NEWS: AMD's Phil Hester Speaks: Next Generation Architecture, Sort of...

[DilTech]

http://www.anandtech.com/cpuchipsets...oc.aspx?i=2565

I'll let the link speak for itself.

[eBoy0]

*whistles*
DDR4, FBD2, HyperTransport 4.0
and Multi-Core Architecture

[DilTech]

Multi-core is quad-core as far as I know. It'll start with the opterons and move on to the A64's.

You can definitely tell by that chart AMD is looking towards the future. Next year an 2007 will be a very interesting time for pc enthusiasts. It'll be nice to see what conroe is like from intel camp, and what the on-die pci-e controller will do for team AMD....

[Ugly n Grey]

Not quite, in that context muti core is more "multi CPU segment", so a server might get a half dozen specialized coprocessors where a consumer proc gets 1 or 2 etc... the day of the generalized core is coming to an end, I think we'll see one more round of multi (standard) cores per CPU then it's all going to become specialized.... Cell processing led the practical way there in my opinion.

All very interesting though, may both corporations compete hard to get it done....

[nn_step]

So you expect to own a 20 core CPU one day?

[perkam]

So you expect to own a 20 core CPU one day?

Once the no. of CPUs becomes that high, the architecture will be referred to differently...just like pipelines isnt the right way to refer to the pixel and vertex shading abilities to the new gen gpus by ATI and nvidia.

Perkam

[[XC] leviathan18]

if you use 2 cell you almost have 20 cores :P

[nn_step]

need more better cores and quad x1800xts for folding...

[perkam]

We want roadmaps, with firm dates and code names and features, and we've yet to see or hear it. So the best we can report today, is what we told you back at IDF: with the exception of some minor updates as well as the migration to DDR2, the Athlon 64 micro-architecture will remain unchanged throughout 2006.

Wrong...

Hasn't Anandtech heard of the AMD K9?

Perkam

[Ugly n Grey]

Originally Posted by nn_step

So you expect to own a 20 core CPU one day?

I think the term 20 core needs a lot of explaining before I say "yes"

I can see owning a specialized chip that may have four or more standardized cores and a number of specialty cores for different types of coprocessing, yes. In fact, it's inevitable.

[nn_step]

Originally Posted by Ugly n Grey

I think the term 20 core needs a lot of explaining before I say "yes"

I can see owning a specialized chip that may have four or more standardized cores and a number of specialty cores for different types of coprocessing, yes. In fact, it's inevitable.

think of the X2 as of having 2 cores then imagine 10 times as many and being able to buy it for under $500..

[slider99]

Originally Posted by perkam

Wrong...

Hasn't Anandtech heard of the AMD K9?

Perkam

Are you sure there will be a K9? I haven't heard much about it since I read this. Initially, wasn't the K9 supposed to be the dual-core (opterons, for 1207 perhaps), basically "the bastard offspring of K8"?
http://www.theinquirer.net/?article=18967

[onewingedangel]

last I remember k9 was abandoned (at least as a codename), and the successor to dual core k8 (which may now be called k9 if you wish) will be k10.

[turtle]

I imagine it'll go 1-2-4-8, or something like that, as they progress to 65nm, then 45nm and then lower tech, staying with the K8 main architecture, but just adding new ram support (ddr2 in 06, then ddr3 in 07) and minor improvements for core instructions.

The way they mention each core having specific duties (like floating point) is very interesting. The fact IBM and AMD are close buddies (strained silicon for example) makes me believe we'll see cores acting similar to the ways cell does, only designed more towards a all-around computer usage, not only gaming like cell is and X2 will be before anything else. That alone sounds interesting. The fact the cores could divide a workload instead of being redundant (and only really good for multitasking like they are now) really sweetens the pot.

Can you imagine a 45nm AMD (IBM enhanced) chip with 8 cores (programable for seperate instructions at the same time for the same application) running at 4-5ghz a piece, perhaps with 16mb of total cache? I wouldn't be surprised to see it in 5 years.

Think of decoding a movie using all the cores at once (stacked), then as you open another program/game the amount of cores needed to run it efficiently detach from that operation and both move fluidly, managing the loads so they're both efficient. If Vista64 and the software designed for it are programmed to utilize all available cores and assign them specific operations on the fly as they become available/unavailable and distributing loads...This could be extremely cool stuff.

[[XC] leviathan18]

cell is not for gaming.... they are going to be used is supercomputers made by ibm

the future is the cell type processor..... that is for sure i preffer FBDIMM rather than ddr2

[onewingedangel]

a major advantage of fb dimms is that you can have an on die serial memory controller and use either ddr2, gddr3, ddr3 etc. modules on a fb dimm, as long as the memory controller to memory buffer interface is fast enough. So if fbdimms are all they are meant to be, we may never see ddr3 as a standard for memory controllers.

[mtx]

Man, I rather them increase the core speed to 5GHz rather than going to multicores.... I noticed a big jump from going NF2 to NF3, but I doubt I'll notice such a huge jump in performance if I go X2.

[STEvil]

i was saying something about a hardware arbitration device some time back, dunno if anyone remembers it.. and doubt fewer paid attention.

[rcofell]

Originally Posted by onewingedangel

a major advantage of fb dimms is that you can have an on die serial memory controller and use either ddr2, gddr3, ddr3 etc. modules on a fb dimm, as long as the memory controller to memory buffer interface is fast enough. So if fbdimms are all they are meant to be, we may never see ddr3 as a standard for memory controllers.

Well, alot of it depends on costs, from the looks of things servers will be going fbdimms, but desktop systems will stick with normal ddr/ddr2/ddr3, at least for the forseeable future. And since ddr3 is supposed to be electrically compatible with ddr2, it shouldn't take much work to switch over.

If I understand it correctly, the biggest reason for fbdimms is to simplify all the traces to memory, which plays a really big part in servers, since they address so much memory.

[Magnj]

Originally Posted by turtle

I imagine it'll go 1-2-4-8, or something like that, as they progress to 65nm, then 45nm and then lower tech, staying with the K8 main architecture, but just adding new ram support (ddr2 in 06, then ddr3 in 07) and minor improvements for core instructions.

The way they mention each core having specific duties (like floating point) is very interesting. The fact IBM and AMD are close buddies (strained silicon for example) makes me believe we'll see cores acting similar to the ways cell does, only designed more towards a all-around computer usage, not only gaming like cell is and X2 will be before anything else. That alone sounds interesting. The fact the cores could divide a workload instead of being redundant (and only really good for multitasking like they are now) really sweetens the pot.

Can you imagine a 45nm AMD (IBM enhanced) chip with 8 cores (programable for seperate instructions at the same time for the same application) running at 4-5ghz a piece, perhaps with 16mb of total cache? I wouldn't be surprised to see it in 5 years.

Think of decoding a movie using all the cores at once (stacked), then as you open another program/game the amount of cores needed to run it efficiently detach from that operation and both move fluidly, managing the loads so they're both efficient. If Vista64 and the software designed for it are programmed to utilize all available cores and assign them specific operations on the fly as they become available/unavailable and distributing loads...This could be extremely cool stuff.

Yea but everything is relative. I mean a movie at that point will be a 30Gig HD or BlueRay DVD. I mean it will all look and sound and feel better, maybe even smell who the hell knows right but yea... to the future

[nn_step]

and may we not get ripped off in the process..
and nothing like RIMMs please God...

[Mikesta]

Seriously come on guys! AMD is either being very quiet and/or secretive (and sneaky) or they are resting on their laurals.

At a glance Yonah looks to be at least the equal of a manchester cored X2.

I suspect memron/conroe will be significantly better. AMD will barely have realised 65nm Processors when Intel starts pumping out 45nm based cores WITH HEAT LEAKAGE PROBLEM SOLVED!

Even 65nm Pentium D's might be interesting now that they consume around 15-20% less power.

Seriously AMD needs to get off it's ass. I Was expecting that AMD would revise the K8 Architecture. I was hoping it would at least try and shrink L2 Cache Latencies or something.

A K8 that has virtualisation runs DDR2 and is on a 65nm process at the end of 06/early 07 is pretty piss poor.

I seriously expect Conroe/memron to kick the K8's ass, so AMD better have something much better then the type of processor I just mention in my last sentence.

[onewingedangel]

amd are a good 8 - 12 months behind intel on process technology at this point, but that really on target historically.

[Ugly n Grey]

Originally Posted by leviathan18

cell is not for gaming.... they are going to be used is supercomputers made by ibm

the future is the cell type processor..... that is for sure i preffer FBDIMM rather than ddr2

Cell is not for gaming? Come on cell is for eveything, they are sticking it in game consoles and what not. It's a customizeable processor....It's for whatever you want it to be for.

[(sin)morpheus]

Originally Posted by Magnj

Yea but everything is relative. I mean a movie at that point will be a 30Gig HD or BlueRay DVD. I mean it will all look and sound and feel better, maybe even smell who the hell knows right but yea... to the future

Smell'o'vision already failed awhile back, remember? I think it will be awhile before someone picks it up again and tries to make it work.