Pulled to a new post for clarity.
Supercomputers, and Apple's forthcoming super computers
Guess i'll have to use "Powerful to the core" another time...
SI units, just to be clear:
1
Tera(e.g. 1 Tera
flop (
FLoating point
Operations
Per
Second)
1 million = 1 mega = 1 thousand thousands– 1,000,000 i.e. 1,000 to the power 2
1 billion = 1 giga = 1 thousand million – 1,000,000, 000 i.e. 1,000 to the power 3
1 trillion =1 tera = 1 million million - 1,000,000 ,000,000 i.e. 1,000 to the power 4
1 Quadrillion = 1 peta = 1 thousand trillion – 1,000,000 ,000,000 ,000 i.e. 1000 to power 5
#1 of the Top 500 Supercomputers
Now: Blue Gene/L – capable of just under 500 trillion operations – 0.478 Petaflops (478 teraflops).
Very soon: IBM's new Roadrunner supercomputer which runs at >1.0 petaflops (>1,000 teraflops)
Roadrunner's power is equivalent to the PS3 Folding@home project world record (1 petaflop)
With supercomputers being assembled through standard racks of smaller servers, with high speed connections, why couldn't smaller supercomputers use Snow Leopard and off the shelf equipment (SAN, XServe, Snow Leopard Server, etc)? If you look at the extrapolation graph attached of predicted power of these supercomputers it's a straight line on a log graph. And they’ve been keeping pretty much on a straight line for a bit.
So what might be around the corner?
Intel
Intel's Tera Scale Computing’s 80-core silicon can do 1 Teraflop - at 1 teraflop, it’s at 3.13Ghz, and using only 24 Watts!
So 1,000 Tera Scale chips = 1.000 teraflops (i.e. 1 petaflop) =1x Roadrunner supercomputer
So, how many Nehalem Snow Leopard Macs does it take to make a Petaflop? Like eckers I have the stats to hand. Feel free to post what the flop potential is of a dual socket Nehalem Snow Leopard rig!
Where does the 80 core fit into Intel?
Intel's Tera scale Computing Research Program
(TsCRP) is a worldwide effort. They're looking to scale multi-core architecture to 10s-100s of cores, embracing a shift to parallel computing.
It includes the 80 core processor department, and Universal Parallel Computing Research Centers (UPCRCs)
The hardware research vision:
- Scalable multi-core architectures that integrate streamlined processor cores and accelerators with fast, energy-efficient, modular core-to-core infrastructure e.g. 80-core prototype processor, Tera-scale Emulator, Dynamic Thermal Management, Task Queues.
- Memory sharing & stacking - e.g. 3D Stacking, Cache Quality of Service.
- High Bandwidth I/O & communications balancing the compute demands with I/O & network demands within the platform power and cost budgets e.g. High-speed Copper I/O, Silicon Photonics, I/O Accelerators.
The software research vision:
- Model-based applications e.g. Visual Computing that'll use use tera-scale capabilities More i.e. Ray Tracing, Physical Modeling, Media Mining, Enhancement
- Parallel programming tools e.g. Transactional Memory, Accelerator Exoskeleton, Ct Data-Parallel Programming
- Thread-aware execution environments that provide real-time performance & power management across cores and scale with increasing thread and core counts. e.g. Many-core Run Time, Tera-scale Virtual Machine
Intel has created 2 Universal Parallel Computing Research Centers (UPCRC), through working with Microsoft, University of California (UC) Berkeley & the University of Illinois (UIUC). They're there to look into new ways to program software for multi-core processors. Parallel computing brings together advanced software and processors that have multiple cores or engines, which when combined can handle multiple instructions and tasks simultaneously, but you knew that already
)
They are going to
- tackle the challenges of parallel computing (e.g. programming for processors with multiple cores), thus being able to carry out multiple sets of program instructions at a time
- accelerate accelerating developments in mainstream parallel computing, for consumers & businesses in desktop
& mobile computing.
- advancing parallel programming applications, architecture and OS software.
Microsoft & Intel stumped up $20 million over 5 years, $8 million put up UIUC, & UC Berkeley is getting together $7 million. (As many as 20 universities - including MIT, Stanford competed for the funding)
Lots of
news about the place.
EETimes,
Microsoft,
NYTimes,
Wiredand so on.
The industry is in a little bit of a panic about how to program multi-core processors, especially heterogeneous ones... To make effective use of multi-core hardware today you need a PhD in computer science. That can't continue if we want to enable heterogeneous CPUs
Chuck Moore, a senior fellow at AMD
(Sidenote: "The academic community has never really recovered from Darpa’s withdrawal" Daniel A. Reed, director of scalable and multicore computing at Microsoft, who will help oversee the new research labs - NY Times
Both Intel and Microsoft executives said the research funds were a partial step toward filling a void left by the Pentagon’s Defense Advanced Research Projects Agency, or Darpa. The agency has increasingly focused during the Bush administration on military and other classified projects, and pure research funds for computing at universities have declined. )
Image of Tera chips from
here others from Chipzilla and the Supercomputer place...
Berkeley's project is looking at the >16 core market. So that's stuff for 2010 - not that far off considering! Uni's that didn't get the funding are still looking at the problem. So how the **** is Apple ahead on this? They didn't boast per se, they just hinted that they had everything under control, not to worry, and enjoy Nehalem. Win-Tel is pouring $2 million a year into this. You've got huge academic uni's like MIT and Stanford on the case - how could Apple be ahead - and from the noises Apple made - ahead by a fair bit? (Patent's rear their head again - if Apple is ahead like it is with the iPhone in some areas, boy is that going to be potential fun for them)
Power to the people - distributed processing
With an aim to build China's most powerful supercomputer, The ICT, SSC, & DIIC Ltd have agreed a joint project to build the fastest supercomputer system in China, to turn on in December 2008. It's the trade off - If you wait just a few more months, your money will get you a whole lot more boom for your buck.
1920 nodes of Tyan S4982 quad-socket server motherboards & sprinkled liberally with (presumably) ~7,680 quad-core AMD Opteron processors. They want a peak of ~230 TFlop/s - Which would get you into the top 10 worldwide currently.
So doesn't Apple have a go helping get some computationally challenged, cash laden organisation, with some Intel chips, whatever dual/quad-socket server motherboards are going, and some Mac OS X Snow Leopard beta juice?
Another possibility is though of somehow testing the feasability of using standard Snow Leopard Mac Pros of consumers, ala Folding Project or SETI... With 3/4 of the world's fastest supercomputers using Intel, it isn't too much of an extrapolation it might seem. The rest being mainly IBM Power processors, and AMD Opterons. I wonder what a tripped up maxed out blinged up Mac system with Intel Nehalem or Westmere chips could do on Mac OS X SL...
//Bits
In terms of Nehalem chips, apparently, more information will come in August at Intel's development forum. (including such terms as High K+strained metal gates...)
Can compilers mitigate the requirement for in-order operation?
Hyrbid SLI and CrossFire support from AMD and NVIDIA?
If the GPU can execute x86 instructions they could find a place in clusters/supercomputers.
The 45nm Penryn chip being used in Nehalem is Gilo - not just 32nm shrink of nehalem but a completely integrated gfx core?
G45?
Memory on the chip (One of the strands of research at the Intel parks)
Edit: 22nd July 2008:
Virginia Tech has assembled a Top 100 Mac cluster - 29 Teraflops based on 324 Mac Pros, at the Center for High End Computing Systems (CHECS) within the Virginia Polytechnic Institute and State University.
Using a Mellanox 40Gb/s InfiniBand tech piece of kit to interconnect the machines. So it's possible to do it with Mac Pros currently.
Roll on Nehalem and Snow Leopard! The 1st comment on the 9to5mac page was that you could do this cheaper with GPUs (Fastra project, NVIDIA Tesla etc). However, this supercomputer could potentially be very easily able to ramp up it's power through sequential upgrades and an OS change. I'd imagine there will be a similar project, with Mac Pros that have mainboards capable of more upgrades.
Example of current Apple hardware usage
http://www.apple.com/uk/pro/science/universityofnottingham/
Research into ophthalmology and other aspects of the human body, at the University Hospital, Nottingham -
It has a 5-node Apple Workgroup Cluster for Bioinformatics. (4 compute-only Xserve G5 cluster nodes with 2GB of RAM each, plus an Xserve head node, which will soon be upgraded to hold 4GB of RAM. In addition to this, two Xserve RAIDs give the system a total of 4TB of storage with the hardware linked together by a gigabit Ethernet switch.)
For the raw power we get from the solution, it’s very good value. Especially with the addition of the Xserve RAID arrays for mass storage, which have been of great benefit to us — ultra-reliable and very easy to set up and to manage user access. I only have to restart the machine when there are system updates, so I consider the cluster to be very low maintenance overall”.
Issues? "Tighe has only once needed help from the AppleCare Premium Service and Support team, and they swiftly diagnosed a failed component and immediately arranged for a replacement to be shipped out to the university."
SL might very well help bring a new level of power to bioinformatics and other labs which have computational areas. My pet example would be X Ray Crystallography - what was once left running on an old server - can now be done quicker pretty much by a souped up Mac Pro. With the right tech, any given Lab could yomp ahead if it got into Snow Leopard early - Once you've got crytals, and data, a fair bit of time is spent crunching the numbers, through an iterative modelling process. It could be cut down and even outsourced soon!
