Where to begin? How about with Bobcat:
Bobcat
Bobcat is already a winning architecture. The focus needs to be on refining it and keeping it competitive. The 28nm die shrink is paramount in this regard, as it will reduce power consumption and allow for higher clocks. The Radeon 7 series is also almost upon us, which will further improve the chip's dominance in the netbook graphics market. Integrating the AES instruction set would be a logical choice since many mobile applications have sensitive data and precious clock cycles shouldn't be wasted on decryption where possible.
Why have AMD not yet developed a Bobcat SoC? With an integrated memory controller, graphics chip and CPU, the next move is to integrate the rest of the chipset, such as a USB controller, SATA, LAN and audio onto one module, reducing size and power consumption. The netbook and low power PC market would love it; smaller devices with more space for batteries, cooling, storage and lower shipping and storage costs.
It could be presented amazingly: a presentation with a gamer playing TF2 in 720p, then revealing that the system is the size of a portable HDD, with a 2GB SODIMM, some flash memory and a passive cooler. My guesses say that on a 28nm process with some tweaks, such a chip could run at 1.4 or 1.6GHz, which should be enough for an older game such as HL2 or derivatives, if only at low detail.
Bobcat's reach can easily go beyond the netbook: think about all the retail locations that use terrible old POS systems on P3 and P4 computers. Those rigs will fail sooner rather than later, and what better to replace them with than a low cost rig that is smaller, produces less heat and noise, dramatically lower power consumption (if your store chain has 1000+ POS tills, that's a lot of money a year) and is 100% software compatible with your existing software.
Embedded markets are growing, such as cars, interactive maps and news feeds at transit points, ATMs and more.
The hybrid ARM/x86 market is one that AMD should be having moist dreams over. A product like the Eeepad Transformer is a potential dream come true, if the logistics work that a tablet can run Android on ARM while not docked, then enable a legacy x86 setup when docked with some kind of virtualization or software layer that enables Windows or Linux with full speed and support.
There have been many attempts to kill x86 with (at the time) superior architectures, including those by AMD and Intel, however these have never worked since the market wants support for legacy apps, support for legacy hardware drivers and is so used to the x86 architecture. It's why Linux hasn't killed Windows and why you see so many embedded devices running Windows, too. AMD needs to cash in on this.
Fusion and Bulldozer
My first move in the consumer CPU space would be to make sure the engineers had all they needed to make refining BD a success. AMD has proven in the past that they are capable of making excellent products, both when Intel are performing well and when their competition is very weak. It's poor direction that seems to have held them back. If the rumours about automated design reducing 20% performance are true, maybe that should be fixed. Nothing about the module design is inherently wrong; it just needs fixing up in its implementation.
Apparently Anand said that the L3 cache on the FX chips didn't do much for the performance (due to latency maybe?), so why not remove or reduce it to save costs and implement a much faster, quad channel DDR3 controller instead. Quad channel 2133MHz DDR3 would be closer to the Socket 2011 coming out very soon. This would, of course require a new motherboard socket. With the space saved by the large L3 caches, a 6 module, 12 core chip could be made, providing yield and heat issues are dealt with.
AMD have gained a great reputation for being upgrade friendly. AM2 boards could even support some AM3 chips, never mind AM2+ ones. However, the AMx socket is getting a bit long in the tooth now, and with the introduction of Fusion chips, maybe combining them would be the right choice. Let's call it Socket FM3 (FM2 sounds like it should have DDR2 RAM). Using a combined socket for both Fusion and non-Fusion chips has its benefits:
* Fusion chips need more memory bandwidth to feed both the GFX and the CPU. HPC likes more fast memory. Win for both.
* AMD obviously wants to harness its competitive advantage over Intel (one of the few it has left) in its graphics department, so even enthusiasts CPUs will have integrated graphics at some point. They may be disabled at the user's choice, but AMD has the potential (providing they invest in OpenCL, but I'll touch on that later) to have OpenCL in all of its products and with dual graphics, a head over Nvidia.
* OEMs will only have to buy one or 2 batches of boards over various different types, saving money and making logistics easier.
* Users will be able to upgrade easier. If I had to buy a new board and new CPU, I'd move to Intel right now, but if my mobo didn't have to change, I'd buy an X4 or X6.
With a new socket and matching chipsets, AMD can have a future-proof setup with lots of choice and counter Intel's move of changing boards every year or 2. Let high end CPUs have lots of PCI-E x16 slots and quad channel memory while lower end ones have less PCI-E and only dual channel RAM. Make all new mobos support SATA6 and USB3 (duh!) to be fully up to date.
Trinity doesn't need much input; if Bulldozer can be refined with lower latency and shorter pipelines while improving IPC somehow; Radeon 7 series is a solid improvement over Radeon 6 and GF improves its 32nm process and yields, Trinity will be successful. Quad channel memory at higher frequencies will definitely help though.
In between
I'm stepping into unknown waters here (I don't claim to know a lot about CPU design) but to create the kind of mobile product that AMD have never been able to compete with Intel with, engineers should research to see how plausible it would be to move BD to a 28nm bulk process as opposed to the 32nm SOI they have now. I don't know anything about the benefits of SOI, or how it compares to bulk manufacturing, but if a 28nm BD chip is smaller and consumes less power, even at the cost of some performance, it might be competitive in a Fusion module against Intel in segments where Intel would otherwise be using integrated graphics (Macbook Air? 13" MBP?) but I'm out of my comfort zone a little here.
Software development
AMD is sitting on something special; a powerful CPU business and a powerful GPU business. They are the 2nd best at one and either the 1st or 2nd best at the other. And now more and more of their products will be shipped as one. They need to encourage software developers to take advantage of this.
Give devels everything they need and the incentive to create products that make the most of AMD's unique position in the market. Make sure there is the best documentation possible, in house software engineers to provide (free) support and have employees work on contributing to well known open source projects. Have coding competitions for who can make the best apps using AMD's BD and OpenCL technologies in harmony. Gives prizes to the winners. Give hardware to developers to make apps on. Intel are also going to be using OpenCL; hit Nvidia where it hurts and beat Intel with their superior GPU tech while preaching that it's an open, widely adopted standard.
Marketing
Make AMD a household name: TV advertising is expensive, but to beat Intel out of people's homes, you have to break into their living rooms and their minds. AMD are an unknown name outside of tech circles; they're a knock off or "2nd best" to general users - there's no trust or association. Marketing needs to build that by pushing the advantages of AMD to users: lower prices, great graphics performance, innovative technology.
Game Consoles
As said above, AMD need to keep their relationship with MS for the next Xbox. Their graphics chips are still strong and competitive, so that's good. Even if they have to give huge discounts, it's worth it for the long term relationship.
IBM
AMD has quite frequently worked with IBM, directly and indirectly. The SOI process most AMD chips are made with was pioneered by IBM, GF are part of IBM's Common Platform and as far as I've heard IBM have an x86 license, meaning that should AMD eventually fail and need to be bought out, IBM would be one of the few potential buyers who could keep making x86 chips, as AMD's agreement with Intel voids the x86 license if the company is bought out.
... And those are my thoughts. Any input? Note that I have little knowledge of CPU tech, and this is just hopeful speculation.