HASWELL core I5: Performance Preview

[Homeles]

Voltage regulator isn't going add .1w to the TDP. Voltage regulators are just a couple of transistors and diodes.

Pretty much every site that has covered Haswell and every leak regarding the architecture disagree with you.

And "a couple of transistors and diodes?" You are grossly understating the circuitry involved.

 

[Lepton87]

I havent seen anything on the performance of our next generation of processors but it seems to strike me a bit odd that it is running at 2.6GHz at 83w. We have the Intel® Core™ i5-3330S running at base 2.7GHz before turbo at 65w. Also the voltage is almost 25% lower than our 3rd generation Intel Core processors (1.1v down to .8v), it looks like it being downclocked but heck I really don't know.

So which Intel employees are privy to such information? Engineers? I guess Paul is a sure bet

 

[mrmt]

I think a 10-core SR will match a 6-core 12-thread IB part. However, AMD wont make it. I doubt they will continue with FX series. 6-core Kaveri will probably walk over a 3570K in MT workloads, why bother with FX..

And how much power you would need for such a monster?

It's not only the core, but also the cache, the uncore, beefier memory controller, etc. All this costs both power and extra die dize, things that AMD cannot afford right now. And I'm not even counting for the diminishing returns of extra cores.

And remember that Steamroller won't have to deal with Haswell, but with Broadwell. AMD will be full two nodes behind Intel in 2014, not a good prospect for a die/power hog like a 10C Steamroller.

 

[parvadomus]

And how much power you would need for such a monster?

It's not only the core, but also the cache, the uncore, beefier memory controller, etc. All this costs both power and extra die dize, things that AMD cannot afford right now. And I'm not even counting for the diminishing returns of extra cores.

And remember that Steamroller won't have to deal with Haswell, but with Broadwell. AMD will be full two nodes behind Intel in 2014, not a good prospect for a die/power hog like a 10C Steamroller.

The same as 8350. Its 28nm plus power tweaks, and new arch. But as I said AMD wont do it, I think they will stick to 6 or 8 cores at most.

 

[pm]

So which Intel employees are privy to such information? Engineers? I guess Paul is a sure bet

As far as I'm aware, all Intel employees have access if they know where to look - there's actually lots of internal websites that list out all sorts of cool things that aren't generally common knowledge... it's not like there's a Google of Intel's internal website so it's difficult to know where to look unless someone tells you. In this case there's an internal website called "mark" (I have no idea why) that lists everything there. All of the specs for Haswell products including all of the product names and specifications are listed there.

 

[Homeles]

Its 28nm

It's 28nm bulk. Their 32nm process is SOI. They're taking a step forward, while simultaneously taking a step back.

 

[BallaTheFeared]

I'm still hopeful it will be faster than my old i5-2500ks and use less power doing it.


Also hoping to see some bclk and uncore overclocking for a bit more "tweaking" fun.

 

[Arkaign]

If it were even remotely possible to make a 300W chip that was twice as fast as a 2700K @ 5GHz (in IPC terms) then you can bet intel would be selling it. But that is like asking for an 7 GHz wolfdale back in 2007. (7 Ghz is about how fast a wolfdale would need to be to produce a 5GHz sandy bridge super pi single thread score.)

If you were to design a custom 200W peltier which perfectly fit a 4770K's die (in 3 dimensions), I bet you could theoretically run it at near 7Ghz, giving you a ~300W chip that is nearly twice as fast as a 2700K. It would be close, but surely in the ballpark of twice as fast.

I was speaking somewhat hyperbolically

But yeah, I understand, I just wish a little more emphasis on performance was there. Maybe it will show up in the OC limits even if IPC and stock clocks are mediocre.

 

[Phynaz]

Pretty much every site that has covered Haswell and every leak regarding the architecture disagree with you.

And "a couple of transistors and diodes?" You are grossly understating the circuitry involved.

Got anything to back that up? Voltage regulators are very simple circuits.

 

[OCGuy]

Check the roadmap, Haswell fills the mainstream and bottom part of Premium.

 

[Homeles]

Got anything to back that up? Voltage regulators are very simple circuits.

Why don't you tell me why we put heatsinks on them if they're drawing less than ".1W?"

Tell me that this is a simple circuit:
http://www.xbitlabs.com/picture/?src=/images/news/2012-12/haswell_power_1.png

Then tell me that this 106mm² chip will be drawing less than 0.1W:
http://www.xbitlabs.com/picture/?src=/images/news/2012-12/haswell_power_2_550.png

Doesn't matter if it ends up on Haswell or not. Its existence defeats your ludicrous argument. Individual MOSFETs may be simple, but Intel's doing a lot more than sticking one on Haswell and calling it a day.

 

[BallaTheFeared]

I always thought it was because they regulated the voltage by having the power pass through them, not that they actually draw power, but it still goes through it.

Like voltage through a wire, it's not drawing the power but it sure will get hot.

But I'm probably wrong.

 

[Homeles]

I always thought it was because they regulated the voltage by having the power pass through them, not that they actually draw power, but it still goes through it.

Like voltage through a wire, it's not drawing the power but it sure will get hot.

But I'm probably wrong.

That doesn't really fit in with the the first law of thermodynamics. That heat has to come from somewhere, be it electrical resistance through wires, leakage, or normal power consumed when switching a transistor.

 

[inf64]

Like voltage through a wire, it's not drawing the power but it sure will get hot.

This HAS to be sigged! Priceless post, golden :biggrin:

 

[BallaTheFeared]

/\

Did I make a stupid?

 

[Homeles]

/\

Did I make a stupid?

Well, wires can act as a heatsink and get hot due to the components they are connected to drawing a lot of power.

 

[ShintaiDK]

/\

Did I make a stupid?

VRM efficiency on regular VRM is around 85-90%. So if your CPU uses 77W, you might draw around 85W. Those 8W difference are lost in the VRM.

However Intels on package VRM is much more efficient. And size is the key. Less resistance to pass.

 

[386user]

http://vr-zone.com/articles/meet-haswell-s-graphics-thump-card-crystalwell-gt3-edram-/19604.html

disappointing if true

 

[BallaTheFeared]

VRM efficiency on regular VRM is around 85-90%. So if your CPU uses 77W, you might draw around 85W. Those 8W difference are lost in the VRM.

However Intels on package VRM is much more efficient. And size is the key. Less resistance to pass.

Right the heat is part of the inefficiency.


This is why Intel is going with onboard vrms, they'll be more efficient and have more phases/more control over individual areas.

The on package vrms for Haswell should really help push it's efficiency.

I guess the question is does the cpu now use more power because of the power loss is occurring inside the chip instead of on the board. I'm guessing it's just semantics, since there will be less loss with the vrms in the chip. Plus with the design Intel is using they'll have far more control over each part of the chip and how much voltage it gets and when.

CPU usage might be higher than what it would be if the vrms were on the board (though this is unlikely if Intel succeeds at gating and more precise voltage control with it), however platform usage will go down because of it.

 

[lagokc]

Even if integrating the VRMs means lower total system power usage wouldn't that still mean more heat that needs to be removed from the small CPU area? Seems like a bad thing for casual overclockers, though the more precise voltage regulation might actually benefit people looking to set record clockspeeds.

 

[Piroko]

I don't think the actual vrms are that much more efficient - apart from being able to right-size them better and switch faster between stages. With the larger regulators I know you typically trade between efficiency, regulation stability and price.

 

[Lepton87]

As far as I'm aware, all Intel employees have access if they know where to look - there's actually lots of internal websites that list out all sorts of cool things that aren't generally common knowledge... it's not like there's a Google of Intel's internal website so it's difficult to know where to look unless someone tells you. In this case there's an internal website called "mark" (I have no idea why) that lists everything there. All of the specs for Haswell products including all of the product names and specifications are listed there.

Thanks for the extensive answer, so IntelEnthusiast just doesn't care enough to check.
Not so enthusiastic after all, is he? Half of this forum would get an orgasm just perusing through that stuff.

 

[frozentundra123456]

Thanks for the extensive answer, so IntelEnthusiast just doesn't care enough to check.
Not so enthusiastic after all, is he? Half of this forum would get an orgasm just perusing through that stuff.

Probably what he cares most about is keeping his job, not losing it by giving out confidential information before the NDA is lifted. (Not sure if you were just joking, but I assume not, so I gave a serious answer.)

Actually I am very impressed by his posts in general, very upfront about his affiliation and gives guarded and carefully thought out answers, not some type of fan-boy pro-Intel comments.

 

[pm]

Thanks for the extensive answer, so IntelEnthusiast just doesn't care enough to check.
Not so enthusiastic after all, is he? Half of this forum would get an orgasm just perusing through that stuff.

I didn't say anything like that. I said that I thought most employees have access but the site is not very well known - in fact, I wouldn't even want to guess how few people that I work with know of the existence of that internal website but it's a very low number.

And I agree with FrozenTundra. I have access to the site, I actually worked on the design of Haswell, and I'm not going to say much of anything about it. The SKU product line has not, to the best of my knowledge, been announced, and I have absolutely no intention of bringing down lots of negative attention on me by people within my company by disclosing it here on Anandtech. IntelEnthusiast does a fantastic job in my opinion discussing the things he can discuss - he's certainly better about it than me... my general tactic is simply not to say anything and err on the side of caution/paranoia.

 

[mrmt]

IntelEnthusiast does a fantastic job in my opinion discussing the things he can discuss - he's certainly better about it than me... my general tactic is simply not to say anything and err on the side of caution/paranoia.

+1 here. I never ever discuss a post containing the company I work, worked or had direct involvement in financial operations/M&A. My thumbs up to those who can walk this thin ice without falling.