A question about BlueGene's ASIC

Techno+

Regular
Hi,

I was reading about the BlueGene SC at wikipedia - http://en.wikipedia.org/wiki/BlueGene when i stumbled upon this sentence

"Each Compute or IO node is a single ASIC with associated DRAM memory chips. The ASIC integrates two 700 MHz PowerPC 440 embedded processors, each with a double-pipeline-double-precision Floating Point Unit (FPU), a cache sub-system with built-in DRAM controller and the logic to support multiple communication sub-systems. The dual FPUs give each BlueGene/L node a theoretical peak performance of 5.6 GFLOPS. Node CPUs are not cache coherent with one another."

the diagram that explains it

http://en.wikipedia.org/wiki/Image:Blue_Gene_L_ASIC.png

Now the PowerPC 440 is a General Purpose 32 bit RISC CPU, how can you put can such a compute node ASIC when it is made out of General-purpose hardware?
 
Why not? There are plenty of SCs that use standard of-the-shelf Intel/AMD "general purpose" CPUs. In fact most of the MPP SCs use standard CPUs that's why their DP FP performance per CPU is so low. The non-general purpose ASICs are the ones with high DP FP capability like the Clearspeed ASICs or NEC's vector processors.
 
Yeah basically
Virginia Tech's X runs PPC 970, its just a ton of G5 Apple's
Lonestar (some uni in texas runs it) is Pentium 4 Xeons

additionally, the CPUs in BlueGene are dual PPC 440's, its a somewhat similar design to Sun's UltraSPARC T1 setup, the difference is that we're talking 2 cores, and each has an FPU, instead of 32 cores sharing one FPU and a different memory crossbar

hell, Cell is basically the BlueGene CPU streched out
www.top500.org
look at the current list
the majority of systems are running on AMD Opteron or Intel Xeon or Itanium 2, along with IBM's POWER4 and POWER5 architectures

calling POWER4 a "general purpose" chip is sort of like calling Pentium 4 a "server" chip imo though, its a higher dollar server RISC chip, similar to HP's PA-RISC, DEC's Alpha, and MIPS Technologies' R series chips, its a commodity chip in the sense that its mass produced and individually "weak", however its not a consumer grade chip

I'd venture further to draw the distinction from modern cluster computing and "actual" supercomputing, mainframe computing, etc in that a "true" supercomputer by original definition had entirely custom ASICs, so the processors were only for 1 or 2 models of system

POWER4 is found in hundreds of various systems, everything from G4 series Apples to IBM rack servers, to clusters, to BlueGene, etc, the POWER4 is a fairly versatile chip

compared to say, the ASICs used for the Cray 2 which were only really used in the Cray 2, and inspired elements of the Cray 3, they weren't "general purpose" they were specialized for their task

in short i'm saying that a lot of terminology referencing system components is deprecated or "mis-used" contrasted to its original intent

for example in 1987 a supercomputer would be understood as a large system comprised of maybe 16 processors, each one custom built for the system, and characterized by having a large amount of very fast RAM

in 2007 a supercomputer is classified as a large system comprised of potentially hundreds of thousands of processors, very densly packed, and some sort of specialized backbone I/O which is rarely unique to the machine, and the machines are instead defined by their software applications
for exapmle MOSIX, OpenMPI, MPICH, Solaris 10/Sun Hypervisor, IBM VIVA/VIVA2, UNICOS-xx etc

mostly because "commodity" hardware is usually so close to the best of the best, when you look at price and thermals
compare the Cray X1 MCM chip to the Itanium2, the X1 is a fairly specialized chip that it only appears in the X1 platform machines, while the X1 is like 3 or 4 times faster than the Itanium at FP ops in a 1:1 comparison, its much more expensive, proprietary, etc while the Itanium is much closer to conventional computing (through software emulation x86 code is possible with very minimal performance cut) along with its EPIC architecture, meaning it goes for higher ILP efficiency, and it costs a lot less, $1000 a chip is a lot better than whatever 5 figure price the X1 takes (I don't know the exact price of an X1, but i know it isn't cheap, given that X1E systems are trading in the 8 or 9 figure range and they aren't "world record" worthy systems at that price point)
 
Back
Top