Intel Nehalem, Bloomfield has 8MB of cache

jmke · 24th October 2007, 15:15

To all of our surprise the future Nehalem processors with four cores and eight threads will have 8MB of cache memory. Yorkfield has 12MB, or should we say two times 6MB as this is still dual chip stitched together chip. Each core in Yorkfield has 3MB cache and it looks that Nehalem will have 8MB.

http://www.fudzilla.com/index.php?op...11&Ite mid=35

24th October 2007, 16:00

Sounds about right. Shouldn't be a need for a huge cache with an IMC, but looks like they kept some of the cache for HT use. I'm very curious how big the die is compared to a 45nm Yorkfield...

24th March 2008, 08:02

so exactly how will an integrated memory controller help speed.... i imagine it will be easier for the cpu to communicate with the ram.... but how much faster will this make it... i know amd has been doing it but don't know that much about it. if you guys could explain or point me in the right direction of an explanation it would be greatly appreciated. thanks.

jmke · 24th March 2008, 10:39

reduce latency, faster data exchange = speedier CPU

24th March 2008, 15:32

The entire point of thise insanely large 12mb L2 caches, even 16MB caches is to keep the CPU fed with data to crunch.

If you look at AMD K7 processor verses AMD K8 processor benchmarks you will see the improvement integrating the memory controller will bring... it is the biggest reason for why AMD's K8 easily defeated the Pentium.

For example... Imagine you and two other people in a building... You need to complete a project but lack the information. To get the info you must ask the 2nd person to go talk to the 3rd person to get your info and bring it back to you. If you keep havingto ask questions or you find you don't have all the info you can see how slow this would make completing your project...

The middle guy has to act as the go between for you to get the information you need... Until now the middle guy has always been the chipset, the CPU could never directly talk to or update the memory.

geoffrey · 24th March 2008, 20:05

I don't suspect so see performance boost like with K7->K8, with C2D reducing memory latency's at FSB400 doesn't help that much.

24th March 2008, 20:47

It was my understanding that the FSB has an inherent latency penalty already, and that aggressive memory prefetchers only hid the latency issue.

So increasing the FSB would not compensate any for the inherent latency penalty. FSB is just a general bus where everything in the system including other processors use it together, they all share the same FSB. Just like with PCI bus, all the cards used the same PCI bus and would hinder each other's performance.

With QPI not only will there be a dedicated point-to-point interconnect between the CPU and chipset, there will bea dedicated connection directly to the main memory. No waiting on other cores and other processors to be done with the FSB anymore. Since Nehalem will also be a native quad design there won't be any added penalty for having cores mesh coherency traffic over the FSB either.

Intel chips still scale amazingly worse compared to AMD Opterons in dual/quad socket servers, even if they still overall perform better against AMD chips... this could only be from the FSB/lack of IMC.

C2D is far more data hungry than K8 ever could hope to be. Nehalem will be even more so, Intel widened the instruction pipe even further to accomodate 4 cores executing 8 threads... I suspect this may contribute to the gains seen

Quote:

Originally Posted by Anandtech

Nehalem allows for 33% more micro-ops in flight compared to Penryn (128 micro-ops vs. 96 in Penryn), this increase was achieved by simply increasing the size of the re-order window and other such buffers throughout the pipeline.

With more micro-ops in flight, Nehalem can extract greater instruction level parallelism (ILP) as well as support an increase in micro-ops thanks to each core now handling micro-ops from two threads at once.

geoffrey · 24th March 2008, 22:58

For server/workstation, the bandwidth of the CSI bus is just huge, combined with the on-die memory controller Nehalem will certainly offer a nice gain compared to C2D. On the tech side of things, Nehalem is spectacular, but not for home users... probable not enough data to feed.

25th March 2008, 17:01

We haven't run out of data yet...

Even single-threaded applications are going to see a performance improvement.

Thought this was interesting: http://blogs.zdnet.com/Ou/?p=1025 Especially for Shanghai's results.

geoffrey · 25th March 2008, 18:09

To bad they forgot to mention clock rates

Still, at this moment dual core is still favorable over Quads, unless software dramatically changes over the following half year I don't think many home users will need it, maybe in 2 years or so... Theoretically, single core apps may be faster, but will you notice the difference compared with the E8x00 which is doing all ready very good? If so, where?

24th October 2007, 15:15	#1
jmke Madshrimp Join Date: May 2002 Location: 7090/Belgium Posts: 79,022	Intel Nehalem, Bloomfield has 8MB of cache To all of our surprise the future Nehalem processors with four cores and eight threads will have 8MB of cache memory. Yorkfield has 12MB, or should we say two times 6MB as this is still dual chip stitched together chip. Each core in Yorkfield has 3MB cache and it looks that Nehalem will have 8MB. http://www.fudzilla.com/index.php?op...11&Ite mid=35 __________________

24th March 2008, 10:39	#4
jmke Madshrimp Join Date: May 2002 Location: 7090/Belgium Posts: 79,022	reduce latency, faster data exchange = speedier CPU __________________

25th March 2008, 17:01	#9
Kougar Posts: n/a	We haven't run out of data yet... Even single-threaded applications are going to see a performance improvement. Thought this was interesting: http://blogs.zdnet.com/Ou/?p=1025 Especially for Shanghai's results. Last edited by Kougar : 25th March 2008 at 17:03.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Power gating and turbo mode: Intel talks Nehalem at IDF	jmke	WebNews	0	20th August 2008 16:04
NVIDIA Brings SLI Technology to Intel Bloomfield CPU Platforms	jmke	WebNews	1	14th July 2008 23:13
First Intel Nehalem has 8MB cache	jmke	WebNews	3	27th May 2008 01:53
Intel Core 2 Extreme QX6950 coming Q4, 3.33GHz, 12MB L2 cache	jmke	WebNews	0	27th July 2007 16:46
Intel Bloomfield To Hit > 4GHz and Has 8MB Cache	jmke	WebNews	3	19th December 2006 11:44
Most efficient L2 cache usage: Intel shared or AMD dedicated?	jmke	WebNews	0	23rd October 2006 14:04
Intel Preps 3.50GHz Server Chips with 8MB Cache	jmke	WebNews	0	14th January 2005 10:58
Intel Itanium 2 to Get Additional Clock-Speed, Cache	jmke	WebNews	0	25th May 2004 16:21
Intel Heads Towards 4MB Cache in Desktop Processors	jmke	WebNews	0	17th March 2004 09:36

24th October 2007, 16:00	#2
Kougar Posts: n/a	Sounds about right. Shouldn't be a need for a huge cache with an IMC, but looks like they kept some of the cache for HT use. I'm very curious how big the die is compared to a 45nm Yorkfield...

24th March 2008, 08:02	#3
jakebot Posts: n/a	so exactly how will an integrated memory controller help speed.... i imagine it will be easier for the cpu to communicate with the ram.... but how much faster will this make it... i know amd has been doing it but don't know that much about it. if you guys could explain or point me in the right direction of an explanation it would be greatly appreciated. thanks.

24th March 2008, 15:32	#5
Kougar Posts: n/a	The entire point of thise insanely large 12mb L2 caches, even 16MB caches is to keep the CPU fed with data to crunch. If you look at AMD K7 processor verses AMD K8 processor benchmarks you will see the improvement integrating the memory controller will bring... it is the biggest reason for why AMD's K8 easily defeated the Pentium. For example... Imagine you and two other people in a building... You need to complete a project but lack the information. To get the info you must ask the 2nd person to go talk to the 3rd person to get your info and bring it back to you. If you keep havingto ask questions or you find you don't have all the info you can see how slow this would make completing your project... The middle guy has to act as the go between for you to get the information you need... Until now the middle guy has always been the chipset, the CPU could never directly talk to or update the memory.

24th March 2008, 20:05	#6
geoffrey [M] Reviewer Join Date: Nov 2004 Posts: 4,127	I don't suspect so see performance boost like with K7->K8, with C2D reducing memory latency's at FSB400 doesn't help that much.

24th March 2008, 22:58	#8
geoffrey [M] Reviewer Join Date: Nov 2004 Posts: 4,127	For server/workstation, the bandwidth of the CSI bus is just huge, combined with the on-die memory controller Nehalem will certainly offer a nice gain compared to C2D. On the tech side of things, Nehalem is spectacular, but not for home users... probable not enough data to feed.

25th March 2008, 18:09	#10
geoffrey [M] Reviewer Join Date: Nov 2004 Posts: 4,127	To bad they forgot to mention clock rates Still, at this moment dual core is still favorable over Quads, unless software dramatically changes over the following half year I don't think many home users will need it, maybe in 2 years or so... Theoretically, single core apps may be faster, but will you notice the difference compared with the E8x00 which is doing all ready very good? If so, where?