Intel's 8-core CPUs will have 2.3 billion transistors

By Wolfgang Gruener

San Francisco (CA) – Intel celebrates its annual research show-off at the IEEE International Solid-State Circuits Conference (ISSCC) with a look at upcoming Nehalem processors. The company is planning a “family” of processors with up to eight cores that will remain in the current dual- to quad-core power envelope. We have known for a while that Intel’s Nehalem processors will be available with up to eight cores and multithreading technology to enable them to run up to 16 threads in parallel. This week at ISSCC 2009, Intel will be providing more detailed information about the Nehalem processor lineup. According to the conference program, 45 nm Nehalem processors are on the way for mobile, desktop and server applications. Besides the fact that these processors will have up to eight cores, it is interesting to note that despite the core count Intel will be able to maintain the power envelope of its current Core 2 processors. On the mobile side, some CPUs will be rated at a thermal design power (TDP) of less than 10 watts, while high-end versions will be listed with a 130 watt TDP. Intel will discuss in more detail a 45 nm 8-core “Enterprise” Xeon processor, which carries 2.3 billion transistors. The company said that it uses core cache shut-off techniques to minimize leakage, but a power consumption of 130 watts in such a scenario is certainly impressive nevertheless. Four years ago, we had 90 nm single-core processors that were scratching at the 130 watt mark and we remember that the firm’s first dual-core processor, the Pentium D 800-series with Smithfield core, was also rated at a TDP of 130 watts. Other news coming from Intel include the description of a monolithic 45 nm 6-core Xeon processor with 1.9 billion transistors, 9 MB L2 and 16 MB L3 cache, which Intel claims is able to exceed a 1 million transactions per minute TPCC in a 8-socket configuration. Also, Intel will highlight a 65 nm quad-core Itanium processor equipped with dynamic frequency switching, which the company says enables a frequency-power optimization without stopping the clock. Surprisingly, there is no news on Intel’s Terascale chip, no new information on clock speed, Teraflops or power consumption. However, a small presentation provides some new insight in the company’s progress with its 32 nm progress. The company claims it has developed a 291 Mb 4.0 GHz SRAM chip in 32 nm high-k metal gate CMOS with integrated power management. The chip is built with 0.171 μm2 6T cells. According to the company, 128 kb subarrays consume only 5 milliwatts of leakage power at 1 volt. Intel showed at similar at the International Electron Devices meeting (IEDM) last December; however, back in December, the 1.9 billion transistor chip ran at only 3.8 GHz. 32 nm shrinks of Nehalem processors are expected to debut in Q4 of this year. It may be interesting to note that 2.3 billion transistor 8-core CPUs have 80,000 times more transistors than the Intel's original 16-bit 8086 CPU first released on 10,000 nm processes just 31 years ago, and one million times more transistors than Intel's original 4-bit 4004 CPU in 1971. UPDATED: February 10, 2009 - 10:06am CST Sandia National Laboratories recently conducted simulations on the benefits and pitfalls of multi-cores. Their simulations show that moving from 2 cores to 4 cores shows a big increase in performance. However, moving from 4 cores to 8 cores barely shows any increase. And moving from 8 cores to 16 cores actually shows a decrease, down almost to the point of a 2-core system in terms of overall compute abilities. While these limitations were imposed primarily due to memory bandwidth limitations, we must realize that this 8-core behemoth (16-virtual core) machine will suffer from the same types of memory limitations on certain workloads. The only applications to see big benefits from 8-core implementations are those where high compute and low bandwidth are needed, and these are fairly rare. In such a case, it would be better to have several individual dual- or quad-core CPUs each connected to their own physical memory along with NUMA (Non-Uniform Memory Architecture) logic employed in the operating system.