performance just because your registers becomes twice as wide. For example x264 contains lots of functions which already fit inside smaller than 256 registers. So increasing the registers don't help in those instances. x264 deals with lots and lots of small numbers.

 

Since I don't know enough myself I decided to ask a x264 developer. His/her answer made things very clear.

 

I also asked about GPU acceleration and the TL;DR was that because GPUs are very parallel but bad when it comes to branches, the encoders which uses GPUs run simplified algorithms which has a lot less dependencies, but that comes at the expense of quality.

His answer that x264 is integer-based already proves he doesn't know what he's talking about beyond a basic level. AVX has integer instructions!

 

16 bytes is 128 bits, so you can work on 2 groups of pixels at once with AVX 256. Again his answer is making no freaking sense.

 

And I might be willing to give him the scalar code excuse for less vectorization, but I'm highly skeptical because more threads are helpful. I think your dev is a very early junior.

 

The GPU acceleration makes sense, but if you're careful you can turn branches into mathematical equations and just solve them anyway and not have to use simpler, messier algorithms.

[LAwLz]

2 minutes ago, patrickjp93 said:

His answer that x264 is integer-based already proves he doesn't know what he's talking about beyond a basic level. AVX has integer instructions!

Read his response again, and this time do it carefully. AVX did not have integer instructions. They were introduced in AVX2. That's why he said that AVX1 did not do much, but AVX2 roughly doubled the performance of some algorithms.

 

4 minutes ago, patrickjp93 said:

16 bytes is 128 bits, so you can work on 2 groups of pixels at once with AVX 256. Again his answer is making no freaking sense.

SIMD just ends up being a lot less efficient when it has to constantly load/unload data. You would spend more time moving things data around than actually processing it. Squeezing in two completely unrelated chunks into the same register would increase the time spent shuffling data (something you want to avoid since moving data between L1 and the registers costs valuable time), and loading two chunks with dependencies would... well... cause issues with dependencies.

 

24 minutes ago, patrickjp93 said:

And I might be willing to give him the scalar code excuse for less vectorization, but I'm highly skeptical because more threads are helpful. I think your dev is a very early junior.

It's one of the lead developers (Gramner). I am fairly sure he knows what he is talking about.

 

1 hour ago, patrickjp93 said:

The GPU acceleration makes sense, but if you're careful you can turn branches into mathematical equations and just solve them anyway and not have to use simpler, messier algorithms.

Well I don't know what GPU manufacturers has been doing, but none of them has succeeded. Maybe it is a very difficult thing to do which is why even the might Intel fails at it (although QuickSync deserves a lot of credit for being as good as it is)?

[patrickjp93]

13 minutes ago, LAwLz said:

Read his response again, and this time do it carefully. AVX did not have integer instructions. They were introduced in AVX2. That's why he said that AVX1 did not do much, but AVX2 roughly doubled the performance of some algorithms.

 

SIMD just ends up being a lot less efficient when it has to constantly load/unload data. You would spend more time moving things data around than actually processing it. Squeezing in two completely unrelated chunks into the same register would increase the time spent shuffling data (something you want to avoid since moving data between L1 and the registers costs valuable time), and loading two chunks with dependencies would... well... cause issues with dependencies.

 

It's one of the lead developers (Gramner). I am fairly sure he knows what he is talking about.

 

Well I don't know what GPU manufacturers has been doing, but none of them has succeeded. Maybe it is a very difficult thing to do which is why even the might Intel fails at it (although QuickSync deserves a lot of credit for being as good as it is)?

That's BS. AVX 1.0 already had integer instructions. Heck he even could have gone back to MMX.

 

No it doesn't when done correctly. That was the entire purpose in creating SSE (Streaming SIMD Extensions). The fact his algorithm isn't hitting memory harder means the CPU can be doing more work, especially since the 256-bit registers aren't being used.

 

And they're not unrelated unless you load them in a moronic order. Have you not heard of block-wise compression? You can also compress between blocks too. It makes perfect sense to have both in one place if you're going to need mutual information for the next step. Again no. If you use SSE correctly in combination with AVX the data will be there before you need it. The only stall is right at the start.

 

The dependencies only matter after each has been processed on its own (I've begun analyzing the algorithm and there's plenty of room to make this go faster), and at that point they'll be sitting in L3 waiting for you (since you can crunch double in the same amount of time, it's not going to fall out of L3).

 

You can be the lead on a project and still not know the whole scope of your problem or its solution. I'm not saying he's stupid. I'm saying he's ignorant.

 

Quicksync is only a set of multimedia extensions. They're being used less efficiently than possible in all likelihood. Alternatively since the iGPU is so close, maybe let the CPU branch predictors handle that and let the GPU go crunch, crunch, crunch. Heterogeneous doesn't require you only use the GPU.

[LAwLz]

Just now, patrickjp93 said:

-snip-

Well you can go talk to the developer over on the IRC if you think you know more than him. I think the problem here is that I did not convey the info properly.

They are over at #x264dev on freenode.

 

He is not ignorant though (he seems to have implemented most of the AVX2 code if you look at the change logs). I am more inclined to think that both you and I are ignorant and don't fully grasp this.

If you think you can do a better job then submit pull requests. I would not be surprised if you are overestimating your ability though.

 

I want you to post the improvements to the algorithm here though, because I am quite tired of you going "I did it better" and never backing it up.

[winningsince1337]

No need to upgrade for the next 8 years then I guess. #thanksintel Oh well... I hope AMD's Zen gives Intel a righteous spanking. They need it

[samcool55]

So if the 6700k is about as good as it's going to get for games. And AMD beats it, that would mean suddenly AMD is king and intel can't do sh*t about it?

Would be amazing tho

[Arty]

4 minutes ago, samcool55 said:

So if the 6700k is about as good as it's going to get for games. And AMD beats it, that would mean suddenly AMD is king and intel can't do sh*t about it?

Would be amazing tho

i mean gaming and cpus are pretty much no longer bottlenecks

 

 

more so for rendering it is a bottleneck.

[samcool55]

Just now, Arty said:

i mean gaming and cpus are pretty much no longer bottlenecks

 

 

more so for rendering it is a bottleneck.

Yes true even sandy bridge can still keep up no problem with a nice OC.

But still if AMD beats it, it would mean it's a more future-proof option. And if amd can offer a better product than intel for the right price, i do see intel's sales going down.

Also Intel told us recently that pc's aren't their main focus anymore, so i won't even be suprised if they decide to adjust prices more than you would expect.

[Arty]

Just now, samcool55 said:

Yes true even sandy bridge can still keep up no problem with a nice OC.

But still if AMD beats it, it would mean it's a more future-proof option. And if amd can offer a better product than intel for the right price, i do see intel's sales going down.

Also Intel told us recently that pc's aren't their main focus anymore, so i won't even be suprised if they decide to adjust prices more than you would expect.

please if amd has higher clocks it won't hurt intel.

 

Most of intel's sales are laptops. 

not pc gamers

[patrickjp93]

20 hours ago, LAwLz said:

Well you can go talk to the developer over on the IRC if you think you know more than him. I think the problem here is that I did not convey the info properly.

They are over at #x264dev on freenode.

 

He is not ignorant though (he seems to have implemented most of the AVX2 code if you look at the change logs). I am more inclined to think that both you and I are ignorant and don't fully grasp this.

If you think you can do a better job then submit pull requests. I would not be surprised if you are overestimating your ability though.

 

I want you to post the improvements to the algorithm here though, because I am quite tired of you going "I did it better" and never backing it up.

I've backed it up many times with other projects. Where have you been? And if I have spare time this semester between everything else I'll dive into it.

 

20 hours ago, winningsince1337 said:

No need to upgrade for the next 8 years then I guess. #thanksintel Oh well... I hope AMD's Zen gives Intel a righteous spanking. They need it

No, software developers need it. We've doubled our compute power per core since Sandy Bridge, and it's due to double again when Cannonlake comes out for consumers. Compilers apart from Microsoft's Visual C/C++ have been able to extract better data and instruction-level parallelism, and yet consumer software is stagnant while professional and enterprise actually is able to eat up Intel's 18-core Xeon flagship. I think the problem isn't Intel.

 

 

[shawno]

Id say this is a good thing for consumption of energy across the world. 

 

At the moment gaming wise, processors havent really netted any solid performance boosts. As we increase resolution the demand of the cpu has decreased as the gpu takes the strain.

 

I hope we see better programming to ensure we could utilise the cpu much more effectively in the future so that boosts in coming cpu builds become more practical. At the moment other than board features (m2, ddr4 ect) i see no reason to upgrade from sandy bridge.

[vanished]

1 minute ago, shawno said:

Id say this is a good thing for consumption of energy across the world. 

 

At the moment gaming wise, processors havent really netted any solid performance boosts. As we increase resolution the demand of the cpu has decreased as the gpu takes the strain.

 

I hope we see better programming to ensure we could utilise the cpu much more effectively in the future so that boosts in coming cpu builds become more practical. At the moment other than board features (m2, ddr4 ect) i see no reason to upgrade from sandy bridge.

As much as we as enthusiasts crave moar power all the time, I agree efficiency is equally important.  It allows for less heat output, longer lifespans, less costly and complex cooling solutions, longer battery life in mobile situations, and creates the opportunity/headroom for intel to just massively factory overclock the chip if they suddenly need to catch up performance-wise in a panic   Not to mention you're right about global power consumption.  I can't remember the exact numbers but I read somewhere that the majority of PC power consumption was due to a minority of machines (ie, all us gamers and power users of all kinds :p)

[AHaskin14]

I'm OK with this. Realistically the chips will still get faster while becoming more efficient, so you can start to put 2 chips into a system you normally wouldn't since you get more performance with still decreased energy

[KaiserTom]

On 2/6/2016 at 5:52 AM, HKZeroFive said:

Moore's Law

I think it's time to stop using this, it's been irrelevant for the past few years and hasn't properly communicated what people are actually referring to for longer than that.

 

Koomey's Law is much more relevant and speaks much closer to what people now mean when they say computers are getting more powerful year by year, especially in terms of something like smartphones or smaller devices in general. Koomey's Law has also been remarkably stable and probably will remain stable for as long as both power and heat is an issue for humans (which I don't foresee unlimited power and/or instant heat transfer happening in the near future).

[Carclis]

What reason would anyone have to upgrade CPU now? Power savings will be minuscule and performance will be worse. Intel is going to have to get creative with this one.

[patrickjp93]

1 hour ago, Carclis said:

What reason would anyone have to upgrade CPU now? Power savings will be minuscule and performance will be worse. Intel is going to have to get creative with this one.

That will be up to programmers to use the compute power already on the table that currently goes unused.

[MarcWolfe]

On 2/15/2016 at 0:35 PM, patrickjp93 said:

That will be up to programmers to use the compute power already on the table that currently goes unused.

Yep. They even say it here that your on your own for programing now.   Introduction to OpenMP - Tim Mattson (Intel): http://www.youtube.com/playlist?list=PLLX-Q6B8xqZ8n8bwjGdzBJ25X2utwnoEG

[Albatross]

I don't like the idea of my products getting slower/less performance for the sake of energy proficiency but whatever I guess.

[Unhelpful]

12 minutes ago, Albatross said:

I don't like the idea of my products getting slower/less performance for the sake of energy proficiency but whatever I guess.

100% unrelated... WOO WHEEL OF TIME!

[Colonel_Gerdauf]

29 minutes ago, Albatross said:

I don't like the idea of my products getting slower/less performance for the sake of energy proficiency but whatever I guess.

I think there is odd wording there. It is not going to get worse performance over time, but rather it places a higher priority on the energy efficiency. What this means is that we would eventually get performance marginally better than a 6700K while consuming power to the level of a cheap smartphone.

[Albatross]

3 hours ago, Colonel_Gerdauf said:

I think there is odd wording there. It is not going to get worse performance over time, but rather it places a higher priority on the energy efficiency. What this means is that we would eventually get performance marginally better than a 6700K while consuming power to the level of a cheap smartphone.

The moment the article says it will reduce speed in favor of energy proficiency though and either way it means customers are going to see a preference (and not of their own choosing) of energy proficiency over performance.

[Colonel_Gerdauf]

5 minutes ago, Albatross said:

The moment the article says it will reduce speed in favor of energy proficiency though and either way it means customers are going to see a preference (and not of their own choosing) of energy proficiency over performance.

As has been mentioned, it is not like Intel or AMD have a lot of options here either. We are reaching technological and physical limits on performance with silicon. It is out of the hands of raw hardware to provide performance boosts, and it is now the job of software optimization to provide it. People have been screaming for "MOAR CORES" while being unaware of the fact that software support for such amount of threads is (still) in a lackluster state.

[patrickjp93]

10 minutes ago, Albatross said:

The moment the article says it will reduce speed in favor of energy proficiency though and either way it means customers are going to see a preference (and not of their own choosing) of energy proficiency over performance.

If a company doesn't make a product which suits your needs, don't buy from it. You could equally argue consumers were always getting a limited choice of improving performance vs. not so very much improving power efficiency for many years.

[vanished]

1 hour ago, patrickjp93 said:

If a company doesn't make a product which suits your needs, don't buy from it. You could equally argue consumers were always getting a limited choice of improving performance vs. not so very much improving power efficiency for many years.

Yeah if the chips do get less powerful, and you want power over efficiency, just stick with what you've got!