Intel to spill Nehalem secrets
Wither AMD? IBM? Sun?
It's the beginning of a new year, and that means chip makers will soon show off their wares at the annual International Solid State Circuits Conference, which runs next week in San Francisco. This time around, Intel pretty much owns the show as far as microprocessors are concerned,
Hosted by the IEEE, the ISSCC is one of two big chip events held each year. The other one is the Hot Chips conference hosted by Stanford University, which is held in the summer. Of course, you could argue that the Intel Developer Forum events hosted by that chipmaker are equally important, considering the dominant market share of the x64 chip on PCs and servers. You know IDF is important because rival Advanced Micro Devices is always luring members of the press across town for a look at its own chips and roadmaps.
This time around, the ISSCC event is a little skinny when it comes to microprocessors. The one thing that everyone is keen on getting a lot more detail on is the forthcoming "Nehalem" Xeon processors, which are based on a new microarchitecture and chip interconnection scheme called QuickPath Interconnect that smells and tastes an awful lot like Advanced Micro Devices Opteron-HyperTransport combo. Intel is also raising the curtain a bit on the next in line in the Core i7 64-bit chips. These are also in the Nehalem family but they're used in laptops, desktops, and entry servers.
A long time ago, Intel promised it would do a point-to-point interconnect for chips and provide a common socket for Xeon and Itanium processors. The interconnect was being implemented to boost system performance, and the shared sockets and interconnect was to cut costs for motherboard makers - and therefore PC and server suppliers. This Common System Interface was delayed. Then it was formally named QuickPath Interconnect last year. And now it's shipping in crimped form inside Core i7-based PCs.
But it remains to be seen if Nehalem and the not-yet-shipping "Tukwila" quad-core Itaniums, due any minute now, will share a common socket and common chipsets as originally expected. Maybe Intel will clear this up at ISSCC next week. And maybe Intel will start talking about "Poulson," the kicker to Tukwila, just to keep Hewlett-Packard's enterprise customers from storming Santa Clara with pitchforks and firebrands.
Intel is making three presentations at the event. One is for an eight-core, 16-thread "enterprise Xeon processor" that's manufactured on a 45nm process and has 2.3 billion transistors. This chip is almost certainly the high-end "Beckton" Nehalem EX processor that taped out last summer and that Intel Digital Enterprise Group general manager, Pat Gelsigner, showed off at IDF around the same time.
The Nehalem EX processors are aimed at four-socket and larger machines, while the "Gainestown" Nehalem EP chips are for two-socket machines. It's the latter Gainestown chips that are expected to ship sometime before the end of March in servers - if the launch hasn't been delayed by the economic meltdown. Nehalems include on-chip DDR3 main memory controllers, and you'll find three memory channels per chip.
The desktop and laptop Nehalem chips that Intel will be showing off at ISSCC next week scale up to eight cores on a single die. They have simultaneous multithreading providing up to 16 threads per chip, and there are three levels of cache memory. These "next generation IA processors" are implemented in 45 nanometer processes as well, with integrated memory controllers and a point-to-point interconnect. Intel says that these chips will scale from under 10 watts to 130 watts in heat dissipation and will turn up in entry-level (meaning single socket) servers.
Finally, Intel is going to be showing off some benchmarks on a six-core Xeon processor with 1.9 billion transistors. The company doesn't specify, but this has to be the top-end Xeon X7640 "Dunnington" processor that the company launched back in September 2008. The Intel abstract for this presentation boasts that the front side bus I/O circuits are implemented in the center of the die - a dead architecture, mind you, now that QPI is nearly here - and that an eight-socket machine using this six-core chip can crank through more than 1 million transactions per second on the TPC-C online transaction processing benchmark test. At last year's ISSCC event, Sun Microsystems' "Rock" UltraSparc-RK processor and Intel's Tukwila Itaniums were the stars of the event, at least where processors were concerned.
The 16-core, 32-thread Rock chips are supposed to weigh in at 410 million transistors and run at 2.3 GHz, dissipating a hefty 249 watts. Sun has been pretty quiet about the Rock chips and their systems since this time last year, but they did confirm last week that they are on track for shipment in systems in the second half of this year.
The quad-core Tukwilas - which were also supposed to ship last year like the Rocks from Sun were - are comprised of over 2 billion transistors and are implemented in a 65 nanometer process that Intel has retired for Xeon and Core processors. Each of the four cores on the Tukwila chip has HyperThreading, giving each chip eight instruction threads. And there's 30MB of L3 cache memory on the chip, up from 24MB on the current "Montvale" Itanium 9000 chips. Intel said a year ago that QuickPath enables processor-to-processor bandwidth of 96 GB/sec and peak memory bandwidth of 34 GB/sec with the Tukwilas.
IBM could have showed off Power6+ (perhaps quad-core) or Power7 (definitely eight-core) processors at next week's ISSCC event, and AMD could have talked about its future "Istanbul" six-core Opterons. And for that matter, Sun could have talked about future 16-core "Niagara" Sparcs, and Fujitsu could have provided some roadmap for its future Sparc64 chips. But alas, these chip makers are keeping their powder dry in these trying times. Maybe they will say more at the Hot Chips conference this summer, particularly as future products get closer to launch. ®