>The semiconductor guys don’t know much about computers
and you'd expect an opposite huh. Can anyone comment on how accurate this statement is?
Hopefully one day I'll finally get to the point of being able to design my own circuits.
and you'd expect an opposite huh. Can anyone comment on how accurate this statement is?
Hopefully one day I'll finally get to the point of being able to design my own circuits.
Modern x86_64 and ARM designers absolutely know a lot about computers, and the expectations of modern software. That’s why we see constant advancements in things like vector instructions, needed for video decoding and photo manipulation, and specific instructions to speed up random number generation and block cipher modes such as AES-GCM. Then, at a higher architectural level we have much better integrated memory controllers, protected memory implementations, and branch predictors than could ever be dreamed of in the 90s. The Intel Core architecture reigned in the cost of pipeline flushing, and led to a 15 year renaissance in efficiency, now being led by AMD’s Zen architecure.
There have of course been advancements as well in semiconductor physics and chemistry, such as Optical Pattern Correction and ridiculous advancements in DRAM dialectric efficiency that isn’t touched on anywhere here; and I would consider being able to physically manifest these near impossible artifacts of technology “knowing something about computers”
Perhaps the designers know a lot, but Intel and AMD seem to have been driven by sales & marketing, not by knowledgeable designers. Look at the list of "features" being regularly added to Intel CPUs - and then look at how many have been depreciated or removed, and consider the large proportion that have never been widely used.
Don't waste those transistors - give me the simpler & cleaner architecture, and a little more cache.
I actually think they're talking about packing general compute cores into a super computer that would be better off having RTL specific to the task. Like AVX 512 (which Intel seems to be moving away from while AMD is embracing?). But that's kind of an early 2000's view of super computers. They have been packing accelerators on them (notably graphics cards) to crank up the teraflops. Whether or not those were effectively exploited, or they actually got the performance, I don't know. In my work I've seen fewer customers head that way, but that's not the pool I swim in. Some machines added FPGAs to the mix. We use them on our custom network cards so we do more of the packet processing outside the kernel. But maybe the weather super-computer should have some RTL for Navier - Stokes (?) that's what I think the weather people use? Or maybe make a custom silicon just for super-computing with a whole bunch of AVX512 units? Or maybe a lot more on-chip separate instruction cache? Might not be useful for CoD or Java but awesome for SC. But I do agree with you their characterization of modern super computers is a little dated.