AMD: the world's first 12-core processor has begun shipping

[SocketMan]

AMD is not targeting a small wprime application with its 6100 series...

I am waiting for people to start complaining about super pie numbers.....
Got to love some of our armchair "computer experts"

[JumpingJack]

let magny IPC = 1, magny frequency = 1, let's say a westmere core with hyperthreading = 1.3 magnycour cores (?), and compare 3.33ghz westmere to 2.2ghz magny

magny performance = 12 (cores) * 1.0 (relative IPC) * 1.0 (relative frequency) = 12
westm performance = _6 (cores) * 1.3 (relative IPC) * 1.5 (relative frequency) = 11.7

?

This is actually, to a first approximation, a nice analysis ... it's hard to tell at this point so many variables, but it looks like to me MC will win some, WSTM will win some, and others will be a tie.

To get a rough (very rough) idea, just take a 4S Istanbul SPEC score, I think WSTM will have an edge in FP, and MC may take a lead in INT.

[JF-AMD]

Westmere will have an edge in FP? I will take any bet from anyone on that.

3 memory channels vs. 4 channels, westmere will not have an edge in FP. Remember also that there will be 12 FPUs in a Magny cours and only 6 in a Westmere.

[mstp2009]

Westmere will have an edge in FP? I will take any bet from anyone on that.

3 memory channels vs. 4 channels, westmere will not have an edge in FP. Remember also that there will be 12 FPUs in a Magny cours and only 6 in a Westmere.

DDR3 vs. DDR2?

4/5 instructions per cycle vs 3?

3.3GHz vs. 2.2GHz?

Mangy Cours might match Westmere in FP, but I would be surprised if it surpasses it. i7 currently spanks Opteron 2400/8400 series in FP workloads. I see AMD just fusing two Istanbul cores together as a stop-gap measure to compete at the high-end until Buldozer cores are available. And even then, Intel has Beckton to counter with (8 cores, 16 threads), due out in 1-2 months.


Oh, and for the record, I run a server farm with mostly Istanbul CPUs, so please don't brand me as an Intel "fanboi".

[JF-AMD]

I think you will be impressed with the FP capabilities of MC. And I am not that concerned with Beckton. Intel is targeting that squarely to be "Itanium III." From what I have seen from a price and power perspective, they are going to be hard pressed to convince people that it is a wise choice for their data centers. The memory buffer design tells me that they didn't learn that much after all of their customers complained about FB DIMMs. And Beckton will only support 1066 memory from what I have read.

[savantu]

I think you will be impressed with the FP capabilities of MC. And I am not that concerned with Beckton. Intel is targeting that squarely to be "Itanium III." From what I have seen from a price and power perspective, they are going to be hard pressed to convince people that it is a wise choice for their data centers. The memory buffer design tells me that they didn't learn that much after all of their customers complained about FB DIMMs. And Beckton will only support 1066 memory from what I have read.

How much memory can you link to a socket for MC ?

For Beckton you can link up to 32 memory DIMMs without loss of speed.

[blindbox]

DDR3 vs. DDR2?

4/5 instructions per cycle vs 3?

3.3GHz vs. 2.2GHz?

Mangy Cours might match Westmere in FP, but I would be surprised if it surpasses it. i7 currently spanks Opteron 2400/8400 series in FP workloads. I see AMD just fusing two Istanbul cores together as a stop-gap measure to compete at the high-end until Buldozer cores are available. And even then, Intel has Beckton to counter with (8 cores, 16 threads), due out in 1-2 months.


Oh, and for the record, I run a server farm with mostly Istanbul CPUs, so please don't brand me as an Intel "fanboi".

DDR3 vs DDR3

http://www.xtremesystems.org/forums/...24#post3995524

I think AMD slides also showed Socket G34 is DDR3.

[ajaidev]

How much memory can you link to a socket for MC ?

For Beckton you can link up to 32 memory DIMMs without loss of speed.

I thing MC's can use around 64GB's of ram in a 8P setup '48core" i could be wrong

[duploxxx]

How much memory can you link to a socket for MC ?

For Beckton you can link up to 32 memory DIMMs without loss of speed.

You can use 3 dimms / MC so you end up with 12 dimms/socket. And if you compare the MC of Nehal against Magny you will see that the Magny is able to sustain a bit faster memory speeds when using 2 dimms/MC and 3 dimms/MC Nehalem will drop to 1066 with 2 and 3 dimms/MC in both SR/DR while Magny will only drop to 1066 when using 3 dimms DR

again Beckton is a different price range.

I thing MC's can use around 64GB's of ram in a 8P setup '48core" i could be wrong

Magny is only up to 4P so that is 48 cores and also 48 dimms so it should give a total of 384GB ram with DR 8GB dimms but those are very expensive......

Opteron 6xxx series pricing -- source

Type Clock L2 cache L3 cache Cons. (ACP - TDP) List price
Opteron 6128 1,5 GHz 8 x 512 kB 2 x 6 MB 75 - 115 W €254
Opteron 6134 1,7 GHz 8 x 512 kB 2 x 6 MB 75 - 115 W €489
Opteron 6136 2,4 GHz 8 x 512 kB 2 x 6 MB 75 - 115 W €692
Opteron 6168 1,9 GHz 12 x 512 kB 2 x 6 MB 75 - 115 W €692
Opteron 6172 2,1 GHz 12 x 512 kB 2 x 6 MB 75 - 115 W €917
Opteron 6174 2,2 GHz 12 x 512 kB 2 x 6 MB 75 - 115 W €1078

8 cores for €254
Even the highest-end 12-core seems more than reasonable at €1078

funny as usual a list from a certain source is going around as the final. Well I can tell you some speed allocations are wrong and are actually based on an old doc

DDR3 vs. DDR2?

4/5 instructions per cycle vs 3?

3.3GHz vs. 2.2GHz?

Mangy Cours might match Westmere in FP, but I would be surprised if it surpasses it. i7 currently spanks Opteron 2400/8400 series in FP workloads. I see AMD just fusing two Istanbul cores together as a stop-gap measure to compete at the high-end until Buldozer cores are available. And even then, Intel has Beckton to counter with (8 cores, 16 threads), due out in 1-2 months.


Oh, and for the record, I run a server farm with mostly Istanbul CPUs, so please don't brand me as an Intel "fanboi".

westmere x5680 is a 130W tdp part so you will have to compare against the 6176SE part so it's 2.3 ghz

Magny match Westmere in FP is not fusing just 2 istanbul it will more then tripple the memory bandwith of an istanbul

[ajaidev]

Semi accurate says these are the prices:

8-core models

* 6124 HE, 1.8GHz, 65W TDP $529
* 6128 HE, 2.0GHz, 65W TDP $599
* 6128, 2.0GHz, 80W TPD, $309
* 6134, 2.3GHz, 80W TDP $599
* 6136 2.4GHz, 80W TDP, $849

12-core models

* 6164 HE, 1.7GHz, 67W TDP, $879
* 6168, 1.9GHz, 80W TDP, $849
* 6172, 2.1GHz, 80W TDP, $1,149
* 6174, 2.2GHz, 80W TDP, $1,349
* 6176 SE, 2.3GHz, 105W TDP, $1,599

http://www.semiaccurate.com/2010/02/...evealed-early/

if these are indeed true i am very very interested in a 6128 2P machine

[ajaidev]

Magny is only up to 4P so that is 48 cores and also 48 dimms so it should give a total of 384GB ram with DR 8GB dimms but those are very expensive......

Ya i meant 4P thats why i said 48 core

and i also meant quad 64 meaning 4*64= 256GB i do remmber reading somewhere that there will be 32 DIMMs so 256Gb will be the max that it supported

[mstp2009]

Magny match Westmere in FP is not fusing just 2 istanbul it will more then tripple the memory bandwith of an istanbul

I do a lot of FP crunching, and most FP workloads are not memory bandwidth constrained. It's just not an issue for over 99% of workloads, and is a non-factor for us crunchers. It's just raw SSE computational power that matters.

[Mech0z]

Do these CPUs use socket F / 1207 ?

[N19h7m4r3]

Do these CPUs use socket F / 1207 ?

They use a new socket, G34. It's much wider than the older one.

[Mech0z]

They use a new socket, G34. It's much wider than the older one.

So there is no motherboards out for it yet?

[N19h7m4r3]

Not that I know of. They have been shown, but I haven't seen them sold yet.

[informal]

In fp there will be no problems for MC Vs Westmere. Both 10h and i7 have same flops theoretically while i7 had an advantage of SMT (gives 0-30%) and DDR3. MC will answer this with much more cores ,2x more,at lower clocks but enough to exceed high clocked Westmere in throughput and with DDR3 support + qudchannel it will address the Westmere's 3-channel DDR3 controller. Simple yet very efficient.

edit: oh and the latest pricing of the MC posted at SA website is wrong too..

[JF-AMD]

How much memory can you link to a socket for MC ?

For Beckton you can link up to 32 memory DIMMs without loss of speed.

12 DIMMs per socket.

I was unaware that there were 32 DIMM per socket configs for Beckton.

The most I have seen are 12 DIMMs per socket and max speed of 1066.

Do you have a link?

[JF-AMD]

Semi accurate says these are the prices:

8-core models

* 6124 HE, 1.8GHz, 65W TDP $529
* 6128 HE, 2.0GHz, 65W TDP $599
* 6128, 2.0GHz, 80W TPD, $309
* 6134, 2.3GHz, 80W TDP $599
* 6136 2.4GHz, 80W TDP, $849

12-core models

* 6164 HE, 1.7GHz, 67W TDP, $879
* 6168, 1.9GHz, 80W TDP, $849
* 6172, 2.1GHz, 80W TDP, $1,149
* 6174, 2.2GHz, 80W TDP, $1,349
* 6176 SE, 2.3GHz, 105W TDP, $1,599

http://www.semiaccurate.com/2010/02/...evealed-early/

if these are indeed true i am very very interested in a 6128 2P machine

That information is not semi accurate, the accuracy rate is lower

[terrace215]

Semi accurate says these are the prices:

8-core models

* 6124 HE, 1.8GHz, 65W TDP $529
* 6128 HE, 2.0GHz, 65W TDP $599
* 6128, 2.0GHz, 80W TPD, $309
* 6134, 2.3GHz, 80W TDP $599
* 6136 2.4GHz, 80W TDP, $849

12-core models

* 6164 HE, 1.7GHz, 67W TDP, $879
* 6168, 1.9GHz, 80W TDP, $849
* 6172, 2.1GHz, 80W TDP, $1,149
* 6174, 2.2GHz, 80W TDP, $1,349
* 6176 SE, 2.3GHz, 105W TDP, $1,599

http://www.semiaccurate.com/2010/02/...evealed-early/

Anyone know if those are actually TDPs, as opposed to AMD "ACP" numbers?

105W appears to be the SE "ACP" wattage for 6-core parts, which is why I ask.

[JF-AMD]

westmere x5680 is a 130W tdp part so you will have to compare against the 6176SE part so it's 2.3 ghz

The X5680 is a workstation only part, typically priced well above reasonable rates and consuming tons of power. We won't see it in the server world.

29% of the 2P market is priced at $3K or below. How do you build a <$3K 2P server with $3200 worth of silicon.

62% of the 2P market is priced from $3-6K. Basically more than 90% of the market price points could never be achieved if you are using those procs.

[Movieman]

12 DIMMs per socket.

I was unaware that there were 32 DIMM per socket configs for Beckton.

The most I have seen are 12 DIMMs per socket and max speed of 1066.

Do you have a link?

The one quad Beckton board I've seen is 8 dimms per socket and correct on the 1066 memory.

The X5680 is a workstation only part, typically priced well above reasonable rates and consuming tons of power. We won't see it in the server world.

29% of the 2P market is priced at $3K or below. How do you build a <$3K 2P server with $3200 worth of silicon.

62% of the 2P market is priced from $3-6K. Basically more than 90% of the market price points could never be achieved if you are using those procs.

I'd disagree with this, power consumption less than previous generation 45nm Gainestown chips.

$3200 for chips
$500 for board
$300 for memory
$4000.00 sub total
That leaves $2000.00 for video,hard drives,case,PSU..Tight yes but doable.

[duploxxx]

The X5680 is a workstation only part, typically priced well above reasonable rates and consuming tons of power. We won't see it in the server world.

29% of the 2P market is priced at $3K or below. How do you build a <$3K 2P server with $3200 worth of silicon.

62% of the 2P market is priced from $3-6K. Basically more than 90% of the market price points could never be achieved if you are using those procs.

hehe don't worry, I wasn't going to compare at all high-end, it was just a remark to the " virtual performance calculation", there are however a few OEM that actually sell those X range in certain rack servers.

the next range of AMD G34 systems that we will ship this year will be based on the 6134 or 6136 will depend a bit on the final asp

[kl0012]

The memory buffer design tells me that they didn't learn that much after all of their customers complained about FB DIMMs. And Beckton will only support 1066 memory from what I have read.

FBDIMM has one mem buffer per slot while Nehalem-ex will have one mem buffer per channel so the power consumption of mem subsystem will be much lower.

The most I have seen are 12 DIMMs per socket and max speed of 1066.

In fact Intel claims up to 16 dimms per socket.
http://download.intel.com/pressroom/pdf/nehalem-ex.pdf

[duploxxx]

FBDIMM has one mem buffer per slot while Nehalem-ex will have one mem buffer per channel so the power consumption of mem subsystem will be much lower.


In fact Intel claims up to 16 dimms per socket.
http://download.intel.com/pressroom/pdf/nehalem-ex.pdf

depends how much that buffer will consume (3xsmall against 1 big) and even more base cost.... yes your memory will be cheaper and less consuming but the base board will be way more expensive and i don't think you will be able to plugin those buffers each time you buy a cpu