2 socket/4 core AMD FX system coming

[Yousuf Khan]

A+E Interactive: AMD Puts Pressure On Intel In PC Gaming With New Socket
For Two Dual-Core Processors
http://blogs.mercurynews.com/aei/2006/06/amd_puts_pressu.html#more

 

[Nate Edel]

A+E Interactive: AMD Puts Pressure On Intel In PC Gaming With New Socket
For Two Dual-Core Processors
http://blogs.mercurynews.com/aei/2006/06/amd_puts_pressu.html#more

So is that basically a quad-core in the Intel/Smithfield model? (Two dies,
two cores each, on a single "chip" in the retail sense of chip.)

Or is that just a double socket, analogous to getting two Opteron 280s
today?

Either way, pretty cool.

 

[Mark]

A+E Interactive: AMD Puts Pressure On Intel In PC Gaming With New Socket
For Two Dual-Core Processors
http://blogs.mercurynews.com/aei/2006/06/amd_puts_pressu.html#more

There was the suggestion that AMD was going to release something to try and
compete with the Conroe at the top end, but if this is it, then it's
disappointing. Most games are only single threaded, or dual threaded at
best. So effectively you will have 2 cores not doing anything.

 

[YKhan]

So is that basically a quad-core in the Intel/Smithfield model? (Two dies,
two cores each, on a single "chip" in the retail sense of chip.)

Or is that just a double socket, analogous to getting two Opteron 280s
today?

Either way, pretty cool.

Although the story I quoted didn't make it clear, AMD had a technology
presentation today, and they explained it. It's a 2-socket system, two
AM2 sockets to be precise. This means that AM2 is multiprocessor
capable, and I bet it's likely that its predecessor Socket 939 was also
multiprocessor capable. It's a hidden feature that was never revealed
till now. Multiprocessors were something only available to Opterons
until now, I figured.

Yousuf Khan

 

[lyon_wonder]

A+E Interactive: AMD Puts Pressure On Intel In PC Gaming With New Socket

For Two Dual-Core Processors
http://blogs.mercurynews.com/aei/2006/06/amd_puts_pressu.html#more

about as desperate as when Intel came up with the Gallatin Xeon-based
P4EE almost 3 years back.

 

[George Macdonald]

about as desperate as when Intel came up with the Gallatin Xeon-based
P4EE almost 3 years back.

Is it as desperate as the Paxville attempt at a dual core Xeon?

 

[GSV Three Minds in a Can]

There was the suggestion that AMD was going to release something to try and
compete with the Conroe at the top end, but if this is it, then it's
disappointing. Most games are only single threaded, or dual threaded at
best.

Fires up 'Oblivion' and looks at thread count .. 12 .. yeah, right
...... damn, these facts surely do spoil a good speculative argument.

Not that I'm claiming they can all execute in parallel, or even that
it's a shining example of multithreading (alt-tab to desktop and it gets
in a real twist, unless you use task manager 'switch to' to get back to
it.)

 

[Evgenij Barsukov]

Fires up 'Oblivion' and looks at thread count .. 12 .. yeah, right .....
damn, these facts surely do spoil a good speculative argument.

Not that I'm claiming they can all execute in parallel, or even that
it's a shining example of multithreading (alt-tab to desktop and it gets
in a real twist, unless you use task manager 'switch to' to get back to
it.)

I wonder if anybody has found a review on how the games in average are
getting into parallelizing the code. That should be a substantial trend
now, considering that 2x became a mainstream for high-end systems.

Basically such review would not even require access to the code -
somebody would run a set of different newly released games on single CPU
vs multiple (2x, 4x) and see how performance difference is.
Looking on number of threads could also be instructive.

Regards,
Yevgen

 

[Evgenij Barsukov]

Fires up 'Oblivion' and looks at thread count .. 12 .. yeah, right .....
damn, these facts surely do spoil a good speculative argument.

Not that I'm claiming they can all execute in parallel, or even that
it's a shining example of multithreading (alt-tab to desktop and it gets
in a real twist, unless you use task manager 'switch to' to get back to
it.)

There is an interesting article discussing multi-processor in games
with design architect of Unreal:
http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2377&p=3

It is somewhat dated, but prediction he gives is that by second half of
these year most game-making teams will be seriously working on
parallelizing critical portions of their code.

Regards,
Evgenij

 

[Nate Edel]

Although the story I quoted didn't make it clear, AMD had a technology
presentation today, and they explained it. It's a 2-socket system, two
AM2 sockets to be precise. This means that AM2 is multiprocessor
capable, and I bet it's likely that its predecessor Socket 939 was also
multiprocessor capable. It's a hidden feature that was never revealed
till now. Multiprocessors were something only available to Opterons
until now, I figured.

Sweet.

I was under the impression the 939 was identical to 940 except for not
requiring registered memory, and would have supported multiprocessors had
the other things needed been there (not to minimize the importance of the
lack of a 2/800-series opteron for it or any 2-socket motherboards.)

 

[The little lost angel]

I wonder if anybody has found a review on how the games in average are
getting into parallelizing the code. That should be a substantial trend
now, considering that 2x became a mainstream for high-end systems.

Basically such review would not even require access to the code -
somebody would run a set of different newly released games on single CPU
vs multiple (2x, 4x) and see how performance difference is.
Looking on number of threads could also be instructive.

Xbit lab did a review of Opteron vs Xeon and had one benchmark on
Quake 4 (that's relatively modern no?) showing 25% increase from
multiple cores with SMP support enabled but only on dual cores SMP
strangely enough and not from a Single Core SMP.

http://www.xbitlabs.com/articles/cpu/display/opteron-xeon-workstation_15.html

 

[nobody]

about as desperate as when Intel came up with the Gallatin Xeon-based
P4EE almost 3 years back.

Why desperate? Dual socket was a must-have in workstations and
servers for years. AMD should be commended for bringing it into
mainstream first rime in years since SMP using mainstream P3 chips
became history. The midrange workstation and entry level server crowd
will jump on it, as well as power users that _do_ multitask and always
have a lot of stuff running in the background. These applications
don't need registered RAM, and don't require ECC, though could use it,
especially servers.

Unlike Xeon-based P4EE, these chips would not need extra cache or
anything comparing to mainstream A64, just enabling existing coherent
HT - or it was enabled all the time, just not announced until now? So
what's desperate? Revealing that A64 and Opteron are cut from the same
wafers? This "secret" was known since K8 came out. And while
Gallatins with all the extra cache and insane price had hard time
competing against even midrange A64, these duallies will leave in the
dust single socket Conroe, and will be competitive against dual
Woodcrest, maybe even more so than dual Opterons because of faster
non-reg/ECC memory.

The only unknown part of it is pricing. If dual socket A64 board will
offer significant savings compared to Opteron board, and also any A64
will work in it, not just special SMP-enabled, higher-priced variety,
this setup will be a hit across multiple sectors of users. What the
heck, maybe even SMP on a budget will be an option - remember Celeron
and Duron-based duallies? I bet the setup consisting of the cheapest
possible dual socket board (around $100, anyone?), a couple of
Semprons, and 2 RAM sticks connected to only one socket out of 2 might
become popular as a poor man's SMP. OTOH, if the price of such a
system will be comparable to dual Opteron, there might not be many
takers.

NNN

 

[Tony Hill]

Sweet.

I was under the impression the 939 was identical to 940 except for not
requiring registered memory, and would have supported multiprocessors had
the other things needed been there (not to minimize the importance of the
lack of a 2/800-series opteron for it or any 2-socket motherboards.)

In the case of Socket 939 it was actually quite significantly
different than Socket 940, and it most likely was not multiprocessor
capable. I looked at the pin-outs before and they didn't have the
second hypertransport pin-out marked, and that would be needed for
even just a 2-processor system. There aren't even enough "no
connection" pins that could be used for this.

FWIW you can find the comparison between Socket 939 and Socket 940 in
the following two documents:

http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/31411.pdf

http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/31412.pdf


Of particular interest for this discussion is the fact that Socket 940
has L0, L1 and L2 pins, while Socket 939 only has L0 pins. On the
flip side, 939 has a couple extra "no connection" pins, some
(seemingly useless?) "strap" pins and a bunch of extra power and
ground pins.

As for socket AM2, I don't think the docs have been published yet, so
I can't say one way or the other. Certainly the possibility for a
second Hypertransport link does exist though.

 

[Yousuf Khan]

There was the suggestion that AMD was going to release something to try and
compete with the Conroe at the top end, but if this is it, then it's
disappointing. Most games are only single threaded, or dual threaded at
best. So effectively you will have 2 cores not doing anything.

Why would it be disappointing? This looks like it's a slightly modified
2-way Opteron system (modified because it won't require server-grade
registered RAM), that's gotten so popular in the server world. They've
obviously been hiding things from us, it looks like the Athlons have had
multiprocessor capabilities all along, just like the Opterons. Each
Athlon is connected to each other via Hypertransport and each one has
its own bank of memory. They didn't cheap out and make all of the memory
available through one Athlon or the other, both will have their own
banks, which means their aggregate memory bandwidth will be doubled just
with the addition of a second processor.

Yousuf Khan

 

[Yousuf Khan]

Sweet.

I was under the impression the 939 was identical to 940 except for not
requiring registered memory, and would have supported multiprocessors had
the other things needed been there (not to minimize the importance of the
lack of a 2/800-series opteron for it or any 2-socket motherboards.)

I personally never suspected that the only difference between Socket 939
and 940 was simply the removal of Registered RAM pin. I thought it was a
much more thorough working over of the pins and their functions. But of
course now in retrospect, what else could it have been?

Yousuf Khan

 

[Yousuf Khan]

In the case of Socket 939 it was actually quite significantly
different than Socket 940, and it most likely was not multiprocessor
capable. I looked at the pin-outs before and they didn't have the
second hypertransport pin-out marked, and that would be needed for
even just a 2-processor system. There aren't even enough "no
connection" pins that could be used for this.

However, were there a bunch of "reserved" or "not defined" pins
available? Those are usually where hidden functions reside.

Yousuf Khan

 

[Tony Hill]

However, were there a bunch of "reserved" or "not defined" pins
available? Those are usually where hidden functions reside.

Nope, they didn't have anything listed as "reserved" or "not defined",
just the above-mentioned "no connection" pins, and there wasn't enough
of them for another 16-bit/16-bit (in/out) hypertransport channel
(which requires about 70 pins total). I suppose there were enough for
a 8-bit/8-bit or narrower hypertransport channel, and they might have
had just enough to squeak in a 16-bit/8-bit channel, but not the
full-fledge connection as in the Opteron chips.

I think there are something like 55 unused pins in Socket 939. The
bulk of the old Hypertransport pins from Lane 1 and Lane 2 in Socket
940 were converted to be power and grounding pins in Socket 939.
Actually I suppose they weren't technically "converted", since the two
chips have totally different pin-outs, but that's where the difference
in pins went.

 

[Yousuf Khan]

I think there are something like 55 unused pins in Socket 939. The
bulk of the old Hypertransport pins from Lane 1 and Lane 2 in Socket
940 were converted to be power and grounding pins in Socket 939.
Actually I suppose they weren't technically "converted", since the two
chips have totally different pin-outs, but that's where the difference
in pins went.

I wonder if power and ground are now the new "reserved" pins? You don't
want it used, and while you're not using it, you want it to be kept in
logical ON state, so you call it power. Similarly, other pins which you
want to remain in the logical OFF state, you call those ground. Makes
you wonder.

AMD is introducing Socket S2 or something for Turion laptops. It's 640
pins, so they've blown off hundreds of pins. And apparently it's still
capable of dual-channel DDR2 operation, despite the fact that it's got
less pins than the older AMD Socket 754 DDR single-channel-only
connector. Hell it's got 300 pins less than Socket AM2, which is also
dual-channel DDR2. Again makes you wonder.

Yousuf Khan

 

[Nate Edel]

AMD is introducing Socket S2 or something for Turion laptops. It's 640
pins, so they've blown off hundreds of pins. And apparently it's still
capable of dual-channel DDR2 operation, despite the fact that it's got
less pins than the older AMD Socket 754 DDR single-channel-only
connector. Hell it's got 300 pins less than Socket AM2, which is also
dual-channel DDR2. Again makes you wonder.

Possibly. Could they also have had higher expected power consumptions when
they designed the original socket 754 and 939 designs? It takes a lot of
power in/out pins to get modern desktop power consumptions in at low volts
and amps per pin.

 

[Mark]

Why would it be disappointing? This looks like it's a slightly modified
2-way Opteron system (modified because it won't require server-grade
registered RAM), that's gotten so popular in the server world. They've
obviously been hiding things from us, it looks like the Athlons have had
multiprocessor capabilities all along, just like the Opterons. Each Athlon
is connected to each other via Hypertransport and each one has its own
bank of memory. They didn't cheap out and make all of the memory available
through one Athlon or the other, both will have their own banks, which
means their aggregate memory bandwidth will be doubled just with the
addition of a second processor.

Yousuf Khan

It is disappointing as the Conroe is good for the home market, so how does
releasing a server chip help?