AMD K8L - this is not a tweaked K8 ...

SugarCoat said:
Within a month or two. Windsor (dual) and Orleans (single). DDR2 memory support with Pacifica and Presidio technologies. Late 2007 looks to be modified architecture, DDR3 support, dual and quad core line ups some of which will contain L3 cache, I/O virtualization and support for RAID 5 (Opterons). We may see AMD FX parts break 3.4GHz.

In otherwords, nothing new and interesting for quite a while. Dont forget they have yet to break from 90nm which will be the main goal this year, getting production primarily to 65nm.

Expect Intel to have DDR3 and 1333FSB Dual core chips this time next year and quad core on the way as well. There is some speculation that the Extremely Expensive or EE edition of Conroe may be distinguished by a 1333FSB.

Then in 2008 or 2009 Intel will be releasing desktop parts with their own onboard memory controllers.
Click to expand...

It actually appears that in Q4 amd will beginn replacing *some* of their X2 parts with 65nm ones while the rest of the line will be at 90nm.
K8L could very well be a early Q2 release.
AM2 will compete just fine under the time.

And in 08-09 i guess AMD has been sitting on their asses doing nothing again ;)
All guys should check out AMDs analyst day June 1st for more info.
 
pcchen said:
Actually incorporating memory controller is not necessarily a good move in server market. For example, the biggest problem with Opteron is its (lacking of) memory RAS. For smallish servers, memory RAS is not that important. However, it limited the possibility of Opteron for higher-end servers. On the other hand, it's quite possible to provide very nice memory RAS for Xeon platforms, such as IBM's X3.

Of course, for AMD's target market, Opteron does not really need that kind of memory RAS. However, after the initial success of Opteron, AMD needs to put something better into Opteron for higher-end markets, thus the better memory RAS for K8L. Of course it's still not in the same level of X3.
Click to expand...

what's memory RAS?
and won't the FB-DIMM support be nice enough?
 
Blazkowicz_ said:
what's memory RAS?
and won't the FB-DIMM support be nice enough?
Click to expand...
fb-dimm have increased latency, so AMD will wait...

I think (on paper) K8L seems to address many of "narrow" places in its server cpu (like memory mirroring, poisoning etc.)
 
Last edited by a moderator:
nutball said:
:???: Huh? Tukwila isn't a desktop part and never will be (it's an Itanium!). The interesting news about Tukwila is that it sits on CSI, which will also be introduced for Xeons around that time AFAIK. When (if!) CSI will make it to the desktop... is anyone's guess I think.
Click to expand...


never said that chip was a desktop part, and csi is key.
 
chavvdarrr said:
I think (on paper) K8L seems to address many of "narrow" places in its server cpu (like memory mirroring, poisoning etc.)
Click to expand...

Yes, these are among the weakest part of the K8 core on server market.

ECC is the most basic kind of memory RAS. Chipkill is a better kind. For higher-end servers, memory mirroring, data poisoning, hot spare and/or hot swapping are important. IIRC current K8 only supports ECC and Chipkill. On the other hand, IBM X3 supports memory mirroring, hot-swapping, bit steering (for Chipkill), and scrubbing (actively test DRAM for error). Of course, X3 is very expensive, but if you can afford it, you can have it.
 
Memory latency is nothing to all current AMD processors. That's because they have that on-chip memory controller. That allows them to handle all memory requests and the cache fill as optimal as possible. And that is important. That's why Intel benefits from large L2 caches and a very fast memory bus. Because they cannot optimally fill the L2 cache memory. So they have larger latencies they can hide less well, and they have to transport more data from the memory to the L2 cache.
 
You must log in or register to reply here.
Back
Top