Hacker News Reader

Comment by Osiris | Hacker News Reader

On a plane flight about a year and a half ago, I sat next to a Professor of Electrical Engineering who worked with various chip companies and startups on chip design. The discussion was very interesting.

In particular, we talked about Intel and AMD. He claimed that if AMD's architecture was produced using Intel's fabs, the chips would be better than Intel's.

I'm not sure if he had inside knowledge or it was just his industry experience, but I found it interesting that AMD's primary disadvantage is fabrication technology and not chip design. They have to design their chips better because they are competing often 2 processes behind Intel and still compete.

View on HN

> In particular, we talked about Intel and AMD. He claimed that if AMD's architecture was produced using Intel's fabs, the chips would be better than Intel's.

I would have believed this maybe 4 or 5 years ago, but these days I don't think that's the case anymore. There was a time maybe 7 or 8 years ago when AMD was hands down better, then Intel started doing good work with the original Core 2 and then Nehalem, innovating on both the architecture and process sides. Now AMD is kind of where Pentium 4 Prescott was in that their latest architecture has to be pumped up to insanely high clock speeds to be competitive with Intel's midrange parts. I don't think shrinking the process node would help there, because they just don't have the IPC, and I don't think going down to 22nm would help AMD push their clocks that much higher that they would suddenly be competitive. And power consumption would probably go through the roof if they tried that.

That said I'm probably going to build another AMD desktop just to support them though an Intel would be way better for my needs, only because without them competition would die on the CPU front. At the price points I'm looking at I could get AMD's fastest CPU which competes with Intel's mid range and it'd at least do everything I need it to do, but the Intel would probably use less power. I'm kind of curious if AMD's Bulldozer architecture benefits parallel compiling because my hand wavy intuition says that compiling should be mostly integer code, and their processing modules do have two integer units so vaguely speaking it might be able to perform close to how an 8 core standard system would perform.

Intel is essentially still using the Pentium II architecture, where AMD has refreshed theirs recently with the bulldozer. Intel's fab advantage (32->22nm gives a massive thermal advantage) is getting wasted on putting more GPU on die. Probably because they can't put more cores on there without blowing their thermal threshold. The only way to add more cores then, is for cores to share more component: share the sequencer (more vector instructions), share the FP pipeline (aka bulldozer module), etc.

Running bulldozer modules on a 22nm fab would definitely beat the Intel Xeons on the majority of the server applications. Increasing the cache sizes and increasing the set associativity (amd is at 4, Intel at 8) will already give it the edge.

With regard to the sharing of FP pipelines: this is not new, the Sun Sparks have been doing that for ages. As we add more cores and parallelism, we will go more towards finer-grained parallel architectures like the spark. The bulldozer is only a shy step in that direction, not willing to commit too much in loss of sequential performance.

It was a while ago, but I remember one of my professors said that AMD hired all the design people from DEC when Compaq brought them. He said AMD's Opteron processor was highly influenced by DEC's Alpha processor. That little advantage led to AMD's golden years when their processors we're a better value then Intel's. It was even enough to overcome Intel's obvious fab advantage. Remember the Pentium 4 Extreme editions? It was a desperate attempt to keep AMD at bay by stuffing it with cache.

Something Intel is essentially still doing, if you compare the cache sizes and set associativity.

That's strange. Intel's 32nm Sandy Bridge 2600k is more energy efficient overall and has better single thread performance than AMD's 28nm FX-8350, while having almost equal multi-threaded performance.

AMD's FX-8350 is a 32nm part. I don't know to what extent one company's 32nm process can be better than another, but but if there is a difference, I'd expect Intel to have the advantage. Given that the AMD part has 1.2B transistors on a 315mm^2 die and the Intel part has 1.16B transistors on a 216mm^2 die, I'd say the fabrication node isn't the whole story. (Even Intel's quad-core Sandy Bridge E part is 1.27B transistors in 284mm^2)

>FX-8350 is a 32nm part.

My bad.

>AMD part has 1.2B transistors on a 315mm^2 die and the Intel part has 1.16B transistors on a 216mm^2 die

Yeah, I recently read somewhere that AMD significantly improved die area and power consumption of next FX cpu by using tools from graphics division. So it looks like the difference is not only in Intel's process advantage.

http://www.anandtech.com/show/6201/amd-details-its-3rd-gen-s...

Two generations away, but eventually we'll be seeing AMD processors that are denser and thus cheaper, but with potentially lower clock speeds. I suspect Intel's got the resources to do a much better job of laying out a chip, but it's hard to say whether that's a process advantage or a chip design advantage.

Reply on news.ycombinator.com