4 ms·
> I still remain amused to guess what kinds of programs were included in the architecture simulations that led them to want so many integer units (12?) and so f
by aidenn0 6y ago
> I still remain amused to guess what kinds of programs were included in the architecture simulations that led them to want so many integer units (12?) and so few floating point ones.
If I had to guess, SPEC Int.
I one had a frank conversation with a CPU designer at a chip company. I was looking at a CPU that was fairly wide superscalar, but had super tiny caches. I observed that this chip was probably only efficient at running Dhrystone. His response was "We win and lose 7 figure CPU sales based off of Dhrystone benchmarks. DMIPS/Watt is the only benchmark that matters for the majority of our sales in dollars"
- boulos 6y agoSure. Specint and the more modern specintrate are obvious choices for “we can’t suck at this”. But the overall architecture was so extremely over “int heavy” that the box was just not great for the supercomputing segment it claimed to want to play in. Maybe large Oracle databases, too? The architecture really was the only game in town for a while for having so many physical memory address bits. I just assume someone “resold” it to the HPC/supercomputing groups as “here, go sell this”.
- bonzini 6y agoNot as bad as Niagara, that thing had what, 1 FPU shared by eight cores? Great for Java, a bit less for PHP that only had floating point arithmetic...
- boulos 6y agoPoor Niagara :). The Niagara design though was clearly intended for web serving/databases and Sun didn't go try selling it into the LINPACK/supercomputing segment. That's the disconnect for me. To be clear, if you had a multi-hundred processor system sitting around (as we kind of did, thanks SGI!), you used it. But... you probably shouldn't have purchased it.