Free Transistors? Use With Caution – IGIC
November 3, 2011, International Gaming Innovation Conference, Orange, CA–Craig Hampel from Rambus wondered if gaming can benefit from the increasing availability of free transistors. Moore’s law forecasts ongoing increases in transistor density and performance and the game hardware makers are taking advantage of this largess.
CMOS is the workhorse technology for consoles and PCs. Only in the last decade has the advantages of scaling been challenged. Some of the problems with continuing scaling will affect the next generation of consoles.
Robert Denard from IBM described the rules for physical and electrical scaling that have driven the industry for the past 30 years. The results are that the cost per function and active power consumption are inversely proportional to the area if power density is a constant. One example is the PS2. by increasing integration and scaling, the main chip went from 600 mm2 to 86 mm2 in 6 years while maintaining constant performance.
Scaling drives economics. The cost of the silicon drives other costs and increasing the integration reduces power and the number of I/Os in a system. In ’94, a GPU contained about 1M transistors. Through scaling, now a GPU offers about 30 times the performance and over 1B transistors. The PS3 shows similar performance increases with constant power consumption. The Playstation’s are using the maximum platform power possible, so increases in power consumption are not possible like in the GPUs.
Since ’04, the next generation consoles have been power limited. Due to the power limits, higher performance is not possible through increasing the clock frequency. As a result, designs are moving towards multi-core architectures. The challenge is to program those many cores. Technology offers more transistors per die, but many of them have to be unpowered most of the time to meet power limits.
To continue future scaling within the power limits will force development costs to dominate all other issues. The cost per transistor will continue to drop over time. A million transistors now costs about $0.05 or under $1 for a PS2 chip. The new limits will change the ways that people look at silicon and game platforms.
The low cost of distribution for the silicon means that hardware will become the same as software, and will change the business models to one more like a software business. In the future, game platforms might be a proliferation and collection of peripherals that communicate with a server in the cloud.
At the same time, semiconductors will use new materials for the process and architectures will continue adding cores. Increasing parallelism while reducing frequency results in a small net gain compared to increasing the clock frequency. The power limits will force more of the silicon to be dark, and unpowered until needed. An alternative is to make more of the chip reconfigurable, so the same transistors perform many functions.
Ultimately, security will become the limiting factor as more games have virtual goods and the game content and user functions become server based. Here is a good use for dark silicon; security uses low duty cycle operations, just at boot and for verifying transactions. For freemium games (free to play, pay to upgrade), secure hardware enables modular upgrade capabilities.
The subscription model can also apply to the hardware, as resources can be enabled or virtualized depending upon factors like game level or paid upgrades. Secure and cheap hardware can allow distribution of resources on a pay-as-you-go or pay for premium model. You need to have hardware security to release the secure keys through trusted hardware. The software needs to be protected by the hardware.
Special purpose hardware will be developed for application specific functions. The developers need more robust tools to do this type of design. At the same time, some functions might become redundant in an upgrade mode, so the design tools need to understand this overlap. Power management is a critical issue, especially when entire blocks can be active or inactive between product release and end of life. The challenge is to plan for functions in advance to prevent obsolesce while including functions as embedded and upgradable features.
