Larrabee not quite dead

[Phynaz]

Looks like Larabee's grandchild, Knights Corner, may be real.


http://www.hpcwire.com/hpcwire/2011...coprocessor_at_supercomputing_conference.html

 

[podspi]

Knights Corner, he said, would arrive in 2012 using Intel's newly hatched tri-gate 22nm process node. With perhaps an indirect inference to NVIDIA's and AMD's GPU computing prowess, Neal-Graves noted that they'll be able to use their 22nm technology to deliver cheaper, faster and more power-efficient silicon than their competition, adding, "That's really going to bring the performance to the table that we really need for these types of solutions."

Wow! I feel like the ramp up to 22nm for all of Intel's products is occurring really quickly. Am I just imagining this, or are they changing the way they allocate capacity? I feel like it used to be only their top-of-the-line CPU got the new process. Now we have mid-range, Atom, and whatever you want to call this

 

[garagisti]

Erm, Larrabee as it was intended IS very much dead.

However, you'd be right if you said, Intel has a new variation to offer sometime in future.

 

[zephyrprime]

Larrabee minus any graphical ability would actually be quite useful for cloud servers, super computers, web server farms, and that sort of thing.

 

[Idontcare]

"You end up with two kinds of customers, one highly satisfied with [AMD and Nvidia graphics] accelerators because despite the tedious porting process, their results are very good," Goh said.
"Others feel their time spent on porting [their apps to AMD and Nvidia chips] doesn’t justify the performance and there is a huge part of this second group for whom MIC is useful—and ultimately some of the first group may want MIC, too," he said.

http://www.eetimes.com/electronics-news/4217092/OEMs-show-systems-with-Intel-MIC-chips

I do like the compiler convenience.

50 cores though, I really expected more than that given that this is a 22nm variant. If they had 32 cores on 45nm then 50-64 cores would have been expected for 32nm.

They are probably bandwidth limited though, 50 cores and all.

 

[lol123]

Actually, Larrabee was never cancelled and Intel CEO Paul Otellini has made several statements to the contrary. Knights Corner has been in the roadmaps for release on 22nm all along, and at least according to Anandtech none of the graphics-specific features have been taken out from the design.

"… In terms of Larrabee, we did not stop the project. If we made any mistake with Larrabee, we probably should not have talked about something that was high risk and long term. We have not stopped the project. We have shipped STVs out. We're looking at how and when to bring it to market. It still has very very high promise in areas of throughput computing and in terms of a general reprogrammable graphics engine using small IA cores. We still like the idea. But we've taken the risk associated with a new architecture out of our roadmap over the next few years so we have the flexibility to stay competitive while still working on it."

It definitely made sense to cancel the first implementations of the architecture though. Anyone who thought that Larrabee could be turned from a Powerpoint product as a completely new microarchitecture into a consumer-oriented end user video card in just a few years was insane to be honest.

 

[Idontcare]

Anyone who thought that Larrabee could be turned from a Powerpoint product as a completely new microarchitecture into a consumer-oriented end user video card in just a few years was insane to be honest.

Sure, but in their defense they were kinda guided to have those expectations by the same folks who want to be taken seriously when they say Itanium will not be killed, that Intel was getting in the mobile phones (Xscale era), was getting into HDTV, and that Atom is taking over the mobile sector before ARM gets there...

If your point is to say that even the Intel folks are insane and should be given no more credibility than that afforded an insane person then I'm onboard with your prognosis :thumbsup:

 

[TuxDave]

Wow! I feel like the ramp up to 22nm for all of Intel's products is occurring really quickly. Am I just imagining this, or are they changing the way they allocate capacity? I feel like it used to be only their top-of-the-line CPU got the new process. Now we have mid-range, Atom, and whatever you want to call this

Yes and it's a very facinating process development as well. Before the process was defined for the highest performing part. Now with Atom in play, you have two conflicting metrics that the process has to find a middle ground for.

 

[Kingkazma]

i would buy this for rendering.

gaming? no

but rendering? yes.

 

[Phynaz]

Yes and it's a very facinating process development as well. Before the process was defined for the highest performing part. Now with Atom in play, you have two conflicting metrics that the process has to find a middle ground for.

Wondering if that results in having to make additional compromises in the process in order to make it more 'general'.

 

[Idontcare]

Wondering if that results in having to make additional compromises in the process in order to make it more 'general'.

Intel might be unique, it would not be the first time, but the model adopted by the rest of the industry for IDM's with a need for dual-purpose (or multi-purpose as in 3 or more) flavors of process technology on a given node is that they delineate them by xtor targets and designs and divy up the work into so-called sub-nodes.

So you have a low-power subnode, a midpower subnode, and a high-performance subnode.

They'll all share the same BEOL (in its usually lego connect fashion, skipping some metal levels if they are not needed and so on), and they can mix-and-match xtors (adding masksets) if need-be.

But you'll have your FEOL that is entirely designed for the low-power sipper IC's and then your FEOL that is entirely designed for the high-performance stuff.

The juggling and compromises that come into play are more in terms of priorities for development milestone timeline. Who gets the R&D budget to procure more R&D wafers for faster/wider learning cycles in the fab on the pilot line, etc.

It's not the kind of compromise that I think you are thinking of in which the high-performance guys and the low-power guys drive a development program that results in a universal xtor that can be the jack of all trades.

At least this is my experience from the rest of the industry (specifically TI, AMD, NatSemi, Moto/Freescale,Philips/NXP, Lucent/Agere, UMC, TSMC, IBM, Chartered, SMIC, Samsung) but I readily admit I have zero information or experience about how Intel is managing this side of their new dual-purpose process development model. They so many other things their own special way, who knows maybe this is something new for the industry to get use to as well!

 

[drizek]

Yes and it's a very facinating process development as well. Before the process was defined for the highest performing part. Now with Atom in play, you have two conflicting metrics that the process has to find a middle ground for.

Wasn't 32nm supposed to start with Atom in late 2009? Whatever happened to that?

 

[GammaLaser]

Wasn't 32nm supposed to start with Atom in late 2009? Whatever happened to that?

Atom has always lagged behind in terms of process technology, only recently has Intel announced that it will now move Atom at "double the pace of Moore's Law" [1]by moving it to a new process generation every year until it lines up with the rest of the x86 development.

 

[TuxDave]

It's not the kind of compromise that I think you are thinking of in which the high-performance guys and the low-power guys drive a development program that results in a universal xtor that can be the jack of all trades.

Yeah I wouldn't know what to do if we were forced to use an ultra low power xtor for high performance parts. On the other hand, there are still some conflicting metrics that cause Core vs "other" grief.

 

[Nemesis 1]

http://newsroom.intel.com/servlet/JiveServlet/download/2152-4-5220/ISC_Intel_MIC_factsheet.pdf

 

[Idontcare]

http://newsroom.intel.com/servlet/JiveServlet/download/2152-4-5220/ISC_Intel_MIC_factsheet.pdf

System Config: Intel Shady Cove Software Development Platform

Really!? Shady Cove? They couldn't come up with a less sinister sounding platform name?

What's next? Benchmarking results optimized with Intel's latest edition of Dark & Smokey Back Room Deals Compiler Suite?

Also we can see why Larrabee the discrete GPU was cancelled. Their 45nm Aubry Isle C0 silicon is running 1.2GHz w/2GB GDDR5 @ 3.6GT/s and the card pumps out only 925GFlops SGEMM.

(for reference, a GTX480 gets around 840 Gflops SGEMM, a GTX580 gets 950)

 

[Mopetar]

Wasn't 32nm supposed to start with Atom in late 2009? Whatever happened to that?

There was no competition for it at that time and it was probably more profitable to use their 32 nm production capacity for their other product lines.

 

[Nemesis 1]

Really!? Shady Cove? They couldn't come up with a less sinister sounding platform name?

What's next? Benchmarking results optimized with Intel's latest edition of Dark & Smokey Back Room Deals Compiler Suite?

Also we can see why Larrabee the discrete GPU was cancelled. Their 45nm Aubry Isle C0 silicon is running 1.2GHz w/2GB GDDR5 @ 3.6GT/s and the card pumps out only 925GFlops SGEMM.

(for reference, a GTX480 gets around 840 Gflops SGEMM, a GTX580 gets 950)

I laughed so hard at this I almost wet myself. So you don't think its enough. Its on 45nm and being reworked for 22nm working with intel cpus using AVX and what ever else Intel has in mind it should do rather well . I found the whole paper interesting . Of special interest to me was how fast developers were able to get their programms up and running on it .

 

[GammaLaser]

Of special interest to me was how fast developers were able to get their programms up and running on it .

Yeah, one of the big selling points is its programmability/ISA familiarity, so no surprise that they emphasized it so much.

 

[podspi]

Really!? Shady Cove? They couldn't come up with a less sinister sounding platform name?

What's next? Benchmarking results optimized with Intel's latest edition of Dark & Smokey Back Room Deals Compiler Suite?

Also we can see why Larrabee the discrete GPU was cancelled. Their 45nm Aubry Isle C0 silicon is running 1.2GHz w/2GB GDDR5 @ 3.6GT/s and the card pumps out only 925GFlops SGEMM.

(for reference, a GTX480 gets around 840 Gflops SGEMM, a GTX580 gets 950)

Are you being sarcastic? Or am I missing something?? GTX480 is nothing to sneeze at...

 

[Idontcare]

I laughed so hard at this I almost wet myself. So you don't think its enough. Its on 45nm and being reworked for 22nm working with intel cpus using AVX and what ever else Intel has in mind it should do rather well . I found the whole paper interesting . Of special interest to me was how fast developers were able to get their programms up and running on it .

At 22nm it most certainly will be enough, as a 45nm product that was intended to intersect the discreet GPU market that held 40nm product boasting of already comparable results (and Larrabee was still a year out from product introduction at the time) we see that it most certainly was not going to be enough.

Are you being sarcastic? Or am I missing something?? GTX480 is nothing to sneeze at...

Well I wasn't being sarcastic, that was my point, the GTX480 was more than competitive with, and a year ahead of Larrabee in getting to the market, so Intel really had no choice based on these numbers to throw in the towel at 45nm.

Sure they could have released it anyway, but at best it was going to come up well shy of reaching the halo product at the time which is not what you want to do if you need the product to command high ASP's to enable 50% gross margins.

It'll do well, gross margin wise, in its targeted space. HPC is usually big budget stuff. Look at the price of Tesla.

 

[podspi]

Sure they could have released it anyway, but at best it was going to come up well shy of reaching the halo product at the time which is not what you want to do if you need the product to command high ASP's to enable 50% gross margins.

It'll do well, gross margin wise, in its targeted space. HPC is usually big budget stuff. Look at the price of Tesla.

I keep forgetting about those pesky margin requirements... I would love to have seen how it actually would have performed.

 

[Nemesis 1]

Ya the last 2 fall IDF were boring this years should be more interesting . If Intel can get Knights Corner performance up to were they are pleased at reasonable power consumption it should do really will paired with SB-e and AVX. I am looking to a 3x improvement over what this PDF shows and I believe I am being conservative. Not long now comparred to when PAT started hyping this thing

 

[dealcorn]

Wow! I feel like the ramp up to 22nm for all of Intel's products is occurring really quickly. Am I just imagining this, or are they changing the way they allocate capacity? I feel like it used to be only their top-of-the-line CPU got the new process. Now we have mid-range, Atom, and whatever you want to call this

It is not your imagination, it is not an allocation decision in the sense you mean and Intel has communicated this to the financial community because what you are seeing is the result of a multi year commitment to increase capital spending so it can transition pretty much everything to their current generation fab process really fast. Intel's sense is that it has the worlds best fab process and it wants to exploit that advantage in the marketplace ASAP. If it works, they make more money and if it does not they suffer the consequences.

22 nm is the start of this new strategy and it really kicks into gear at 14 nm. Time will tell whether this is a multi billion dollar boo boo or home run.

 

[krumme]

It is not your imagination, it is not an allocation decision in the sense you mean and Intel has communicated this to the financial community because what you are seeing is the result of a multi year commitment to increase capital spending so it can transition pretty much everything to their current generation fab process really fast. Intel's sense is that it has the worlds best fab process and it wants to exploit that advantage in the marketplace ASAP. If it works, they make more money and if it does not they suffer the consequences.

22 nm is the start of this new strategy and it really kicks into gear at 14 nm. Time will tell whether this is a multi billion dollar boo boo or home run.

Well do Intel have a choice here?

I doubt this is good profit, using new expensive process for 10usd parts. I think that is what is reflected in the share price.

But what should Intel do? Bobcat on 40nm gets dirt cheap, and the 28nm hkmg will kill the 32nm Atom. So they have to resort to their superior process, if they want to be on the market at all. Its a no brainer, - Intel want to stay on the broad consumer market and keep brand awareness high.

In situations like this Otellini will always think how he can get monopoly like situations (its his job to think like that), and he will say entry barriers. The problem this time is, what is good enough is not constantly raising, and therefore the benefit of Intels process superiority will fade.

But anyway, protecting core business is vital, so Intel did the right thing. But its going to be expensive, and they hate it