As vendors begin to rally around a new memory architecture for future distributed symmetric multiprocessing system design called NUMA – Non Uniform Memory Architecure (CI No 2,704) so a new topology, UMA, Unified Memory Architecture, is set to storm the personal computer world and cut the cost of the next generation of desktops. According to […]
As vendors begin to rally around a new memory architecture for future distributed symmetric multiprocessing system design called NUMA – Non Uniform Memory Architecure (CI No 2,704) so a new topology, UMA, Unified Memory Architecture, is set to storm the personal computer world and cut the cost of the next generation of desktops. According to Microprocessor Report, a personal computer with such an architecture combines main memory and frame buffer on a single dynamic random access memory array, eliminating the traditional stand-alone buffer and producing a saving of between $30 to $60 per personal computer. Several companies are now working on Unified Memory Architecture chip sets in which the system can allocate the exact amount of dynamic RAM required for the frame buffer, assigning a block for the frame buffer and whatever remains for main memory. But the savings, the Report observes, do not come free. For example, if a personal computer with 8Mb RAM, a separate 1Mb frame buffer, and a 1,024 by 768 display with 8-bit colour is redesigned with Unified Memory Architecture, its graphics subsystem will require – in this configuration – 768Kb of the 8Mb RAM ordinarily devoted entirely to the main memory subsystem. That leaves 7.25Mb for the operating system and applications, a 9% shortage that will mean more disk swaps. The Report suggests that this loss of main memory wouldn’t generally cause performance problems because most graphics applications are written for 8-bit colour and most screens are 800 by 600 or 1,024 by 768, and said the $30 saving is substantial in today’s mainstream desktop market, where system OEM customers negotiate over pennies. This explains why most OEM customers are enthusiastic about UMA. Savings will, however, also be implementation-dependent. Options include adding the memory for the frame buffer to main memory, resulting in 9Mb systems with 8.25Mb for the operating system and applications. Or adding a Level 2 cache to reduce the disk swaps. Whatever the implementation details, the Report expects the architecture to trigger a new round of consolidation in the core logic and graphics markets, describing the technique as a paradigm shift from the 16-year-old personal computer architecture originated by IBM Corp. Coupled with a move to get rid of the 13-year old AT bus – the forthcoming distributed direct memory access standard will enable manufacturers to support AT devices on PCI – the architecture, it believes, leads to a new level of integration, with more functions and features packed into personal computers without any significant price increases.