Slashdot Mirror

I completely disagree. The majority of the HPC realm still uses Nvidia only because they know CUDA and not because of any technological advantage. AMD has held the line at not allowing sloppy programming methods into their OpenCL compiler and that has held back a lot of HPC users from jumping ship. You can even see this in many complaints from open source projects, like Blender, where they refuse to produce proper multi-threaded code and rely heavily on the CUDA compiler to do the work for them.

The rest of your complaints, "shitty drivers", "piss-poor memory handling" and "worse performance per watt" are also bogus. I own or manage machines using a large number of Nvidia and AMD video cards, and have seen as many driver issues between the two that neither has come out worse. This is a typical fanboy stereotype that keeps being repeated with no real fact behind it.

Your second complaint is seen a lot in programming forums, but I have never seen anyone do a proper write up of any memory issues with any of AMD's generations and most of the conversations lead me to believe it was an issue of the programmer's personal preference not wanting to learn a second platform with less market share than an actual technical issue. Most of these issues would be alleviate if the programmer would just use a common optimized library and stop trying to redo the work themselves.

Lastly, AMD's offerings have historically produce more performance per watt and their latest offerings continue that trend. This, besides the bit shift ability you mention, is also one reason why AMD was used for Bitcoin mining and supercomputers.

http://www.tomshardware.com/re...
http://www.green500.org/news/g...

Now, my latest personal computer has an Nvidia GeForce GTX 980 inside because I more often need to fix CUDA code and noticed some of the games I wanted to play ran better on it (again, from the game designer's preference and not a technical merit). I personally own eight other video cards across AMD, Nvidia and Matrox (who use AMD GPUs these days) and three generations for testing.

And I am only sticking up for AMD because I admire their push to get people to code for multi-core better. Nvidia has been too conciliatory in the last six years in that respect, which is fine for their revenue stream and market share but not a good thing in the big picture for the broader computer industry. Since Moore's law has begun to slow, we are going to need a massive shift to multi-core optimized applications and we need programmers ready for that day.

AMD seems to be ready with the tough love to get everyone there while Nvidia keeps enabling bad behaviors.

I think you meant multi-gigaflop:

" the top spots on the list have been taken over by machines that combine commodity processors with coprocessors or graphics processing units (GPUs) to form heterogeneous high-performance computing systems.

With all eyes on the new TOP500 number one system, Oak Ridge National Labs' Titan, it was a system belonging to a neighbor at the University of Tennessee that debuted at the top of the November Green500 List. The National Institute for Computational Sciences' Beacon system has set the new energy efficiency bar at nearly two-and-a-half billion floating-point operations per second (gigaflops) per watt. Employing Intel's Sandy Bridge series of Xeon central processing units (CPUs) and four of Intel's Xeon Phi coprocessors per node Beacon achieved a peak 112,200 gigaflops of performance running the LINPACK benchmark while consuming only 44.89 kW of power."

source here

What you're looking for is the Green500 list, where the K computer is #6; behind an ATI GPU+Intel i5 cluster and two NVIDIA Fermi+Intel Xeon clusters (two IBM Blue Gene Q prototypes sit at the top). The first three are fairly small (100+ on the top500), but the NVIDIA systems sit at #5 and #54 on the top500 as well, so it doesn't appear to be a scalability issue. I have no knowledge of the design tradeoffs of the individual systems, but I'd say that it's fairly impressive that both the top500 and the Green500 have so many GPUs in the top 10, given that they're both CPU-dominated lists.

Fujitsu is fishing in the same waters as IBM does with their BlueGene machines: both lines are designed to deliver 20 PFLOPS and both are traditional systems in the sense that you don't have accelerators like GPUs, which are still awkward to program for the average physicist. Thus, to potential buyers the TCO would be interesting. From what I've heard BlueGene/Q is twice as power efficient as the Sparc VIIIfx design, but those were just 8-cores, not 16-cores.

So, assuming comparable total power consumptions and a affordable price tag, Fujitsu could snatch several deals from big blue, perhaps even the recently failed Blue Waters, although my money is on Cray for that machine.

5 of Top 10 most green supercomputers use GPUs:
Green 500 List

Each GPU is very high performance and so high power. Performance / watt is what counts and
here GPUs beat CPUs by 4 to 5 times. This is why so many of the new supercomputers are using
GPUs / heterogenous computing.

Thats a wonderful soapbox speech which has nothing at all to do with supercomputers or the issue at hand. Linpack is no longer the most relevant benchmark and just as we saw in benchmarks for video cards, the system can be gamed to show exceptional performance on the benchmark while failiing to live up to that same standard on real world work. For instance, current and near term systems from IBM are exceptional and place more emphasis on memory bandwidth and power use than just pure compute ability. Focusing just on Linpac is similar to just focusing on clock speed. There are a number of new benchmarks being developed which attempt to quantify performance other than solving sets of dense linear equations.

See for instance graph500 or green500 or hpc challenge

When a foreign computer wins, the benchmark needs to be changed? Now that is gaming the system, American style.

No, actually... we've needed a better metric for a long time. Linpack only measures the performance of a very small subset of problems that these machines are used to solved. There are a lot more constraints to consider than just floating point performance these days - which is why we're starting to see new stuff pop up like http://www.green500.org/ and http://www.graph500.org/. Linpack is still relevant, but it's only part of the overall picture.

That's a graphics card though. It's likely that they're using a more general purpose CPU:
The microwulf (first cluster I could find data for) performs at 58.34 Mflops/Watt, so 58.34 Mflops/Watt * 65 Megawatts = 3,770 teraflops
The most efficient computer on the Green500 gets 536.24 Mflops/W so 536.24 Mflops/W * 65 Megawatts = ~ 34,083 teraflops
And of course, that's assuming they don't have lights or heat..

New measurement for the top500

It's called the Green500 list.

Quite a few people in the industry now are starting to care about power efficiency at the other end of the performance spectrum too. The Green 500 list for example tracks Megaflops per Watt data for the top 500 supercomputers. Judging from this data the Cell processor looks very good.

The reasons for caring about energy efficiency at the high end are of course very different from what ARM is trying to do, which is to maximize performance within a given battery life envelope. For large installations it has more to do with operating cost and environmental concerns.

Linpack is a terrible representation of performance on real HPC applications. Linpack almost exclusively rewards FP ALU throughput, scales almost perfectly on multicores, and requires very little of the interconnect, and has very modest memory bandwidth needs.

Argh... looking at the wrong numbers. It has dropped from the 70s in terms of CPU power per watt to 91. Not very green after all.

http://www.green500.org/lists/2008/02/ranks1-100.php

Why didn't this hit Slashdot sooner? Anyway, it seems to be very "green" as well. It looks as though there is only one computer more powerful than it that beats it on the Green500 complimentary list:

http://www.green500.org/lists/2008/02/ranks1-100.php

Huh? One of the best ways to get onto the list is to figure out a new and innovative interconnect topology. Remember System X from Virginia Tech, it was in the top 10 on the list when implemented and cost a fifth of the next cheapest system. It's not all about money unless you are gunning for #1. Virginia Tech is also the creators of an interesting alternative list, the Green 500, a list of the most power efficient supercomputers.

With the relatively low cost and high availability of computing speed nowadays, the green500 list might become very important, as it is not only the environment-friendliness but also a lot of the running cost that is involved here.

Or, probably the greenest as measured in Flops/Watt.

http://www.green500.org/CurrentLists.html

At 0.95 volts and 3.16 GHz - the clock speed that was indicated at the fall developer forum - the processor provides a data bandwidth of 1.62 Tb/s and a floating point performance of 1.01 TFlops, according to Intel. This will fit nicely into the Green 500, http://www.green500.org/Home.html, lineup, eventually.