How Intel or AMD can make a better processor

Things to make a better processor (My first new Topic)

1) Improve the uArch
2) Change the process

So let’s talk about the uArch. Currently Intel has released the new Core 2 uArch on for them an older stable process. This allowed them to leapfrog the older but well optimized AMD K8 uArch.

What Intel can do to improve the Core 2?

They look for speedpath problems in the uArch. Speedpaths are areas in the uArch that for whatever reason cause a bottleneck in the flow of uOps through said uArch. (This is a gross simplification but the idea is correct). So from the initial A0 stepping that came out of Intel’s fabs until we had the launch of the B2 Stepping that is the commercialized product. Intel fixed bugs in the uArch, fixed speed paths to ensure a high enough number of parts at the highest speed bin can be supplied to all of the eager customers. Also part of this process is looking for hotspots in the uArch that limit the actual speed of the parts that do get binned out.

Intel and all other processor manufacturers have a criteria called guardband. This is the amount of margin from a thermal limit or other operating limit that they must meet in order to sell us the public, a processor. As we can see from the overclocking potential of the Core 2, there is plenty of guardband.

Now Intel is moving ahead and fixing additional bugs and improving on the uArch to be able to release higher speed parts that we all like. We should start seeing the benefit of these improved steppings of the processor giving us higher speeds and lower thermals.

So now we come to the Process side of the equation. As stated the Core 2 is on a 65nm process and Intel is diligently working on the next process node of 45nm. This of course allows them to do several things. 1) Boost speed for the same thermals. 2) Decrease DTP at the same Frequency 3) Produce more parts per wafer allowing them to sell them at a lower cost but still make the same profit on them or at the same cost to us and improve their profit. This is all contingent on them getting the same yield on the new process as the old one. They deed to do this because a new 45nm Fab is costing upwards of $4 Billion or more. They need to sell millions of processors to recoup their investment above and beyond what they can depreciate the Fab for.

Now to make a point. The Core 2 uArch is new and Intel can optimize it and the 65nm process to get faster or higher binned parts out!


What AMD can do to improve the K8?

Well, not much on the architecture side of the house, because the uArch is 3 years old and counting. They have most likely improved it by fixing all of the easy speedpaths problems and thermal hot spots. This means that with such a well optimized uArch they will not be able to gain much more speed or reduced TDP that way. Since I have not heard of any uArch improvements to improve performance in the new Core "Brisbane" then AMD can only improved it by their new process. They need the K8L uArch improvements to gain back as much performance as they can to keep pushing Intel forward.

Now that they have moved to the new for them 65nm process we hope to see the following improvements. 1) Better thermals for the same speed which we have from the AnandTech article. 2) Improved performance by keeping the DTP of the top bin part at 125W DTP as seen on their 90nm process parts. We have not seen this yet meaning they will continue to either tweak the uArch to the new process or modify the process to help the uArch. 3) Improved number of die per wafer enabling them to lower their prices to undercut Intel’s. We have not seen this either at this point because there are no new Brisbane parts available to buy yet on the market.

Unfortunately AMD has just moved to the new 65nm process and yield will always be lower than the older mature 90nm process.

So let’s compare where AMD and Intel are at right now.

Process: AMD just moved to this and yield is most likely a lot less than their older 90nm process yield. Intel has been on 65nm process for over a year and yield should be much better than AMD’s. If AMD has converted their Fab 36 to only manufacture 65nm parts then their overall processor production has probably gone down until they get their yields up to their 90nm process yields. Intel has at least 3 fabs making 65nm processors and AMD only has one. Intel is moving rapidly to their new process node of 45nm.

Conclusion: AMD is way behind and the conversion to 65nm process is currently hurting them. Intel has room to improve the Core 2 uArch on their 65nm process.

uArch: AMD has about come to the end of possible uArch optimizations as seen by the 3.0 GHz FX74 low over-clockability. No reported uArch improvements to the K8 on the 65nm process. Intel is still early in the optimization of the Core 2 uArch. We should see improvement in Core 2 uArch on 65nm in the 2007 year. Penryn Core 2 uArch has seen improvements: 6Meg cache. Updated SSE instructions that I know about and I’m sure there are more.

Conclusion: I’m not expecting AMD to reach much beyond 3.0GHz in 2007 until they improve 65nm process yield and optimize the K8 core to this new process. Penryn core will improve performance for Intel.

This is my thoughts/opinion/consensus on where Intel and AMD are at right now with a look to the future. Please argue amongst yourselves and me.