CPU architecture question

[Maximilian]

How true is this:

http://forums.anandtech.com/showpost.php?p=32833243&postcount=34

I know the pentium PRO is like the grand daddy of all modern CPU's but its not the first time ive seen someone say that nehalem was the last architecture to be related to it.

What made sandy bridge so different? Or is it not all that different and people are mistaken when they say sandy bridge was brand new and completely unrelated to what came before?

 

[Nothingness]

Sandy Bridge out of order execution engine was heavily changed dueto the switch to physical register file. See http://www.realworldtech.com/sandy-bridge/5/

Is that enough to say it's a brand new micro architecture, I'd say no, but it wouldn't be hard to convince me of the contrary 🙂

 

[cytg111]

some1 else has to dime in here, but I remeber back in the days when I had an oc'd 266 K6 part, and a 200mhz ppro still outperformed it in quake2(fpu intensive and all). As I understand it it was the architecture to be - until netburst came along, and when netburst failed they went back on some derivative of the old line (p2,p3,banias), wich in turn turned into core

 

[TuxDave]

What made sandy bridge so different? Or is it not all that different and people are mistaken when they say sandy bridge was brand new and completely unrelated to what came before?

To me, "brand new" means that it was created without using a previous processor as a baseline. I probably would also accept cases where over 50% of the code was deleted and functionality was recoded back in one by one.

Using that, then I would say that it's fairly rare to have a "brand new processor" because there's simply too much complexity to account for. IMO, Sandybridge wouldn't count under those terms.

 

[Exophase]

It's kind of arbitrary where the line is drawn. I remember reading from Andy Glew, a CPU architect that worked on some of these processors, that he too considered Nehalem/Westmere to be the last member of the P6 family. But the odd thing is that the context of the quote was his looking at the change in cache hierarchy, moving to strictly inclusive last level caches. IIRC that was added in Nehalem, which determined the whole mostly inclusive private 256KB L2 + strictly inclusive large L3 design that we still see up to Haswell.

Another major change (IMO) that Sandy Bridge brought is the uop cache, although you could see the progression from Core (added a pre-decode loop buffer) and Nehalem (post-decode loop buffer).

What is definitely not the case is that Sandy Bridge is brand new, every uarch Intel made since Pentium Pro that wasn't a Pentium 4 is clearly an evolution of the previous uarch and probably leaves alone substantially more than it changes. On paper Haswell actually looks like one of the biggest changes.

Netburst may have been 100% new, just about everything I can think of about it was pretty different from previous Intel processors, even the ALUs. And if anyone's wondering, I highly doubt Bulldozer was totally new, given how similar some parts are to older AMD processors.

 

[sm625]

Take away the OoO engine, the multiple pipelines and execution units, the SSE, the AVX, and you basically have the same chip as 20 years ago. The most significant step has probably been multiple cores (Athlon X2 and pentium D). After that comes the big front end microcode changes of the Pentium Pro. 3rd would probably be Nehalem/Athlon 64's memory controller latency reductions or perhaps SSE. It's tough to say with SSE since its been a slow evolving process. But then again that's how its been in general. It's not like they've done anything truly wild or off the wall, like the transmeta design. Had intel actually bought/swallowed/consumed/copied transmeta, chips might very well be a whole lot different today.

 

[zir_blazer]

some1 else has to dime in here, but I remeber back in the days when I had an oc'd 266 K6 part, and a 200mhz ppro still outperformed it in quake2(fpu intensive and all). As I understand it it was the architecture to be - until netburst came along, and when netburst failed they went back on some derivative of the old line (p2,p3,banias), wich in turn turned into core

P6 had a very strong FPU, while the K6 one was archaic (Not even pipelined). Later AMD incorporated 3D Now on the K6-IIs. I recall there was a Quake 2 client compiled to use 3D Now that matched Pentium II performance at the same Frequency, check here.