Intel's 22nm Successors Revealed - Ivy Bridge & Haswell

[Face]

{Credit goes mostly to urbanempire at ocn, thanks for the great sum up }

"Details of Intel's post-Nehalem CPU architecture are emerging. I got this from a French site, but it looks legit.



Some info about Sandy Bridge and Haswell:

Google Translate:

• Sandy Bridge (SNB)

In 2010, architecture Sandy Bridge, formerly called Gesher, will succeed Nehalem. It should be equipped with 8 cores on the same die, a cache of 512 KB L2 by heart and a cache of 16 MB L3 One of the main characteristics Sandy Bridge will be adding the game 'Instruction AVX (Advanced Vectors Extensions), formerly known as the VSSE.

• Haswell (HSW)

Haswell is expected to have these features:

• 22 nm process
• 8 cores by default
• Entirely new cache design
• Revolutionary power saving systems
• Possible on-package vector coprocessors
• The addition of the FMA instruction set

(wiki)

Planned for 2012 or earlier, the family CPU Haswell succeed architecture Sandy Bridge. Gravés to 22 nm, they should include 8-core default, a whole new architecture caches, mechanisms "revolutionary" energy saving and the possibility of board coprocessors vector processing in a single package. Côté instruction set, the sea serpent FMA (Fused Multiply-Add), which allows for simultaneous operation of multiplication and addition via the same instruction, should be implemented.

No other information has filtered by Haswell, but since we are talking about a CPU to be released in 3 generations, it is not surprising. The design of the processor is still in the early stages.

2007: Penryn - 45nm, Quad core (Tick)
2008: Nehalem - 45nm, Quad core with HT, IMC, and QPI (Tock)

2009: Westmere - 32nm, 6-cores with HT, IMC and QPI (Tick)
2010: Sandy Bridge - 32nm, 8-cores with HT, IMC, QPI, and the revolutionary new AVX game instruction technology (Tock)

2011: Ivy Bridge - 22nm (Tick)
2012: Haswell - 22nm (Tock)

(* Not much has been released about the Ivy Bridge and Haswell architectures, but I think this is the first time that their codenames have been publicly released. It's logically to assume, based on the progression that they will contain anywhere from 8 to 16 cores with HT.)

Last but not least is Larabee which apparently will include ray-tracing. Apparently the Larabee ran Quake 4 at 90 FPS at 1280x960. Hopefully more pictures will be forthcoming once IDF starts."

Sources: CanardPlus (translated) via Overclock.net

[perkam]

Inspiration for SandyBridge, I'm sure: http://www.emerchandise.com/images/p...dTYSBS0008.jpg



Interesting to see Intel first going IMC, and now going for large L3 caches

Perkam

[RPGWiZaRD]

I like Intel's tick tock strategy... makes it easier to plan ahead so you'll know what to get. Personally I favor buying the "tick" parts and I might skip Nehalem and wait for Westmere as my next upgrade. By then the extra cores should really start becoming useful in a wide range of apps and DDR3 prices have settled down quite a bit I hope.

Like the big cach sizes in upcoming parts.

[ownage]

Yeah, I like Tick-Tock also.



Intel's goal is to bring out new Processor technology every 2 year and a lower nm after 1 year. So we can expect something entirely new every 2 year without having doubts about what the next CPU's will be.

[y2kbos]

Yeah, I like Tick-Tock also.

I dont like it, as nothing should be set in stone.

Roadmaps have no place in Tech, especially 4 year roadmaps.Tech should be spontaneous.

[bowman]

The only thing that's set in stone is the name and node - if AMD suddenly launches something competitive they can change up the architecture and take more chances if they need to.

[Shintai]

I dont like it, as nothing should be set in stone.

Roadmaps have no place in Tech, especially 4 year roadmaps.Tech should be spontaneous.

Kinda hard to make something spontanious that takes 4-5+ years to develop.

[y2kbos]

Kinda hard to make something spontanious that takes 4-5+ years to develop.

Says who?

The past is the past.

[ownage]

Says who?

The past is the past.

I think you don't understand what he means.

[y2kbos]

I think you don't understand what he means.

I do, spontaneous wasnt the right word, sure.

[Shintai]

I do, spontaneous wasnt the right word, sure.

So you do know the research and development towards a goal in 2012 has to have been started this year or even before? Yet the tick/tock is not something you like? What do you like?

Or should it be something like throwing 6billion$ a year into R&D with the comment. Do what you like, when you like it. And lets hope we reach a goal one day.

[y2kbos]

So you do know the research and development towards a goal in 2012 has to have been started this year or even before? Yet the tick/tock is not something you like? What do you like?

Or should it be something like throwing 6billion$ a year into R&D with the comment. Do what you like, when you like it. And lets hope we reach a goal one day.

They are comfortable, I dont like that.Do you?

Where was there 4 year plan when they were getting a whiping a few years ago?

Its all very convenient

[Nsomniac]

The only thing really set in stone is the process. 45nm to 32nm requires a fab to be built 4-5years out. Some of logic designs can be updated but for the most part they too much conform to the roadmaps.

For instance if you want your memory controller to handle ddr5 and skip ddr4 that can be done, but you also must change the northbridge chipset, and then you may need to change socket design, and then motherboard manufacturers must change their designs.

But, if you want to add new instructions to say SSE or make more cache available that can be done without changing the map.

[tdream]

I like Intel's tick tock strategy... makes it easier to plan ahead so you'll know what to get. Personally I favor buying the "tick" parts and I might skip Nehalem and wait for Westmere as my next upgrade. By then the extra cores should really start becoming useful in a wide range of apps and DDR3 prices have settled down quite a bit I hope.

Like the big cach sizes in upcoming parts.

I'm liking the ticks too.
DDR3 should be reasonable or maybe comparable to what DDR2 is now as well.

They are comfortable, I dont like that.Do you?

Where was there 4 year plan when they were getting a whiping a few years ago?

Its all very convenient

It cost them a lot to get to where they are now, and they aren't about to let the lead go anytime soon.

[Shadowmage]

I don't think you guys have any experience with the industry... tech takes a very long time to develop. For example, Haswell development officially started some time last summer. Generally, a processor design takes around 4-5 years to come to market.

First, you need to develop the architecture, in which the architects propose new modifications to the design. These modifications are then evaluated based on performance, area, power, and difficulty of implementation. The hardware designers provide feedback regard implementation feasibility, etc.

Next, a design is formalized by the HW designers, who take the general architectural model (which gives a rough idea as to what the module does) and define each module exactly, including all the inputs and outputs. Then they write the RTL.

Then, verification of the RTL is needed, both functional verification (does the module do what it's supposed to do) as well as performance verification (does the architecture perform as well as expected). The VLSI engineers start their place-and-route, and depending on the number of bugs (usually in the thousands), RTL writing and placing-and-routing happens many times. Eventually, some cut-off point is determined, and the part tapes out.

After the chip comes back, more verification is needed (usually many more bugs are found in this phase), and the chip may need to be revised multiple times (A0 is the initial tape-out and is usually buggy). Speed paths are found and potentially fixed, as well as hot spots. Eventually, based on some data gathered from marketing, the product is then released to the market while still buggy. Every single chip in production today is buggy; the only question is if the bugs are serious

[Helmore]

I'm liking the ticks too.
DDR3 should be reasonable or maybe comparable to what DDR2 is now as well.

It cost them a lot to get to where they are now, and they aren't about to let the lead go anytime soon.

I prefer the Tocks, but what you like depends on where you're coming from. I plan on buying Nehalem after Christmas, but I'm coming from an AMD Socket 939 system and then Nehalem makes more sense for me. If you have a Yorkfield in you current system then that's a different story for you.

For me the advantages are a better upgrade path, as Westmere will probably (hopefully) be compatible with the X58 chipset and LGA1333 socket. DDR3 will have matured some more by then as well and future memory increases should also be easy.
Although Deneb is shaping up to be a reasonable alternative for me, but that only depends on pricing.

[drjunk]

They are comfortable, I dont like that.Do you?

Where was there 4 year plan when they were getting a whiping a few years ago?

Its all very convenient

I can see both sides of the argument....true, its good to know whats around the corner...but if intel was under the same pressure as they were a few years ago,....i think people would be overclocking there Nehalems by now.

[initialised]

I dont like it, as nothing should be set in stone.

Roadmaps have no place in Tech, especially 4 year roadmaps.Tech should be spontaneous.

Ever heard of Moore's Law?

[[XC] Synthetickiller]

I like Intel's tick tock strategy... makes it easier to plan ahead so you'll know what to get. Personally I favor buying the "tick" parts and I might skip Nehalem and wait for Westmere as my next upgrade. By then the extra cores should really start becoming useful in a wide range of apps and DDR3 prices have settled down quite a bit I hope.

Like the big cach sizes in upcoming parts.

I agree with you. Nehalem looks tempting, but there's nothing my q6600 can't handle still. On top of that like you said, ddr3 is not at a good price point for me. I want good 2x2 gig kits for 150 or less.

I want a 32nm shrink w/ 6 cores. Heck, if I can hold out, 22nm isn't too far off from that. Its a waiting game, but at the same time, i want good, cheap, mobos that OC like crazy.

Price point does it for me.

[xlink]

I prefer the Tocks, but what you like depends on where you're coming from. I plan on buying Nehalem after Christmas, but I'm coming from an AMD Socket 939 system and then Nehalem makes more sense for me. If you have a Yorkfield in you current system then that's a different story for you.

I prefer ticks.the diference between the 65nm tock and the conroe tick was significant. As is the difference between the 45nm tock and the nehalem tick.

the srinks just didn't do too much though. 10% higher clocks, <10% more performance/clock... <20% performance boosts don't impress me as much as 20-100% performance boosts.

[y2kbos]

Ever heard of Moore's Law?

I have and it remains the problem, the industry needs to think more outside the box now and stop talking about Moore's never-existing Law.

[gurusan]

I have and it remains the problem, the industry needs to think more outside the box now and stop talking about Moore's never-existing Law.

yeah, the "industry" is comprised of mindless automatons who never think outside their little box.

Intel should really hire an xs member from the Interwebz. Tech would really move along then.

[CraptacularOne]

They are comfortable, I dont like that.Do you?

Where was there 4 year plan when they were getting a whiping a few years ago?

Its all very convenient

What do you think they were doing? They were planning the Core 2 architecture a few years ago. It's not like the Core 2 architecture was a "spontaneous" development, where one R&D guy just shouted "EUREKA!" Then they rushed off to develop the Core 2 line. They spent years developing the technology.

Sometimes the competition has a better plan or a more efficient design. AMD had the performance crown for a while with their Athlon line, but look at what happened to them. They apparently didn't have as good of a strategy or road map and look where it got them. Struggling to be competitive, no real hope of regaining their lead any time soon and reeling from the market share lose they took. If they had a strategy in mind with an honest effort to advance their architecture they wouldn't be in the situation they are in now. When they moved from s939 to AM2 it should have been more than simply a memory controller revision to DDR2, because honestly, that's all that it brought to the table. The first AM2 CPUs didn't evn benefit from a smaller die size as they were still on 90nm. It was only after another year into the life of AM2 that we saw 65m, parts. But even those did not bring a new architecture to the table. It was only when they released the Phenom line thet they made an honest effort to move forward.

Don't hate the strategy that is clearly leading the pack. After all, they are in the lead so they must be doing something right

[Syn.]

I dont like it, as nothing should be set in stone.

Roadmaps have no place in Tech, especially 4 year roadmaps.Tech should be spontaneous.

Whooo.... are you talking about the same thing as rest of the people because what you are saying is very out-there. In fact its so out there that you might even call it stupid.

[gurusan]

Whooo.... are you talking about the same thing as rest of the people because what you are saying is very out-there. In fact its so out there that you might even call it stupid.

ignorant is what it is.