SGI is officially committed to Linux and seems to be the company that comes closest to doing "the right thing" by working WITH the Linux community instead of trying to roll it's own kernel.
SGI has also sold off the vector part of Cray to Tera, hopefully that will be a better match than Cray-SGI was. And hopefully the new Cray company will create some more interesting hardware than the decidedly luke-warm SV1.
Define "smoke"? It seems like you want to compare I/O capacity, something that SGI is VERY good at. Of course an array of FC disks will outperform an array of UW-disks provided they all share the same channel, hardly a fair comparison. There are fibrechannel adapters for SGIs and trust me, they work great (>90MB/s on one channel from 10 striped 9GB disks). Yes, SGIs are expensive for the FLOPS they provide and does usually not make sense to use as workstations unless you do modelling or visualization, but they make excellent I/O monsters. It all depends on the problem you want to solve, there is no single system that is best at everything.
The recent SGI announcments are reasonable and I think that their strategy is sound. The big questions is : who's left at SGI? With the 'slightly' different job market for graphics hardware designers and the strange decisions of SGI in the past few years I'm worried that they have lost most of their best engineers, and SGI is nothing without their engineers.
On another note : 88 million polys/s on IR? Where does this come from? An IR pipe does 10 million polys/s AFAIK. The big benefit is bandwidth and the fact that basically all of OpenGL is implemented in hardware (a glVertex3f call turns into 7 (!) instructions on an IR machine), offloading the CPUs to do other stuff. But SGI must do something radical to keep hold of high-end graphics, recent delays in the future graphics line does little to inspire confidence.
How does pipelining help you retire more than 1 instruction per cycle? I'm sure that it is heavily pipelined, but AFAIK that only helps with reducing the CPI if it is larger than 1. I'm also sure that the chip is superscalar but 50GFLOPS sounds unrealistically high, the Fuzion 150 runs at 200Mhz, is MP SIMD with 1500+ processing elements and still does approx. 3GFLOPS. You'd also need to have 200-300GB/s bandwidth to sustain that kind of speed (assuming a modest 16 bit-sized floats).
Suppose the chip runs at 500Mhz (just for the sake of argument), then it would have to retire 100 floating point operations per cycle to reach 50GFLOPS. I *DON'T* thinks so.
Slashdot Mirror
Read up on the MTA concept. Sun stuff is only used
for I/O and not calculations, the MTA processor is
what's interesting.
SGI is officially committed to Linux and seems to be the company that comes closest to doing "the right thing" by working WITH the Linux community instead of trying to roll it's own kernel.
SGI has also sold off the vector part of Cray to Tera, hopefully that will be a better match than Cray-SGI was. And hopefully the new Cray company will create some more interesting hardware than the decidedly luke-warm SV1.
MIPS-Linux have been running on Indys for quite some time IIRC.
Define "smoke"? It seems like you want to compare I/O capacity, something that SGI is VERY good at. Of course an array of FC disks will outperform an array of UW-disks provided they all share the same channel, hardly a fair comparison. There are fibrechannel adapters for SGIs and trust me, they work great (>90MB/s on one channel from 10 striped 9GB disks). Yes, SGIs are expensive for the FLOPS they provide and does usually not make sense to use as workstations unless you do modelling or visualization, but they make excellent I/O monsters. It all depends on the problem you want to solve, there is no single system that is best at everything.
The recent SGI announcments are reasonable and I think that their strategy is sound. The big questions is : who's left at SGI? With the 'slightly' different job market for graphics hardware designers and the strange decisions of SGI in the past few years I'm worried that they have lost most of their best engineers, and SGI is nothing without their engineers.
On another note : 88 million polys/s on IR? Where does this come from? An IR pipe does 10 million polys/s AFAIK. The big benefit is bandwidth and the fact that basically all of OpenGL is implemented in hardware (a glVertex3f call turns into 7 (!) instructions on an IR machine), offloading the CPUs to do other stuff. But SGI must do something radical to keep hold of high-end graphics, recent delays in the future graphics line does little to inspire confidence.
How does pipelining help you retire more than 1 instruction per cycle? I'm sure that it is heavily pipelined, but AFAIK that only helps with reducing the CPI if it is larger than 1. I'm also sure that the chip is superscalar but 50GFLOPS sounds unrealistically high, the Fuzion 150 runs at 200Mhz, is MP SIMD with 1500+ processing elements and still does approx. 3GFLOPS.
You'd also need to have 200-300GB/s bandwidth to sustain that kind of speed (assuming a modest 16 bit-sized floats).
Suppose the chip runs at 500Mhz (just for the sake of argument), then it would have to retire 100
floating point operations per cycle to reach
50GFLOPS. I *DON'T* thinks so.