Asustek has unveiled its first supercomputer, the desktop computer-sized ESC 1000, which uses Nvidia graphics processors to attain speeds up to 1.1 teraflops. Asus’s ESC 1000 comes with a 3.33GHz Intel LGA1366 Xeon W3580 microprocessor designed for servers, along with 960 graphics processing cores from Nvidia inside three Tesla c1060 Computing Processors and one Quadro FX5800
“Designed for Windows Vista”
What!?
You mean like not all supercomputers run Linux? 😉
What !!!!
Do you mean that I need a supercomputer to run windows7 ? Microsoft need to optimise their code then !
😛
actually, if you install vista, you’ll need a graphic card and will only be able to install 3 nvidia tesla cards with 720 cores. several competing companies have similar offers:
http://www.nvidia.com/object/tesla_supercomputer_wtb.html
you can get a decent personal number chruncher with 4 tesla cards for about US$8000
Looks like a single core isn’t going much above 3 GHz any time soon.
I hope in 5 years time a ‘supercomputer’ isn’t just 16 times as many 3GHz processors, shoved into an AT tower case.
And what’s wrong with that?
64+ cores, 1 hard disk.
64+ cores, 1 ethernet card.
64+ cores, 1 keyboard.
There is only so much parallel processing can help in a serial world.
multi-core hard disks… coool! 🙂
a 234 ghz core won’t help you either if you need to access these very slow components.
If they are slow. I agree. But a 234 GHz core would be a huge performance boost if serial IO can keep up to a 234 GHz bus.
But 100x 2.34 GHz cores would see no advantage to having data arrive at 100 times the processor speed.
Moore’s Law becomes Moore’s Curse: cores cores everywhere, but not a Hz more processing power 🙁
Multi-channel SSD’s could solve the disk issues. Imagine a drive or device that could handle 16 (or 32/64/whatever) read or write operations simultaneously from separate threads. It would almost require a kernel on the device to manage the requests and take care of dependencies, but it is within the realm of reason. It could effectively give the OS DMA type access to the hard drive.
Multiple and/or multiplexed Ethernet ports could fix the network issues. Many motherboards come with two Ethernet ports now and if the OS would “automagically” divide the network traffic between them it could boost performance. Include smart routers and switches that are “multi-port PC” aware and you have a considerably faster network. Now add to it some sort of technology that could multiplex different streams (from separate threads) across the same wire and you could have a very significant improvement to performance in a multi-core world.
I can imagine plenty, but there is a good reason why they are not rolling off the production line. This is a much harder problem than it first appears.
I suggest a google of Amdahl’s Law. Even with some of the devices you describe, the scope for actual improvement in processing speed is very limited.
The modern computer will remain serial at heart for many years yet 🙁
Processing speed isn’t the bottleneck in your average computer these days. Stick a 10K RPM highly cached hard drive in an old PC and you’ll breathe new life into it without ever touching the processor.
Having said that, there are a great deal of processor bound applications out there that could utilize more cores than they currently do. Programming models and languages haven’t exactly caught up withe rapid multi-core expansion of the hardware industry. It’s all about rethinking the problem from the ground up.
New programming languages can help a great deal in this regard. If you still use a procedural language on a multi-core computer then there is only so much you can expect to gain. Procedures are, by definition, sequential. If, however, you restructure your application into a collection of service processes that respond to requests you will achieve a much higher utilization of the multiple cores (for a loss of efficiency most likely, but it could be a worthwhile trade-off).
Not everything can be done that simply, of course. Gaming, for instance, can only be parallelized so far. You could thread the sound, graphics, AI, physics, and several other things. But some of those components (graphics in particular) are currently far too linear up to the point the information is sent to the GPU.
RAM can also be a bottleneck in any SMP environment. If every CPU core is forced to ultimately access and manipulate the same bank of memory then you lose efficiency with more parallelism. Maybe it’s time for multi-channel memory so that every core has a direct and independent link to main memory, or at least a portion of it. This would assume a functional OS in this environment and applications written to make use of it. I personally think it is the SMP model that will ultimately hold back massively parallel systems on the desktop.
It isn’t SMP thats holding back massive parallelism from the desktop, its *cost*.
There are already alternative/better models to SMP out there, like NUMA or cluster-computing, they’re just more expensive to impliment.
Besides cost, the other (big) problem with massive parallelism is that in reality there are few computing jobs that can easily be divided into so many small ‘chunks’. Most of the work people tend to do on their desktops simply doesn’t benefit from it after a certain point.
More appropriately:
64 cores: 1 keyboard, 1 mouse, 1 ethernet card, 3 sound outputs, 3 disks, 2 monitors, 3 bittorrents, and that new game (with separate and multiple threads for physics, sound, graphics, AI, input, network).
Currently, adding cores is easier than ramping up clock speed. Most tasks where performance is key will benefit from multithreading as much, if not more, than simply higher clock speed.
Looking in the task manager, I have 650 threads going, and the only thing I’m doing is cooking a late breakfast. That would be a pain, but luckily I’ve got multiple heat sources.
It’s not a serial world. DOS died a long time ago.
Only if those tasks can be decomposed into smaller pieces for parallel execution. Some tasks simply don’t decompose well.
95% of them are sleeping.
If all 650 of those threads were actually doing something at the same time, you’d be able to cook your breakfast on your CPU, or what was left of it…
Inside your computer, it mostly still is.
What does DOS have to do with this?
Besides, believe it or not, its still being used.
What’s wrong with an AT tower case? I hope we can always put our desktop computers in them. Would you prefer the future to be the all in one imac style? Or the super small mac mini?
I like my cases to have space to fiddle around with their innards without busting my knuckles are requiring them to have cirque du soleil flexibility.
Nothing! 🙂
But I am hoping that distributed computing will take off in a big way. Instead of having 1x 65535-core processor that doubles as central heating, computers can be broken down into simpler, interchangeable pieces.
I really don’t see the “news” aspect here. Systems like this have been available from a slew of vendors like Colfax International and Penguin Computing for years now (heck, even from Dell). Plus it’s really a bad time to release a “tesla system” since NVidia’s new Fermi core just debued and it’s general processing power via CUDA is far more powerful and capable than the current generation teslas.
What is the deal?
IIRC the Radeon 4890 had more than one TFlop in a real test, so it would be interesting in which case this “super computer” reaches these figures, so how they would compare.
AFAIK Radeon cards reach more GFlops than Nvidia cards –> there was a very interesting thread on this issue on beyond3d.com.