Intel "Haswell" Speculation thread

[bronxzv]

It was a forum invention in terms of Haswell, Broadwell etc.

The exascale slides at IDF Spring 2012 [1] were showing several configurations of on package DRAM so IMO it will come soon, probably already with (at least some variants) of Haswell, also Anand mentioned eDRAM fo Haswell in one of his Ivy Bridge article, no to mention all Charlie's articles at S|A

[1]: BJ12_ACAS003_100_ENGf.pdf (see slide 17) downloadable from intel.com/go/idfsessionsBJ

 

[Fjodor2001]

Yes. Its always bottlenecked. Same for discrete cards or Trinity/Llano.

But SB/IB/HW/BW etc got access to the LLC for fast access with smaller working sets that way.

But is it bottlenecked to the point where going from 16->40 EUs does not improve performance much? Hence the need for IGP specific eDRAM / L4 cache?

See this article where overclocking the IB HD4000 IGP by 39% only achieved a 16% FPS increase.

 

[ShintaiDK]

But is it bottlenecked to the point where going from 16->40 EUs does not improve performance much? Hence the need for IGP specific eDRAM / L4 cache?

See this article where overclocking the IB HD4000 IGP by 39% only achieved a 16% FPS increase.

The answer is in Llano/Trinity. So no, more EUs will still boost performance significantly.

 

[Olikan]

So Fudzilla is saying that GT2 graphics on Haswell performs 3x the GT1 on Ivy Bridge. Back with Ivy Bridge Intel did similar comparisons. They compared Ivy Bridge GT2 with Sandy Bridge GT1 and said it was 3x as fast.

lol, the intel fanboys here almost ate me alive when i said the very same thing

 

[mikk]

eDRAM on Intel chips is a forum invention.

Intel itself named it eDRAM in the driver.

 

[sefsefsefsef]

40 EUs w/o a super large eDRAM cache would be very, very sad. Intel wouldn't do that. I'm back on board believing in the eDRAM.

 

[IntelUser2000]

But is it bottlenecked to the point where going from 16->40 EUs does not improve performance much? Hence the need for IGP specific eDRAM / L4 cache?

See this article where overclocking the IB HD4000 IGP by 39% only achieved a 16% FPS increase.

Yes it is. The LLC cache sharing doesn't give that big of a performance boost, which is why Intel decided that with Ivy Bridge the iGPU can get away with a dedicated L3 cache and turn off the LLC cache + Ring as often as possible to save power(you can't put lot of things in 4/8MB buffer, even assuming LLC is all used for iGPU). See how in reviews Llano is bottlenecked by memory bandwidth? HD Graphics 4000 is much more sensitive to memory config because it has much more compute power than HD Graphics 3000.

Now what if we move to an enhanced 20EU in the GT2 Haswell chip? What about enhanced 40EUs with the GT3? Remember even the 20EU should end up 40-50% better performing than GT2 Ivy Bridge due to each EUs being more powerful on Haswell.

 

[ShintaiDK]

Haswells EUs aint much better than IBs. Its more or less just the amount of them. So you aint gonna get 40-50% with the same amount of EUs for sure. IB had the redesigned EUs. Haswell will just ahev more of them.

 

[IntelUser2000]

Haswells EUs aint much better than IBs. Its more or less just the amount of them. So you aint gonna get 40-50% with the same amount of EUs for sure. IB had the redesigned EUs. Haswell will just ahev more of them.

Much better? Probably not. An improvement? It definitely will be.

Ivy Bridge GT2 is 2x as fast as Ivy Bridge GT1. Haswell GT2 is said to be 3x Ivy Bridge GT1. That means the 20EU Haswell GT2 ends up being 40-50% better performing than the 16EU Ivy Bridge GT2.

Ivy Bridge iGPU co-issue rate is 60%, meaning each units are capable of 1.6x Sandy Bridge iGPU in practice, and 2x theoretical. Say they up the co-issue rate to 100%, making it 2x Sandy Bridge iGPU, and 1.25x Ivy Bridge iGPU.

Combine that with 20EUs you end up with ~25EU equivalent in Ivy Bridge terms, or a 50% boost in effective compute. Of course compute performance doesn't directly translate to performance. However, enhancements in other areas are expected as well. It's that boost in compute + tweaks.

 

[cbn]

Intel's main goal is to produce lower power CPUs for the mobile/tablet market it seems. Witness the flatter xtor performance/power that IDK (IIRC) posted elsewhere. That's what is limiting high end performance for enthusiasts.

Can you post the link? (Please)

 

[IntelUser2000]

BTW, the 22nm power/performance curve Intel gave for 22nm wasn't for the traditional drive current metric, but rather RO delay. I don't know how relevant that is to frequency, and I'm pretty sure some people can give us further insights into that.

 

[ShadowVVL]

Is haswell's main performance improvements focused at igpu or cpu or is it unknown at this time?

 

[BallaTheFeared]

I predict it will be faster than IB/SB and people with 1366 who held out saying SB/IB weren't fast enough will buy it while many people with IB/SB will now be the 1366/1156 users and say it's not fast enough and not buy it and I also predict that some people with the best will upgrade for no reason other than to have the title.


Run on.

 

[ShintaiDK]

Is haswell's main performance improvements focused at igpu or cpu or is it unknown at this time?

Its focus is IPC, iGPU and VRM integration.

 

[lehtv]

IPC will increase only by around 10%, I wouldn't call that a focus.

 

[ShintaiDK]

IPC will increase only by around 10%, I wouldn't call that a focus.

Where did you read 10%?

 

[lehtv]

http://www.fudzilla.com/home/item/28318-haswell-to-be-10%20-percent-faster-than-ivy-bridge

Sorry, it was "at least 10%". In any case, the focus is definitely on the IGP, GT3 being expected to be several times faster than HD4000. The difference in IPC pales in comparison

 

[ShadowVVL]

IPC will increase only by around 10%, I wouldn't call that a focus.

hmm thats a bit disappointing, well if intel will sell another x6 at $600 maybe I will jump on IB E or haswell E

 

[lehtv]

I might be jumping on a cheaper 3770K when haswell is released, I'll be using discrete graphics anyway

 

[ShintaiDK]

http://www.fudzilla.com/home/item/28318-haswell-to-be-10%20-percent-faster-than-ivy-bridge

Sorry, it was "at least 10%". In any case, the focus is definitely on the IGP, GT3 being expected to be several times faster than HD4000. The difference in IPC pales in comparison

Fudzilla is an epic newsjoke. Ones IQ drops by a point every time you read his nonsense.

Now Intel tells its partners to expect that Haswell should end up at least 10 percent faster than Ivy Bridge based cores at the same clock. These numbers are based on pre-silicon projections that Intel always does before it gets the working prototype back.

Intel had running Haswell chips for quite a while. Even ES around.

We are at ES2 stage now:

 

[Borealis7]

Xbitlabs is saying Intel will talk elaborately about Haswell at IDF next month, which isn't a surprise but it means we'll have to wait a month to get new info.

 

[lehtv]

ShintaiDK I don't understand your post. Is that supposed to falsify the 'at least 10% faster' claim? Are you saying the performance will be less than 10% faster? :hmm:

And does your post try to defend the idea that IPC improvement is a focus? Remember the IGP will probably improve by several hundred % over Ivy Bridge.

 

[ShintaiDK]

ShintaiDK I don't understand your post. Is that supposed to falsify the 'at least 10% faster' claim? Are you saying the performance will be less than 10% faster? :hmm:

And does your post try to defend the idea that IPC improvement is a focus? Remember the IGP will probably improve by several hundred % over Ivy Bridge.

Im just saying Fud is a bad site.

If I had to guess I would say 15-20%

 

[BenchPress]

If I had to guess I would say 15-20%

I'm sorry but that's a completely uneducated guess.

Here are the facts. The Haswell architecture will heavily focus on performance/Watt, due to the ultrabook market they wish/need to expand into. Aggressively increasing the IPC compromises performance/Watt, so that is unacceptable. Sandy/Ivy Bridge are already highly tuned for performance/Watt, so there's little or no headroom for further improvement. Last but not least, they are adding AVX2 and TSX to improve performance by exploiting DLP and TLP, instead of ILP, since such technology vastly improves performance/Watt.

So please don't make any guesses without presenting any theory to support it. Exactly what architectural changes would improve IPC by as much as 15-20%, without compromising performance/Watt?

Mark my words. For some workloads Haswell will be over twice as fast for a given power envelope. Any IPC enhancements will play a negligible role in that.

 

[inf64]

I agree with BenchPress. In AVX2/FMA tuned workloads,Haswell may end up as ~2x faster than SB or IB. Anywhere else it will be around of what FUD is saying(FUD quotes intel's press materials,he is not making it up). There is just no room for 20% IPC improvement.Heck SB got 2x the load/store BW and is around ~8% faster at the same clock vs Nehalem in non-AVX workloads.