Intel Corporation's Multicore Architecture Briefing

HardOCP News

[H] News
Joined
Dec 31, 1969
Messages
0
From the front page (click the link for images):

Intel Corporation today discussed upcoming leading edge microprocessors and technologies. Intel's 45nm high-k metal gate leading manufacturing technology is enabling the industry to move to multicore processors in all market segments, and Intel discussed future products with four, six, eight and many computing cores coming to the market.

Gelsinger disclosed details around Intel's multi-processor (MP) servers based on Intel's 6-core processor codenamed "Dunnington" and Intel's new Itanium processor codenamed "Tukwila." Gelsinger discussed current hot enterprise topics, including virtualization and the new SPEC power benchmark for measuring server energy efficiency, in which Intel-based systems hold all of the top 20 spots. He disclosed a number of technical features on two important products for Intel, Nehalem, Intel's next generation processor family and Larrabee, a future Intel product with many cores.
 
Thanks for the linky. :)

Looks like AMD has truly awakened the giant. Nehalem and Tukwila look to be monsters. However, Larrabee is what most interests me. Pity there's not much info there than what we already know.
 
If their "vpu"s are as powerful as their cpu lineup has been I can see some major changes in the pc-gaming industry coming our way.
 
Down in the Nehalem section, it mentions these cache quantities...
32K L1 Data, 32K L1 Instruction (per core)
256K L2 Data/Instruction (per core)
(up to) 8MB L3 (shared)

Conroe/Kentsfield: 128K L1 trace, 128K L1 data (per core) & 4MB L2 shared (per die)
Penryn/Yorkfield: ?K L1 trace, ?K L1 data (per core) & 6MB L2 shared (per die)

I wouldn't have forseen that big of a drop in cache size (at least in the lowest levels). It would even represent a decrease in higher level shared cache, from Penryn's 12MB shared to the 8MB shared on Nehalem.
 
Very exciting stuff on every front. I'm really pretty surprised that Intel doesn't seem to be just sitting around waiting on AMD to catch up. My only complaint is sub-par availability of their current tech.
 
I wouldn't have forseen that big of a drop in cache size (at least in the lowest levels). It would even represent a decrease in higher level shared cache, from Penryn's 12MB shared to the 8MB shared on Nehalem.

Penryn has 6MB shared per dual-core die (Wolfdale & Wolfdale-DP). When you make a Harpertown or Yorkfield, you get 12MB total, but still only 6MB per die.
 
Yeah, unified cache is way better for multithreaded apps - otherwise 12MB caches the same stuff twice.

What I'm really waiting for is quickpath - memory bandwidth will be sick :D

Of course none of this is really "news", but it is nice to see nothing coming off the roadmap as the dates get closer.
 
I think some of those pics would make really cool desktop wallpapers. I have always like the way chips look when photographed. Think we could get some 1920x1200 versions of those pics?
 
Anyone can confirm if Dunningtom will be available for other server platforms besides Caneland? I'm noticing some indications of this possibility on news sites but nothing concrete so far.
 
AMD has awaken the blue gigant...they should have seen this comming :D
That 16 threaded "Nehalem" running in the wild made me sad about my Q6600 :eek: *L*
 
Conroe/Kentsfield: 128K L1 trace, 128K L1 data (per core) & 4MB L2 shared (per die)
Penryn/Yorkfield: ?K L1 trace, ?K L1 data (per core) & 6MB L2 shared (per die)

That's not right.
Firstly, Core2 does not use tracecache, but conventional code cache (Netburst used tracecache).
Secondly, it is 64K combined L1 cache, divided in 32K code cache and 32K data cache (per core, Nehalem numbers are also per core).
As far as I know, Penryn has the same amount of L1 cache, only larger L2 cache.

So L1-cache is not changed from Conroe/Penryn to Nehalem.
The L2-cache may be smaller, but it serves a slightly different purpose. It is now per-core, instead of shared by 2 cores.
Instead, there is now L3-cache shared between cores, and this L2 acts as a buffer between L1 and L3 cache. So this caching system has a different hierarchy (which is probably better suited to 4+ cores than the old Core2 hierarchy).
So the total amount of cache doesn't change by much, but it is distributed differently.

Especially with L1-cache, the amount of cache alone doesn't say all that much. For example, Athlon has 128k L1 cachet total (64k code/64k data), but because of its half-duplex nature and higher latency, it is not competitive with Core2's L1-cache.
The Pentium 4 only had 8k L1 cache originally (for data, and 12k uOp tracecache, which cannot be expressed in bytes directly, because you don't store bytes, you store uOps, of which the size and format is unknown, and not relevant anyway), and later got 16k.... Because of the high clockspeeds that Intel was aiming for, it wasn't possible to use big caches. They needed cache that was fast enough to operate at these high clockspeeds, and not bottleneck the CPU with too high latencies.
 
I am really looking forward for Nehalem. I am already saving up for a Nehalem system to replace my aging Opty. I figure Nehalem + the true next gen Nvidia GPU will be out about the same time early 2009.
 
If their "vpu"s are as powerful as their cpu lineup has been I can see some major changes in the pc-gaming industry coming our way.

There's no doubt that their VPU will be powerful... question is, will it deliver faster graphics?
Firstly Intel is taking a completely different approach from nVidia and ATi, secondly Intel doesn't yet have a good driver base.
They seem to want to offer developers a way to program the architecture more directly, bypassing OpenGL or Direct3D... Probably similar to Cuda, except that nVidia didn't mean Cuda to be used for rendering itself, but rather to be used in combination with an OpenGL or D3D renderer (or standalone when no rendering is required).

Also, Intel seems to be steering towards raytracing... but there's no way you will run OpenGL or D3D software efficiently on a raytracing backend... So we'll have to see where Intel wants to take this.
 
That's not right.
Firstly, Core2 does not use tracecache, but conventional code cache (Netburst used tracecache).
Apologies. I apparently quit thinking and typed the wrong word.
 
These multi-core cpu's are nice, but when is the software going to catch up w/the dual cores we have now, not ot mention, the triple and quad ones?

.
 
These multi-core cpu's are nice, but when is the software going to catch up w/the dual cores we have now, not ot mention, the triple and quad ones?

.

I think 64-bit support will come first.
More and more people run into the 4 GB barrier now, and they are seeing that Vista x64 is really just the same OS as the 32-bit version, only with better support for large amounts of memory.
I think that is a more pressing issue than performance at this time.
If software really needs the performance of multiple cores, in most cases it already supports it (eg video editing. 3d rendering, gaming).

Besides, Larrabee is not aimed at regular software. It's for special purposes, like Cuda.
Let me ask you this: what exactly are you missing from current software you're using, that multicore would fix?
 
Most newer games at least make use of two cores reasonably well, some even scale very well to 4 cores, and in general the GPU is the bottleneck, not the CPU.

Most new games is a lie. Very select few games take advantage of two cpu's let alone 4 and even then the code to utilize it all isnt always very efficient.

And its true that the gpu is usually the bottleneck but one cant help that if you were to alleviate some pressure off of the gpu and put it onto those idling cores just exactly how bottlenecked games would really be.

Hell for all we know you could find people rolling around with q600's paired up with a 8600 throwing down with the same settings and fps as people with e8400's and a 8800gts. Grant it thats a bit far fetched but I think you get the idea I'm trying to paint here.
 
Most new games is a lie. Very select few games take advantage of two cpu's let alone 4 and even then the code to utilize it all isnt always very efficient.

Is it? Source engine is multithreaded now, UnrealEngine 3 is, the new Crytek engine is, that pretty much covers most shooters. Games like SupCom are also multithreaded, then there's the upcoming Alan Wake etc.
I'd say most new games are covered.

And its true that the gpu is usually the bottleneck but one cant help that if you were to alleviate some pressure off of the gpu and put it onto those idling cores just exactly how bottlenecked games would really be.

It's not possible to (efficiently) combine CPU and GPU rendering. Besides, even the fastest quadcore is still orders of magnitude slower than a modern GPU at rendering (just look at how slow Intel's IGPs are, mainly because they do vertexprocessing on the CPU). Even if you could combine them, the CPU would have virtually no impact on rendering speed.
So your idea won't fly. Which brings it back to games being GPU-limited, and the GPU is the only solution, more CPU cores or better multithreaded code won't have any effect.
 
I think 64-bit support will come first.
More and more people run into the 4 GB barrier now, and they are seeing that Vista x64 is really just the same OS as the 32-bit version, only with better support for large amounts of memory.
I think that is a more pressing issue than performance at this time.
If software really needs the performance of multiple cores, in most cases it already supports it (eg video editing. 3d rendering, gaming).

Besides, Larrabee is not aimed at regular software. It's for special purposes, like Cuda.
Let me ask you this: what exactly are you missing from current software you're using, that multicore would fix?

There's some software I use at work on my Dell Precision M90 laptop, that takes forever to run. It has a 2.16 GHz Core 2 Duo CPU, w/4 gb (3.25 gb used), w/Windows XP Pro. When I check the Task Manager/Performance Meter, only one core is running. Before you ask, the cpu is running at 2.16 GHz, when plugged in.

.
 
There's some software I use at work on my Dell Precision M90 laptop, that takes forever to run. It has a 2.16 GHz Core 2 Duo CPU, w/4 gb (3.25 gb used), w/Windows XP Pro. When I check the Task Manager/Performance Meter, only one core is running. Before you ask, the cpu is running at 2.16 GHz, when plugged in.

Sure, but that's probably specialized software.
We've reached a point where most software is already multithreaded (databases, web/file/mail/etc servers, multimedia processing, 3d rendering, scientific computing etc), or isn't required (webbrowsers, email clients, Office applications etc).
Therefore I say that for most people, it's not an issue at this point. They either have it, or don't need it. 64-bit isn't that far yet. Server software tends to be 64-bit by now, but there's a lot of desktop applications that would benefit from 64-bit aswell, but isn't yet (mostly rendering/modeling/audio/video processing). These applications are generally already optimized for multithreading.

Ofcourse there will always be software that is poorly written and doesn't use multiple cores, but in many cases there are alternatives that do.
 
Back
Top