Quick Path Interconnect - QPI

In order to unlock Nehalem's magic, Intel first had to remove some of the limitations that held back last generation's Core 2 Duo and Quad processors. The first thing to go was the Front Side Bus, the old system interconnect that allowed the CPU, motherboard and memory to talk to one another. With processors getting faster and memory sizes getting larger, the FSB became increasingly choked up transferring data between the CPU and RAM.

Intel's solution is a new point-to-point bi-directional bus called Quick Path Interconnect that transfers data directly between the CPU and the chipset. Every Core i7 processor is actually equipped with two QPI links, which could potentially allow for future multi-processor systems. It's exciting, but we're not quite there yet.

For now though, QPI simply connects the CPU to the motherboard over a set of 20-bit wide connections that operate at either 4.8GHz (Core i7) or 6.4GHz (Core i7 Extreme Edition). Since these links are bi-directional and allow the CPU and the chipset to both send and receive information simultaneously, the end result is 19.2GB/s of bandwidth between the Intel Core i7 920 and the Intel X58 Express chipset. The important thing to take away from all this math is that the Core i7 won't become bandwidth bottlenecked anytime soon, thanks to QPI.

Memory Controller

The other part of Intel's effort to eliminate bottlenecking was to reduce the amount of bandwidth that's transferred over the system bus. Following a move AMD made back in the days of the original Athlon 64, Intel has added a memory controller to the die of the Core i7. Removing the memory from the system bus means that there's no barriers stopping the CPU from sending information to the chipset at full speed.

Intel has of course taken the opportunity to kick the Core i7's memory controller up a notch. The DDR3-exclusive memory controller is designed to work with triple-channel memory, officially supporting speeds from 800MHz up to 1066Mhz.

BCLK and Clock Speeds

With the FSB gone, adjusting the speed of the Core i7 is now done a little differently. The base clock, or BCLK, is the primary means of over/underclocking the Core i7 920. The processor has a locked multiplier of 20, and multiplying that by the standard 133MHz BCLK yields its core clock speed, 2.66GHz. The BCLK also affects the speed of the Uncore (the memory controller and L3 cache), the speed of the DDR3 memory, and the speed of the QPI link.

Overclockers will be able to fiddle around with the QPI and Memory controller multipliers on the Core i7 920, although the CPU multiplier stays locked unless you're willing to pay for the $999 Core i7 Extreme processor.

Intel Socket 1366 Core i7 processors
Processor Models	Thermal Design Power	Clock Speed (GHz)	QPI (GT/s)	Cache	Price (USD)
Intel Core i7 975 Extreme	130W	3.33	6.4	1MB L2 + 8MB L3	$999
Intel Core i7 965 Extreme	130W	3.2	6.4	1MB L2 + 8MB L3	$985
Intel Core i7 950	130W	3.06	4.8	1MB L2 + 8MB L3	$562
Intel Core i7 940	130W	2.93	4.8	1MB L2 + 8MB L3	$538
Intel Core i7 920	130W	2.66	4.8	1MB L2 + 8MB L3	$273

Cache

In order to keep its four cores supplied with a constant stream of useful data, the Core i7 is equipped with 256KB of L2 cache per core, and 8MB of shared L3 cache. Intel's adoption of smaller, individual L2 caches reduces the amount of time each core spends looking for data, and in the event of a cache miss, there's a good chance the large L3 cache will produce a hit. While the memory latency on the Core i7 is greatly improved thanks to the on-board controller, it's still nowhere near as fast as on-die cache.

HyperThreading

To take advantage of all this processor-to-system bandwidth, Intel has brought back an old friend in the form of HyperThreading. This multi-threading technique was actually introduced six years ago with the Pentium 4, but has made a comeback with the Core i7. In standard multi-core processors, the individual cores often have to sit idle while waiting for tasks to be passed to them along the system bus. HyperThreading allows the processor to work on a second task (also called a thread), during that downtime, and swap between the two of them on the fly. This means that even though the Intel Core i7 920 only has four physical cores, it can actually process eight threads at once. Later on in this article we'll be looking at the effects of HyperThreading in applications that really take advantage of parallel processing.

Thermal Design Power

All of this microprocessing goodness is packed into a 231mm² die, which houses the 731 million transistors that make up the Core i7. The Nehalem family is produced on the same 45nm node that was used to produce Penryn, which is both cost and energy efficient.

A core tenet of Intel's tick-tock strategy is that any increase in power requirements has to be matched by an equal rise in performance. During the development of the Nehalem architecture, Intel's engineers raised the stakes even higher, requiring a two fold increase in performance for every increase in power usage.

Intel rates the Core i7 line with a Thermal Design Power (TDP) of 130W, which is coincidentally the same wattage use as the old Pentium 4's I mentioned above. Let's take a look at some power-usage scenarios to see if Intel's power-performance rules have really worked out.