... a computer that can run Crysis on full settings and over 30fps has been built.

The original french article may have gotten the figure wrong, and was probably badly translated.
The french press announcement states 103 GFLOPS of Nehalem performance, and 192 TFLOPS of NVIDIA GPU.
Finally, there are 1068 "nodes" but they may contain more than one CPU, so the actual characteristics are still uncertain.
But it will be a really interesting machine, sure.

Yeah... :P Funny comment man. But very valid.

Luckily the Crysis crisis is abating now. The newest generation hardware is starting to look fit to handle it.

It seems to me that Each big new Crytek game brings about a hardware revolution. The quality of their games overall probably justify it. I still have fond memories of when I was finally able to play Far Cry at full settings back in the day. Such a beautiful game that was... :) I still play it every now and again just to go swim. It's the next best thing to actually going to a tropical island.

Supercomputer performance rating is ALWAYS measured in DOUBLE PRECISION floating point operations per second (flops). All scientific computing and mostly all HPC systems do double precision (64-bit).

I'm EXTREMELY skeptical that even a next-generation nVidia GPU could do ANYWHERE NEAR a Teraflop on double precision calculations.

In the "teraflop" advertising you always see from Nvidia or ATI (or even Sony with their Playstation 3), they are always referring to single precision performance. Usually, single precision is performance is roughly 10X more than double precision.

I bet it couldn't run Crysis at full quality settings

funy but tru
it wil prolly b 3 yeers b4 i can aford a computer thet wil run crysis on ful setings at a good framrate (60fps is wat i lik)
i dont think this suprcomputre cuold run crysis wel caus the gam aint writen 2 use al thos cors
thets the prob with the multicors caus softwar aint utelizng the xtra cors

Hi "Loose",

GPUs and CPUs differ, but there is one thing that binds them, and that is FP64 DP. My very first story for TG Daily was performance of ATI FireStream cards, which resulted in FireStream 9170 having 480 GFLOPS in FP32 SP, and around 250 (best case) to 125 GFLOPS (worst case) when using FP64 DP.

It is not exactly unbelieveable that nV could have 500 GFLOPS FP64 DP per card (they currently don't support the format), but only in peak cases.

Currently, Achilles heel of GPGPU is branching, since performance in branching sucks atm. CPUs do brilliant job here... but with further GPU developments, and the fact that GPGPU is $1B market by 2010, it does not take a rocket scientist to believe that both Nvidia and AMD are working in 7th gear to get branching performance sorted. Potentially, GPGPU is a death bell for accelerator market, and for certain FPGA applications.

Only time will tell. But for now, things are looking good for the GPU. Yes, there are dozens of issues, but there is a significant effort put in place and there is a community around GPGPU. We'll see.

All the best,

Theo

A single GT200 will run Crysis in very high setting with everything set to maximum. 1.1 Teraflops should be enough.

It is 1,100,000,000 floating point calculations per seconds.

Heya guys, just wanted to let you know that nV is going to have some very interesting news about GPGPU initiative. That includes them being compliant with IEEE 754 formats. Will dig more information, and then you can expect a new story.

Thanks for reading,

Theo

Whats the good of this mumbo jumbo Supercomputers in our real life ... the world is going to an End still ...!!

Hi,

Concerning the DP vs SP thing :
YES i agree that the BULL statements add oranges to
apples to make chairs.

However, what really matters is THE APPLICATION.
See a thread about this on comp.arch.

There is a fact that the french beast will not be used
to compute weather forecasts (which is a specific niche).
Users of this machine will need it indeed to run
"trivially parallel" algorithms where data don't need
full, double precision. Think : images, pixels, indices...
A LOT of computations with little effect on numerical
stability.

The rest of the future cluster will handle DP with the Nehalems,
but it's another story. I have been confirmed this
personnally by the cluster's manager : the CCRT (hosting
the clusters) has several types of computers to best handle
different kinds of algorithms (vectors, high-granularity
cluster, etc.). And the scientists who create the software
are not beginners.


One more important clarification :

There is already a cluster (called Platine)
operating at 47TFlops.

There will be a NEW cluster, smaller, more powerful,
wich has been announced to reach the (theoretical) peak
of 300TF.

The GPU "coprocessors" will complement the FUTURE cluster,
not the existing cluster (so it's not an "extension",
"upgrade", or "addition").
Please ! Check your sources, don't trust amateur websites
and speak directly to the concerned people.
Damn, i have even visited the site where Platine
and its future neighbour are located.

regards,
yg(from) f-cpu.org

That one GT260 chip roughly does half a TFlop, those teslas aparently have 2 fully equipped GF280s on them, and squeeze ~600 GFlop.

Nothing new, 2 crossfire high-end AMD card are also 1Tflops. AMD was the first to announce it even. The NV 'variant' is more in want with scientists because its on IEEE spec.

Belgian astronomers already build a desktop system that did some really good and fast stuff...