- Wednesday, 10 February 2010
- INQ Mobile
- RSS
You are only what you are when no one is looking - Robert C. Edwards
Dual-CPU Nehalem systems pump up the bandwidth
A MONTH FROM NOW, Intel will officially introduce the dual processor Nehalem-EP, aka Gainestown.
Essentially, it keeps the same format and LGA1366 socket as the uni-processor Core i7, just dual CPU support and two QPI links per chip this time - one to talk to the Tylersburg chipset, another for high speed whispering to its twin Nehalem brother.
Even the workstation flavour of the Tylersburg is identical to the desktop one (X58) plus of course an extra QPI link to talk to that second CPU - or the second Tylersburg in dual North Bridge configuration for, say, quad PCIe x16 monster boards. Just count the PCIe lanes on this diagram.
What's so interesting about this machine is not so much the CPU core itself - everyone knows all about the Nehalem microarchitecture now and the core itself is more of an incremental improvement over Core 2 - but probably the most balanced core vs uncore system platform design of any DP machine we can recall.
The 25.6GB/s bidirectional bandwidth on each QPI link is twice the fastest FSB1600 bandwidth previously. And then you add up to 32 GB/s memory bandwidth per processor if simply sticking to the three-channel standard DDR3-1333 on each CPU. Should take care of both workstation and server tasks well, whether cache or memory-dependent. And then we add that scalable I/O bandwidth of one or two Tylersburgs...
Well, we all remember the Skulltrail - it may have been far more of successful if the chipset was not limited by expensive, hot, high latency FB-DIMMs not exactly suited for desktop or even workstation apps, where latency is just important as bandwidth.
In this case, nothing stops the Gainestown, with a memory controller like the Core i7, to support non-ECC high-speed desktop DIMMs on an extreme mobo. And, there will always be applications, desktop, workstation and HPC-alike, that can make good use of the extra bandwidth, not to mention CPU speed uplift possibility.
So, let's hope that we will see some aggressive dual Nehalem boards following the goals of Intel Skulltrail and Asus Z7S WS from the previous generation, but pulling out the stops on an already pretty impressive and scalable platform. µ
Share this:
ShareBusiness Intelligence Software
Customer Relationship Management (CRM)
Enterprise Accounting Software
Enterprise Resource Planning (ERP)
Enterprise Systems Management
32 Pcie Lanes per cpu is close to my calculation for Ultimate. about 40 was my conclusion in January 2005. So Thumper Grows Strong.Yea. 72 is well over mark & this unit should BLAST Away competion to make more such incredibly powerful units, e.g. AMD. Now About AMD, Late Last Night While Working in Lab, thought about AMD new numbers: 4890 or 4970, as reported in theREGISTER. Do You See Numbers: 890 in 4890 & number 970 in 4970? Yes. Its Sign From Got. DaVinci Would Of Coded It Into Cistine Chapel. 890 We know will hit Code Books, yet 970? Whoosssh, More Far Out Possible Players. STeWie Drashek
Very poverful solution ... in theory ... in practice we have ATX-based Core I7 (opening less than 1/2 of it's connection power. With eATX the Nehalem Xeon platform will at most have 4 PCIe x16 and two PCIe x4 slots, that's all!
So what the sence in such connectivity then?
New solution needed!
Perhaps someone should tell the author that the system is dual-socket, not dual core.
wasn't the title enough?
how about
"extra QPI link to talk to that second CPU "
and
"DDR3-1333 on each CPU"
AMD promised but 4x4 platform but failed to deliver it. Seems like Intel is going to do it right this time. AMD has good ideas but poor execution. Intel of late has been using them to transform itself. Soon it will also have Labaree and 32nm chips while AMD does not have real fabs for real men or real ideas.
@Sam
By "do it right this time" you mean tout a platform that most likely requires a $1,000 motherboard to put another $1,000 - $2,000 worth of processors and $500 - $700 of memory.... I wouldn't call it doing it right if 5 - 10 people buy the system despite the performance :)
I have basically been running this system for the last couple of years.... and still do, I love my K8WE.
This thing is basically just a dual socket Opteron system with non-registered memory.
QPI = Hypertransport just 4 years too late.
Core generation+ stuff might be better then amd’s offering but Intel have been following amd for the last 5 years (x86-64 and now Hypertransport)
Previous systems of this type have been severely handicapped by the desire to keep the motherboard functional even with only one CPU populated.
For example, previous multi-socket AMD machines have connected everything but the memory to one CPU. That means all I/O goes to one CPU. Previous Intel multi-socket designs have connected literally everything to one FSB, making the FSB a major bottleneck.
From that block diagram, it looks like that will *not* be the case with this design, and we will be able to use all the PCIe lanes and do I/O to both CPUs.
However, the motherboard will not work properly with just one CPU as some of the PCIe lanes will now be connected to nothing.
There should be very little bottlenecking with this design. Awesome.
@Daryl Quenet
"I wouldn't call it doing it right if 5 - 10 people buy the system despite the performance"
Well, this is not a gamer platform to sbe sold in millions, but there is a need for exellent graphics workstation system - dual Nehalem may be the one (but, afraid, not this year)