Purchase Microprocessor Report Articles Online Weekly collections of Microprocessor Report articles are now available for purchase and download online. Price $50. Click Here to Purchase
If you’re in the server processor business and you either did not ship a dual-core processor in 2004 or are promising one this year, you’re out of the mainstream. The next few years of server design will involve more cores and higher integration, as companies have more transistors to use in 90nm process generation to put more of the server system onto one die.
In 2004, the x86 server processor market continued the transition to the AMD64/EM64T/x64 (AMD/Intel/Microsoft names, respectively) instruction set started by AMD’s Opteron and accelerated by Intel’s Nocona. AMD was making progress with AMD64 (also previously called x86-64) instruction extensions, but it has been slowed waiting for Microsoft’s production release of x64 Windows operating systems, which didn’t happen in 2004 and is now likely in 1H05.
One of the most aggressive and daring designs for multicore, multithreaded processing will be Sun Microsystems’ Niagara processor, with eight quad-threaded cores on one die. The design uses a stripped-down UltraSPARC II core with minimal floating point. Design emphasis was placed on memory bandwidth and a reasonable amount of on-chip cache memory. The challenge for Sun will be software that can efficiently manage 32 threads on one die. If Sun succeeds, expect to see other vendors scrambling to catch up in 2007.
For quite a few years, “low power” was the attribute and specification of simple applications processors, mostly employing a small ARM core, a modestly sized cache, and a few peripherals. But the demands of communications and multimedia workloads have changed this scenario. Features and performance are escalating as service providers and cellphone manufacturers compete for the subscriber dollar. Communications and multimedia applications have made digital-signal processors indispensable and, with them, the special configurations of on-chip memory and DMA serving the DSPs.
Many interesting chips were announced in 2004, but only a handful came out of the oven. With some chip vendors, the delays are the outcome of spreading development cost over more years than we have been accustomed to seeing in the good times. With other vendors, the delays may be the outcome of late starts and the need to preannounce products to retain existing customers. Still other vendors may have silicon, but they or their preferred customers have chosen not to talk about the chips. Chip advantages are becoming system advantages, more so than in the past, and the chips’ availability, differentiating features, and price will have an impact on system sales and on the technical depth of the data that chip vendors will be willing to make public.
