AMD's "Trick" for Summer 2006

[KimB]

Now that I think about it, I think this guy is just confused about the next high-pin socket that AMD's using for their Opterons, the socket that's going to pave the way for their quad-core CPU's.

 

[Geo]

With a marketing name of "4x4" attached to it, Chal? This sniffs like the real deal to me (which is entirely separate from whether it is a good idea or not. . .)

 

[John Reynolds]

It's the real deal. And I've spoken with their lead PR person a few weeks ago, so it's not something I've misinterpreted via 'the' grapvine.

 

[LeStoffer]

FX-62 and higher...

Okay, it is officially announced:

http://www.dailytech.com/article.aspx?newsid=2638

And yes, it will be an expensive experience indeed:

"The new motherboards will be AM2 based, and take new AM2 FX-62 and higher.."

 

[KimB]

Wierd, but I'm still not hearing anything about it that is anything more than your standard multiprocessor system, dropped down to the consumer level.

I find especially odd, for example, this statement from that article:

AMD says that 4x4 is not simply workstation and server technology reduced down to consumer level, but is in fact, a whole new design. The company indicated that the new 4x4 chipset will be able to allow each processor core to have direct access to memory, and to have direct access to each other. The technology is called Direct Connect, and according to AMD, will be able to give huge boost to performance over traditional multi-socket platforms.

But the description that follows "a whole new design" sounds just like current Opteron multiprocessor designs.

 

[nutball]

Could be marketing semantics. "Traditional multi-socket platforms" = Intel's way of doing things. It's hard to see why, if this tech really if the bee knees, it's introduced on a gaming platform first, not in servers and workstations where it actually might make a meaningful difference.

 

[Karma Police]

But isn't it really up do the chipset designers to make a MB that supports two AM2 processors? How long will it take to have nVidia/ATi/Via come out with one after AMD announces they support two AM2 chips?

 

[KimB]

But isn't it really up do the chipset designers to make a MB that supports two AM2 processors? How long will it take to have nVidia/ATi/Via come out with one after AMD announces they support two AM2 chips?

I'm not entirely certain that it requires much of a change in the chipset.

 

[Blazkowicz]

But isn't it really up do the chipset designers to make a MB that supports two AM2 processors?

no, it's actually up to AMD to have a second, NUMA capable HTT link on the CPU. I believe that chipset can stay the same.

 

[Gubbi]

Could be marketing semantics. "Traditional multi-socket platforms" = Intel's way of doing things. It's hard to see why, if this tech really if the bee knees, it's introduced on a gaming platform first, not in servers and workstations where it actually might make a meaningful difference.

I agree, 4 cpus in one socket makes the most sense in workstations.

In servers you probably want the full bandwidth from two sockets (and board size is only a problem in blades).

Cheers

 

[Blazkowicz]

Yeah, but this will to some extent negate the advantage that Opteron has had over Xeon up to now, ie. multiple independent memory busses, one per socket. The secondary pair of cores would be very much second-class citizens, not having their own memory bus. Four cores sharing a single memory bus could present some real scalability issues if your problem doesn't fit in L2 cache.

there are those dual socket 940 MB where only one CPU access its own RAM, you know. I don't know if they are still made, they were a cheap way to run two single core opteron 2xx at the time. Someone with too much time on its hand could run two opteron 265 on such a mobo and compare with a mobo where each CPU has its RAM. While keeping in mind that on AM2 socket you can have more memory bandwith (non registered DDR2 versus registered DDR)

this might actually be AMD wanting a cheap, AM2 based opteron platform (using "normal" RAM being a big part of being cheap). See how they now make socket 939 1xx opteron (no more single socket 940), and socket F looks pretty high end (lots of pins, up to five HTT links, quad channel FB-DIMM). Socket F opteron would be put against Sparc, Itanium ; and dual AM2 socket would go against the Xeon. (with some overlap maybe as someone might want registered RAM, and thus go with a single socket F)

what do you think of that little theory? ^

 

[Rolf N]

I just don't get it. This is the same AM2 socket, right?
So they're basically announcing that AM2 has enough pins for the second (CPU<=>CPU) Hypertransport link, and they expect us to be in awe. Hmm

 

[KimB]

Yeah, I thought socket 939 had that already...

 

[chavvdarrr]

oh, they do use 2 sockets, instead of glueing ... DDR2-800 has 2x the bandwidth of current DDR400 for s939, so basically 4 cores on DDR2 will have same bandwidth available as in current s939 2S systems - with some latency penalty... but how they can call this "new" ?!

 

[KimB]

If they're using two sockets, then each socket will have access to its own dual-channel memory bus. In the future, with quad-core, yes, DDR2 will likely make an impact, for the reason you describe. But this sounds rather different.

 

[AlNom]

Techreport on 4x4...

4x4 will involve dual-socket, enthusiast-class motherboards and dual-core Athlon 64 FX processors

 

[Rur0ni]

Techreport on 4x4...

Yes it's for the enthusiast populace.

 

[chavvdarrr]

If they're using two sockets, then each socket will have access to its own dual-channel memory bus. In the future, with quad-core, yes, DDR2 will likely make an impact, for the reason you describe. But this sounds rather different.

I mean that 2S + 1 common dual-channel memory bus will be more than enough - it will give same BW per core as current s939 2S motherboards. Unless CPUs are clocked much higher (btw anyone noticed that AMD claims 65nm CPus on the way? "in production from Chartered Semi" - so less than 3 months before we see them), there is no need for more BW, and this will make motherboards a bit cheaper

 

[KimB]

I mean that 2S + 1 common dual-channel memory bus will be more than enough - it will give same BW per core as current s939 2S motherboards. Unless CPUs are clocked much higher (btw anyone noticed that AMD claims 65nm CPus on the way? "in production from Chartered Semi" - so less than 3 months before we see them), there is no need for more BW, and this will make motherboards a bit cheaper

Since it's the same socket, each CPU will have its own dual-channel memory controller, and each socket will have the connections for a dual-channel memory connection. So I would consider it very likely that we'd be talking a dual-channel connection to each individual CPU.

 

[arjan de lumens]

Yeah, I thought socket 939 had that already...

Socket 939 never reserved any pins for more than just one HT link, although it easily has enough pins that it would have been possible for AMD to support an additional link if they wanted to (the older socket 940 did provide both 3 full HT links and a 128-bit ddr interface).