Will Haswell be the clincher CPU of this decade?

[Exophase]

This doubling of GPU performance is just going to be for the GT3 variants which are laptop only. The IGPs in the high end desktops aren't going to be that much faster than they are in IB.

APUs aren't nearly as attractive in desktops though..

 

[AtenRa]

AMD's marketshare is what, maybe 30%, and of that 30% what percentage is based on APUs?

100% of Laptops are APUs
~45% of Desktop is APUs

http://www.3dcenter.org/news/genauere-zahlen-zum-absatz-einzelnen-amd-prozessorenarchitekturen

 

[Haserath]

Haswell will definitely be an powerful CPU, and probably the most interesting thing to come out of Intel since at least Nehalem (when they went with an IMC).

But clincher CPU of this decade? I would suggest you temper your expectations; if it does end up being that great then it's better to be surprised than to expect the stars and end up disappointed if they come up short.

I don't think that there will be any "clincher" CPUs this decade. Unless Intel pulls a rabbit out of the hat, we aren't going to see anything like Core2 or Nehalem over the next decade.

Our next best chance for seeing a large increase in CPU performance would some disruptive technology like quantum effects, etc. Something that can substantially decrease xtor switching times and signal wire delays.

I think he means clincher because Haswell will be the last significant increase we see for awhile, on the CPU side at least.

Nehalem was something like 15% better clock for clock over the latest Core 2, but it brought hyperthreading for some extra oomph, it wasn't really all that big of a jump otherwise.

Sandy Bridge was 10-15% better over Nehalem, brought some new instructions, and brought CPU power levels down a lot.

Haswell will be 10%? better than Sandy Bridge, will have new instructions(what else could they bring for big benefits now?), and will bring power levels down again on the CPU side even if TDP for the whole chip isn't lower.

There seems to be a trend of lower CPU performance gains, unless Haswell is actually faster than rumors have said it is.

 

[hokies83]

I think he means clincher because Haswell will be the last significant increase we see for awhile, on the CPU side at least.

Nehalem was something like 15% better clock for clock over the latest Core 2, but it brought hyperthreading for some extra oomph, it wasn't really all that big of a jump otherwise.

Sandy Bridge was 10-15% better over Nehalem, brought some new instructions, and brought CPU power levels down a lot.

Haswell will be 10%? better than Sandy Bridge, will have new instructions(what else could they bring for big benefits now?), and will bring power levels down again on the CPU side even if TDP for the whole chip isn't lower.

There seems to be a trend of lower CPU performance gains, unless Haswell is actually faster than rumors have said it is.

If it is only 10% better then Sandy Bridge that makes it the exact same speed as Ivy..

So i think you were meaning to say Haswell will be 10-20% over Ivy Bridge :thumbsup:

I expect a min of 15% improvement over Ivy Bridge with Haswell. but im sure there will be somethings that are clock for clock very close.. there always is.

 

[Fjodor2001]

I think he means clincher because Haswell will be the last significant increase we see for awhile, on the CPU side at least.

How do you know that? Do you think Broadwell will just be a node shrink and nothing else? And Skylake will be a Tock with no uArch improvements whatsoever?

 

[ShintaiDK]

Skylake is a new uarch, just like Haswell, Sandy Bridge, Nelahem, Conroe.

What I am most intersted in right now is basicly to see how MB makers will handle the ondie VRM in terms of layout.

 

[Idontcare]

100% of Laptops are APUs
~45% of Desktop is APUs

http://www.3dcenter.org/news/genauere-zahlen-zum-absatz-einzelnen-amd-prozessorenarchitekturen

OK, so we'll say 75% of AMD's market share is based on APU products.

According to Mercury Research, as reported by xbitlabs, AMD's unit market share has now dropped below 17%.

75% x 17% = 12.75%

So we are looking at AMD's APUs addressing at most 13% of the total market in terms of units.

What I don't know is how much of the budget segment is represented by AMD's 17% market share. Does AMD own 90% of the budget segment? 80%? 70%?

Numerically speaking, provided AMD owns more than 0.5/0.75 = 66.7% of the budget market, and 75% of their units are APUs then they can claim to be responsible for bringing 3D gaming to the budget buying masses.

 

[Blandge]

The budget buying masses aren't buying AMD APUs though, which is the demographic the poster was referring too.

AMD's marketshare is what, maybe 30%, and of that 30% what percentage is based on APUs?

The majority, i.e. the masses, of today's budget builds are based on Intel's iGPU.

16.1% market share as of Q3.

Market share of Intel increased to 83.3% in Q3 2012, up from 80.6% in the same quarter a year before. Meanwhile, for the first time in several years the share of AMD’s microprocessors on the x86 market dropped to 16.1% from 18.8% in Q3 2011, reports IDG News Service. Market share of Via Technologies was 0.6%.

http://www.xbitlabs.com/news/cpu/di...Market_Conditions_Competition_with_Intel.html

EDIT: Didn't see your last post.

 

[IntelUser2000]

Twice battery life seems exaggerated. Considering the cpu only comprises less than half of a laptop's power draw at idle even eliminating the cpu is not going to double battery life (most of the time a laptop is at idle or near idle-typing documents, casual web browsing, etc).

Yea, but Haswell(at least the Ultrabook bound ULT version) does FAR more than that. CPU core is pretty low already, but the uncore and the I/Os are not. The ULV Ivy Bridge has a 2.2W/2.3W TDP in package C6 and C7, which is the lowest power state. I assume its that high because the System Agent can't be cut so low. They'll lower that to 100mW or so.

They are also bringing in ultra low sleep modes on disk drives, memory, and should save another watt. Add that to faster transitions between C-states, lowered TDP level to 15W that also puts the PCH on package, and integrated voltage regulators which makes it more likely to use lower power states.

 

[IntelUser2000]

Nehalem was something like 15% better clock for clock over the latest Core 2, but it brought hyperthreading for some extra oomph,

Actually it was less than 5-8% in average, with some memory bandwidth intensive applications getting maybe little over 10%. The rest was Hyperthreading. There were cases where Nehalem was even worse in single thread compared to Penryn so it wasn't big of a deal.

If it is only 10% better then Sandy Bridge that makes it the exact same speed as Ivy..

No it isn't. Ivy Bridge has 10% increase with combined clock AND perf/clock increase. Haswell's 10% prediction is perf/clock only, over Ivy Bridge. Of course it may turn out that clock may increase 5%, or even nada over Ivy Bridge.

How do you know that? Do you think Broadwell will just be a node shrink and nothing else? And Skylake will be a Tock with no uArch improvements whatsoever?

We'll see. But if recent PCWatch article is correct, Intel might actually break the Tick Tock cycle to aim Broadwell for lower power. I'm pretty sure we'll see Tock like improvements for Skylake, but I dare say they might do it with some frequency sacrifice for peak perf/watt. Frequency scaling is effectively dead on the high end, and starting on 22nm their process is going away from increasing frequency on the highest clock to increasing frequency on low voltages.

 

[Haserath]

If it is only 10% better then Sandy Bridge that makes it the exact same speed as Ivy..

So i think you were meaning to say Haswell will be 10-20% over Ivy Bridge :thumbsup:

I expect a min of 15% improvement over Ivy Bridge with Haswell. but im sure there will be somethings that are clock for clock very close.. there always is.

I did. I actually was thinking of the realworldtech article at the time and I remembered 10% over Sandy Bridge being in it, but I think it will be 10% or so over Ivy, unless the improvements really only deliver that much.

Ivy was only 4-5% better than Sandy Bridge on average anyway.

How do you know that? Do you think Broadwell will just be a node shrink and nothing else? And Skylake will be a Tock with no uArch improvements whatsoever?

I expect there to be diminishing returns that may leave us with very little improvement.

"Compared to Nehalem, the Haswell core offers 4× the peak FLOPs, 3× the cache bandwidth, and nearly 2× the re-ordering window."

All these additions and we have perhaps 30% gains over Nehalem clock for clock? The core now has more execution resources, better memory management, an addition of a L0 cache, and tons of other little improvements, but the only major gains seem to be from new instructions under certain workloads.
--
One thing that may change this is a new architecture though...

 

[Blandge]

Yea, but Haswell(at least the Ultrabook bound ULT version) does FAR more than that. CPU core is pretty low already, but the uncore and the I/Os are not. The ULV Ivy Bridge has a 2.2W/2.3W TDP in package C6 and C7, which is the lowest power state. They'll cut that to 100mW or so.

They are also bringing in ultra low sleep modes on disk drives, memory, and should save another watt. Add that to faster transitions between C-states, lowered TDP level to 15W that also puts the PCH on package, and integrated voltage regulators.

Exactly. The 20x figure that Intel as been throwing around is PLATFORM IDLE power. Which is exactly what Engmoid was saying uses most of the power. This leaves the screen as the main power draw. Intel has worked with OEMs and component designers to lower, not just CPU power, but platform power as well. A 41% power reduction from the CPU (17W to 10W) and a 95% power reduction on the platform (20x) could easily double battery life assuming the charge capacity is not reduced.

 

[Blandge]

I expect there to be diminishing returns that may leave us with very little improvement.

"Compared to Nehalem, the Haswell core offers 4× the peak FLOPs, 3× the cache bandwidth, and nearly 2× the re-ordering window."

All these additions and we have perhaps 30% gains over Nehalem clock for clock? The core now has more execution resources, better memory management, an addition of a L0 cache, and tons of other little improvements, but the only major gains seem to be from new instructions under certain workloads.
--
One thing that may change this is a new architecture though...

"peak FLOPs" is referring to AVX2 so we'd see 300% (4x like it says) increase in performance considering that Nehalem doesn't even have AVX.

The primary purpose served by widening the pipeline so drastically was to allow AVX2 instructions to reach peak bandwidth.

The 30% increase you are referencing is with legacy instructions.

 

[Haserath]

"peak FLOPs" is referring to AVX2 so we'd see 300% (4x like it says) increase in performance considering that Nehalem doesn't even have AVX.

The only purpose served by widening the pipeline so drastically was to allow AVX2 instructions to reach peak bandwidth.

The 30% increase you are referencing is with legacy instructions.

I wasn't referring to the peak flops. I guess I should've left it out.

 

[mrmt]

100% of Laptops are APUs
~45% of Desktop is APUs

http://www.3dcenter.org/news/genauere-zahlen-zum-absatz-einzelnen-amd-prozessorenarchitekturen

Yeah, you are trolling. As if those AMD apus were all above IVB levels of performance.

 

[Haserath]

Haswell lowers 'Standby power draw' 20x.

Not exactly idle, but if you close the lid, it can sit for 20x longer that way.

 

[IntelUser2000]

A 41% power reduction from the CPU (17W to 10W) and a 95% power reduction on the platform (20x)

You shouldn't overestimate the impact of TDP on light usage though. Haswell brings benefits above and beyond reduced TDP.

I believe 10W variants will end up another 20% or more slower than 15W variants while only offering battery life increases on the really demanding scenarios. You'll still get less than 2 hours anyway so the whole point of lower TDP is to get system designs thinner and lighter.

 

[IntelUser2000]

Haswell lowers 'Standby power draw' 20x.

Not exactly idle, but if you close the lid, it can sit for 20x longer that way.

Yes, but Windows 8's "Standby" is very different from previous OSes. It's what they call Active Standby, where it won't take 2-3 seconds to wake up, but less than 300ms by MS standards, and even doable when the screen is on.

 

[Blandge]

You shouldn't overestimate the impact of TDP on light usage though. Haswell brings benefits above and beyond reduced TDP.

I believe 10W variants will end up another 20% or more slower than 15W variants while only offering battery life increases on the really demanding scenarios. You'll still get less than 2 hours anyway so the whole point of lower TDP is to get system designs thinner and lighter.

We are comparing 10W Haswell to 17W Ivy Bridge, not 10W Haswell to 15W Haswell. Am I wrong in this assumption? Intel has stated that 10W Haswell variants will be equivalent in performance to 17W Ivy Bridge.

And no, the whole point of lower TDP is dependent on the design goals of the OEMs. The beauty of Intel's horizontal strategy is that we have 15 or so companies producing products with different design goals. Price, form factor, battery life, performance, there are several products that are great in each of these categories.

 

[SickBeast]

Haswell will be AMD's death blow.

 

[Blandge]

Yes, but Windows 8's "Standby" is very different from previous OSes. It's what they call Active Standby, where it won't take 2-3 seconds to wake up, but less than 300ms by MS standards, and even doable when the screen is on.

This. Intel added new sleep states to Haswell that allows it to manage power better in Windows 8 while active and in standby. See below.

 

[jpiniero]

How do you know that? Do you think Broadwell will just be a node shrink and nothing else? And Skylake will be a Tock with no uArch improvements whatsoever?

The thinking is that past Haswell, any improvements for the time being are going to be aimed at lowering power consumption or integrating more things into the package or even the core itself. Maybe a faster gpu.

There's also the possibility that Haswell is Intel's last high watt mainstream desktop processor, and future desktops would be "laptops without a screen".

 

[Fjodor2001]

There's also the possibility that Haswell is Intel's last high watt mainstream desktop processor, and future desktops would be "laptops without a screen".

What would be the purpose of that? Are they aiming at enabling small compact desktop PCs such as the Mac Mini? Possibly fanless?

Otherwise, what reasons would the consumers owning a SB/IB/Haswell Desktop PC have to upgrade if the performance will be the same or less compared to their old computer due to the low power => low performance constraints?

 

[Ferzerp]

They have to keep developing the mainstream desktop proc because it is also the Xeon development which is a huge cash cow. The claims of "last 'high' watt desktop" aren't really founded in reality.

 

[Idontcare]

Haswell will be AMD's death blow.