skilzygw said:
Exhibit A)
Dual Core - how convenient they both realease around the same time
Until the 90 nm node, dual core x86 processors would have been completely impractical. Intel barely kept up with AMD's release using a somewhat kludgy multichip package, this was not a sign of coordination with AMD's more polished product.
Exhibit B)
Graphics - Minor jumps ahead from one brand to another
Both vendors go as far as process technology and their design expertise can take them. Both ATI and NVIDIA are the best in their field, which means that barring some mistake (FX vs 9700 anyone?), they will approach a similar ceiling for every hardware generation.
Memory is often a limiting factor shared by both. Cards with either ATI or NVIDIA chips will still use memory modules from a very limited pool of manufacturers such as Samsung, and board partners do not want to go hog wild with exotic layouts.
I seriously was asking this more about the cpu sector than anything else. I mean why does it seem like they are always aorund the same performance even if at different clocks?
Performance is similar despite clock speed differences because of trade-offs in design.
The P4 is an aggressively speculative speed demon that is actually very good at straight line execution with optimal software. However, it suffers from larger penalties due to branching and poorer performance in suboptimal situations.
The K8 relatively conservative compared to the P4, and it is more robust when dealing with funky code. The price is that it cannot be driven to the high clock speeds that the P4 can. AMD was lucky that Prescott turned out to be thermally limited, since it would have scaled beyond 4.5 Ghz otherwise.
There are still compute-intensive applications that only care about clock speed, while others have room for efficiency improvements.
Regardless of the approach, performance improvements take work, money, and risk. Designers must weigh possible performance and feature gains against uncertainties in how future manufacturing processes and software environments will work out. A high-level chip design is set down years before it ever reaches silicon. A lot can go wrong with that.
AMD and Intel have sucked out most of the easy performance gains, so only incremental improvements can be expected. This is particulary true with regards to memory latency. Most of the time, processors are stalled waiting for RAM accesses to go through. Unsurprisingly, the various chips are running from essentially the same pool of RAM types.
Why no radical shift by one vendor? Is it just a matter of intel or someother semiconductor shrinking the process then they can go quad core, faster etc...
Semiconductor scaling is still king when it comes to x86 chips, which rely on massive volumes to pay for design teams to beat their heads and throw transistors against one of the most inelegant and frustrating ISAs still in existence.
Radical shifts may or may not pay off. Even if they do, it is unlikely they will provide a definitive advantage against a competitor that played it safe. Low-hanging fruit in single-threaded execution is basically all gone.
Multicore to a point will provide much better performance scaling, since two cores will provide twice the resources on concurrent code. A single core that could match that would probably be four times as large and extremely hot, since a lot of performance improvements suffer from quadratic (or worse) increases in complexity and power consumption.
At the same time, transistor scaling is massively expensive. Fabs and the research into them are multi-billion dollar investments, so there better be something for them to produce when they are up and running. Pretty much all vendors outside of Intel must cooperate to maintain research and manufacturing growth.
Future nodes will also no longer allow designers to ignore power considerations. The fastest transistor is no longer going to be the coolest running or the smallest. The smallest transistor will not necessarily be the fastest or cool. The coolest running transistor will not be fastest or smallest.
As a result, pushing silicon gets too hard. Multicore gets around this for now because it deemphasizes clock speeds that can lead to a cubic increase in power draw, and can leverage density improvements that are easier to make.