Top 500 supers: China rides GPUs to world domination If the June edition of the bi-annual ranking of the Top 500 supercomputers in the world represented the dawning of the GPU co-processor as a key component in high performance computing, then the November list is breakfast time. The super centers of the world are smacking their lips for some flop-jacks with OpenCL syrup and some …
Given that the GPU supercomputers with their x86 helper chips to run the network stack only achieve relatively low performance but many of them do so with commodity (for HPC) interconnects, compared to some of the very very expensive stuff from the likes of Cray, how do they all stack up on price?
How much cheaper is a TFLOP if you buy it the 10GigE and GPU way on commodity servers than if you blow a substantial chunk of your GDP on IBM or Cray?
"the GPU's flops are relatively inexpensive and the overall machine should offer excellent bang for the buck - provided workloads can scale across the ceepie-geepie of course"
Yes, but how useful is that power to the users? It's all well and good smashing a benchmark but aren't there quite a few restrictions on getting this speed from the system? To do with the size of data used and making sure it stays close to the GPU (all sort of shenanigans with getting it to stay on cache or at least on board etc) rather than moving out towards main memory i.e. the slower CPU only systems may be faster in real world usage no?
These things cost millions, or tens of millions upwards, not using them for anything more than governmental e-peen would be financially odd.
Some of the top ones are owned by government labs are used for nuke modelling, others for weather/climate predicting or rented out to people with a workload, could be genetic research, drug modelling, air flow modelling for the latest car/plane etc... all kinds of things need crunching...
No, you normally can't buy most of the computers on the list, especially when you get near the top. Nearly all of them are at least partially the result of government funding.
Which answers your second question. Many of the supercomputers are purpose-built to handle computationally-intensive tasks like accurate climate modeling (to better predict the weather), nuclear simulations (cuts down on the need for real-world testing), and so on.
The highest system on the list that looks like it is predominantly funded by a commercial organisation is the EDF research system at #37. You may like to suggest that this is partly funded by the French government (like #76), in which case it would probably be #78, which is just described as "IT service provider, Germany".
So yes, if you had the money, and the will to run the infrastructure, you could buy one and HP, IBM or Cray would be delighted to sell you one, but to be honest, what would you use it for! Even the Intel ones are not suited to run Crysis.
On reflection, if you go back 8 or 9 years, you can see a number of commercial organisations with systems in the top 100, including banks and telecoms providers. But this was when you could estimate the power of a system by adding the component parts, not by proving that it could run such jobs. The bank I used to work for had their SP/2 AIX server farms listed, because they were clusters. They were never actually used for any HPC type workload ever.
..had the fastest Jet, the largest submarine, the biggest missile cruiser, the largest Phased Array Radar etc.
On paper all looked very impressive and some of the technology is indeed impressive. But what many laypersons don't realize is that *actual* performance is vastly different from a single performance number.
"Actual Performance" might be defined as "ability to penetrate airspace unharmed and fly back". If that would be the proper definition, the Luftwaffe F4 Phantoms were superior to the Mig25, because they had a system to automatically hack into the Soviet AF airspace control system and spoof their identity. To the soviet controllers it looked like a Mig21 or a Mig23 when it really was a West German Luftwaffe F4. A rogue Luftwaffe pilot and his WSO tried the system and got to Dresden (deep in the east) and back unharmed.
So this computer might have great benchmark numbers, but is it *actually useful* compared to more traditional systems ? I guess it is a one-trick pony useful for dick-length contests and not much more.
Also, a great programmer/engineer/scientist with a normal PC will normally achieve much more than a bunch of retards fumbling with a supercomputer in terms of *useful results*. China didn't impress on the science front so far. Let's see whether this will change.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
