ZAMORA: QUAD CORE WITH SHARED L3 CACHE

[9-inch]

PC Watch has an updated roadmap of upcoming AMD processor including a quad-core 65nm processor with shared cache:
http://pc.watch.impress.co.jp/docs/2006/0503/kaigai267.htm?
You'll need a translator[/i]

 

[Action_Man]

 

[spud]

Word.

 

[hergieburbur]

Just out of curiousity, AM, do you watch for 9-Inch to post or what? You are always like five minutes or so behind him.

 

[Action_Man]

Sort of. If I see one I'll post the keyboard and slinky.

 

[ltcommander_data]

It may be my point of view, but I found that roadmap to be extremely disappointing.

First of all, 65nm processors look to be coming out late Q1 or optimistically early Q1 which is a far cry from the H2 2006 estimates people were pushing about.

Second, despite all the talk about K8L, it's difficult to see where that architecture will appear. All the processors to be launched 2007 are still labeled as K8. There are some improvements such as improved order fetch and decoders which was something mentioned in the Anandtech architecture review as needing to be improved. The later will probably be expanded use of their macro-ops fusion (which is micro-ops fusion in Intel terms) rather than additional decoders. Additional FPUs isn't mentioned at all and without them there's not really much need for more decoders. The other new feature is clock gating for thermals. The 2007 processors look like K8L light rather than K8L. The only interesting thing is the shared L2 cache in the quad core Deerhound. The thing is that it looks like AMD's true and first quad core won't arrive until H2 2007, which may well be 2 quarters behind Cloverton and at the same time as Intel's 45nm true quad core.

The roadmap also doesn't make any mention of K10, even in 2008. In fact, it's only in 2008 that we see the K8 "new core" which seems like when the real K8L will be released. What I found funny is that even though Hypertransport3 is now finalized, AMD won't introduce it in 2007 and will wait for K8L in 2008.

Overall, it really seems AMD is taking their time with things. If this roadmap is to be believed, and since you post it, you must have some faith in it, AMD doesn't have K10 on the horizon and they won't have a new architecture with K8L until 2008. Their quad cores won't arrive until H2 2007. Very interesting indeed.

 

[4745454b]

"like AMD's true and first quad core won't arrive until H2 2007, which may well be 2 quarters behind Cloverton and at the same time as Intel's 45nm true quad core."

So? Does AMD need to keep up with Intel in processes and cores? Even though Intel switched to "faster" DDR2, AMD kept up wonderfully with "old" DDR. Just because your rival moves to something else doesn't mean you have to also. For the record, I am an AMD "fanboy" but I know that Conroe is going to put Intel back in the front again. Maybe K8* or K10 will put AMD back, maybe not. But I don't see the point in worrying about any of these things. Computing, like woodworking, requires you to use the right tool for the task. Intel will be better for some things, AMD for others. Its called the way of life. Maybe I'm not a fanboy after all...

 

[K8MAN]

This is the newest available? I see K7 on there FFS......

 

[Ycon]

I dunno why but I do have a feeling that Intel is already using shared cache and FB-DIMM
Doesnt AMD know what to do now?

 

[mitch074]

Intel: shared L2 cache on Conroe, not P4 (the CPU are on the same die, but not connected)
FB-DIMM: requires dedicated circuitry on the north bridge; the integrated memory controller in AMD's K8 does a bit of the same thing if I'm not mistaken.

 

[Topota_madre]

Shared L2 cache: Know in Core Duo
Fb-Dimm available to Xeon. New chipset. More bandwich than the fatest AMD controler and good in server uses when latency is not a problem. AMD IMC works well when latency is the issue.

 

[mitch074]

Good luck fitting a Core Duo on a desktop (possible, but hard to find the mobo). One can indeed wonder why AMD didn't put the crossbar system before the L2 cache instead of after? Maybe it would create so much cache corruption or cache misses that they considered it not worth it?
Anyway, considering the way the K8 works and the usual L2 cache size on the K8 chips (ie. small) and how good performers they are, a shared L2 cache may have been more trouble than it was worth. On the P4 architecture though, it would have been quite a boon.

I'm not sure about bandwidth: is FB-DIMM faster than the triple 2-ways HT links on Opterons 4xx and 8xx?

 

[iterations]

Core Duo on the desktop is getting more and more popular. The Apple iMacs use them now, and for PCs this review was just published the morning:

http://www.anandtech.com/mb/showdoc.aspx?i=2750

Front page review on Anandtech. Once it is clocked properly, the Core Duo appears to beat the FX-60 in pretty much every case which is pretty amazing when you consider it has a 35W TDP and was originally intended for the mobile space. The reviewer claims it is rock solid, everyday usable at a 2.8GHz overclock with standard air cooling. They also say it took like no effort to do the overclock, they just changed the settings in the BIOS and it was done. Nice.

 

[gOJDO]

I'm not sure about bandwidth: is FB-DIMM faster than the triple 2-ways HT links on Opterons 4xx and 8xx?

Not only faster, but much more easier to implement on standar 4 or 6 layer motherboards.
FB-DIMM is connected through 69 pins, while DDR2 is through 244.
You can add more channels and more modules on every single FB-DIMM channel, therefore enabling more available RAM.

 

[BaronMatrix]

It may be my point of view, but I found that roadmap to be extremely disappointing.

First of all, 65nm processors look to be coming out late Q1 or optimistically early Q1 which is a far cry from the H2 2006 estimates people were pushing about.

Second, despite all the talk about K8L, it's difficult to see where that architecture will appear. All the processors to be launched 2007 are still labeled as K8. There are some improvements such as improved order fetch and decoders which was something mentioned in the Anandtech architecture review as needing to be improved. The later will probably be expanded use of their macro-ops fusion (which is micro-ops fusion in Intel terms) rather than additional decoders. Additional FPUs isn't mentioned at all and without them there's not really much need for more decoders. The other new feature is clock gating for thermals. The 2007 processors look like K8L light rather than K8L. The only interesting thing is the shared L2 cache in the quad core Deerhound. The thing is that it looks like AMD's true and first quad core won't arrive until H2 2007, which may well be 2 quarters behind Cloverton and at the same time as Intel's 45nm true quad core.

The roadmap also doesn't make any mention of K10, even in 2008. In fact, it's only in 2008 that we see the K8 "new core" which seems like when the real K8L will be released. What I found funny is that even though Hypertransport3 is now finalized, AMD won't introduce it in 2007 and will wait for K8L in 2008.

Overall, it really seems AMD is taking their time with things. If this roadmap is to be believed, and since you post it, you must have some faith in it, AMD doesn't have K10 on the horizon and they won't have a new architecture with K8L until 2008. Their quad cores won't arrive until H2 2007. Very interesting indeed.

Sounds like you don't like AMD very much. It's kind of funny how whenever someone says something not totally positive about Intel you jump on them. AMD is going to do what they want. Though it would be funny if they pulled a fast one and widened the K8 to twice the registers and order buffer. 6 IPC would be devastating.

 

[9-inch]

I'm not sure about bandwidth: is FB-DIMM faster than the triple 2-ways HT links on Opterons 4xx and 8xx?

Not only faster, but much more easier to implement on standar 4 or 6 layer motherboards.
FB-DIMM is connected through 69 pins, while DDR2 is through 244.
You can add more channels and more modules on every single FB-DIMM channel, therefore enabling more available RAM.

This is false. FB-DIMM is not faster than current DDR2. The only good thing about FB-DIMM is the possibility to allow a motherboard to have more RAM installed (> 192GB). If AMD really wanted a cooler server environment, they should trash FB-DIMM since it's well known to be a power hog.

 

[gOJDO]

I'm not sure about bandwidth: is FB-DIMM faster than the triple 2-ways HT links on Opterons 4xx and 8xx?

Not only faster, but much more easier to implement on standar 4 or 6 layer motherboards.
FB-DIMM is connected through 69 pins, while DDR2 is through 244.
You can add more channels and more modules on every single FB-DIMM channel, therefore enabling more available RAM.

This is false. FB-DIMM is not faster than current DDR2. The only good thing about FB-DIMM is the possibility to allow a motherboard to have more RAM installed (> 192GB). If AMD really wanted a cooler server environment, they should trash FB-DIMM since it's well known to be a power hog.
Read:
http://www.realworldtech.com/page.cfm?ArticleID=RWT110805135916&p=4

 

[TechnologyCoordinator]

A fascinating find, thanks 9-Inch. PS - Next time use "Title Case" for your thread titles instead of "ALL CAPITAL LETTERS". Imagine skipping the keyboards and slinkies and getting right down to intelligent discussion.

 

[gOJDO]

8O What do you mean....Intelegent disscusion with 9-inch?!? :roll:

 

[Ycon]

9-inch is not after an intelligent discussion...

 

[BaronMatrix]

I'm not sure about bandwidth: is FB-DIMM faster than the triple 2-ways HT links on Opterons 4xx and 8xx?

Not only faster, but much more easier to implement on standar 4 or 6 layer motherboards.
FB-DIMM is connected through 69 pins, while DDR2 is through 244.
You can add more channels and more modules on every single FB-DIMM channel, therefore enabling more available RAM.

This is false. FB-DIMM is not faster than current DDR2. The only good thing about FB-DIMM is the possibility to allow a motherboard to have more RAM installed (> 192GB). If AMD really wanted a cooler server environment, they should trash FB-DIMM since it's well known to be a power hog.


They ARE DDR2, just with a buffer mechanism that communicates with the bus. It allows higher density this way since the buffer is in charge of locations.

 

[SexBomb]

I'm not sure about bandwidth: is FB-DIMM faster than the triple 2-ways HT links on Opterons 4xx and 8xx?

Not only faster, but much more easier to implement on standar 4 or 6 layer motherboards.
FB-DIMM is connected through 69 pins, while DDR2 is through 244.
You can add more channels and more modules on every single FB-DIMM channel, therefore enabling more available RAM.

This is false. FB-DIMM is not faster than current DDR2. The only good thing about FB-DIMM is the possibility to allow a motherboard to have more RAM installed (> 192GB). If AMD really wanted a cooler server environment, they should trash FB-DIMM since it's well known to be a power hog.


They ARE DDR2, just with a buffer mechanism that communicates with the bus. It allows higher density this way since the buffer is in charge of locations.

You've said it all.
After all, 9-Inch wasn't all that wrong since DDR2 isn't faster than DDR2. 8)

 

[nottheking]

Well, I must say, this is perhaps more information we've seen on AMD than we have in years. It does say something rather interesting, though; they're clearly intent on doing something this summer/fall, in sync for Intel's Core onslaught. However, I must say I don't see all that many differences between that and their current lineup. Sadly, the roadmap provides no information on fabrication processes, so we're left to guess when 65nm finally comes in; it's quite possible that the entire later 2006 lineup is all 90nm. Sure, it's served AMD very, very well for the greater part of two years, but I think that some people are starting to grow tired of it.

So this is what AMD has been planning huh?

AMD goes to war against Intel! :lol:

So THAT would be where AMD has been putting all those profits from their 90nm chips. You'd have otherwise thought that, were they actually re-investing it into R&D, they'd have given up 45nm chips, right on the market, by now. But naw, an aircraft carrier's SO much cooler!

 

[hergieburbur]

A fascinating find, thanks 9-Inch. PS - Next time use "Title Case" for your thread titles instead of "ALL CAPITAL LETTERS". Imagine skipping the keyboards and slinkies and getting right down to intelligent discussion.

but the keyboard and slinky are soooo much fun

 

[rettihSlluB]

Well, I must say, this is perhaps more information we've seen on AMD than we have in years. It does say something rather interesting, though; they're clearly intent on doing something this summer/fall, in sync for Intel's Core onslaught. However, I must say I don't see all that many differences between that and their current lineup. Sadly, the roadmap provides no information on fabrication processes, so we're left to guess when 65nm finally comes in; it's quite possible that the entire later 2006 lineup is all 90nm. Sure, it's served AMD very, very well for the greater part of two years, but I think that some people are starting to grow tired of it.
So this is what AMD has been planning huh?

AMD goes to war against Intel! :lol:

So THAT would be where AMD has been putting all those profits from their 90nm chips. You'd have otherwise thought that, were they actually re-investing it into R&D, they'd have given up 45nm chips, right on the market, by now. But naw, an aircraft carrier's SO much cooler!

I hope this one can clear your doubts:
http://www.hkepc.com/bbs/news.php?tid=593736
Still, Greyhound, Cadiz and Zamora cores aren't listed as 65nm processors. It could be that AMD may release these beasts in 45nm process. It does makes sense since they said that 45nm would be available in 2008.