VRM 11 Required For 45-nm Processors
One reason for varying processor support is the voltage regulator circuit of 3-series motherboards. It needs to be VRM 11.0 compliant, which is key when it comes to 45-nm processor support. Let me say that the problem isn't decreasing voltage levels, but strong power fluctuation due to millions of transistors clocking up and down, or switching on and off. Remember that future quad-core processors will be able to dynamically adjust clock speeds for each core individually, and switch cores on and off depending on the workload.
This also means that any 965 motherboard that is VRM 11 compliant can technically support 45-nm processors. VRM 11 says that the circuit is programmed using 8-bit voltage IDs (VID), allowing for 0.00625 V voltage increments. The minimum operating voltage isn't 0.8375 V (as in VRM 10), but goes down all the way to 0.5 V. VRM 11 also comes with the option to share the load across more phases, and the circuit runs so-called dual edge modulation, which means that the controllers send multiple impulses to the transistors while using smaller capacitors. The goal isn't just to provide smaller voltage increments and less voltage for the 45-nm processor generation, but also to provide sufficient power at voltage levels that may switch frequently. This can be done by specifying tight slew rates.
Overclocking: A Good Start At FSB1900
We did not have much time to fine-tune our overclocking settings, but it is safe to say that upcoming motherboard designs (and maybe chipset revisions) will definitely lend themselves to improved overclocking capabilities. As previously mentioned, P35 is great for overclocking. When using MSI's P35 Platinum, we immediately matched the FSB 1900 clock speed with little effort. We are confident that we will see many motherboards that will allow for overclocking that surpasses FSB2000 clock speeds.
Speed It Up: DDR3 Memory Kicks Off
DDR3 memory still is based on double data-rate technology, meaning that data is transferred with the rising and falling of the clock signal, which effectively doubles the transfer rate. However, the memory works with so-called prefetch buffers, which are used to collect data in order to provide it to the interface faster. DDR1 memory works with a prefetch of 2 (DDR mode, no buffering), DDR2 runs on a prefetch of 4 and, you guessed it, DDR3 is based on a prefetch of 8. This is the trick behind speeding up memory performance; and it also explains why the latencies go up again: DDR1 memory works at CAS latencies of 2, 2.5 or 3 clocks. DDR2 runs at CL 3, 4 or 5. DDR3 now has a CL of 5 to CL 8. It simply takes time to access data to fill the buffers. For this reason you shouldn't expect DDR3 to outperform DDR2 right from the start. DDR2-533 at CL 3 cannot outperform DDR1-400 at low latencies for real life applications, either.
Each DDR generation uses higher memory densities, which means that capacities expand with more advanced memory manufacturing processes. The mainstream for DDR1 was 512 MB per module (1 GB total memory capacity). DDR2 has its sweet spot at 1 GB per module (2 GB total memory capacity) and we expect DDR3 memory to be widely used at 2 GB per module (4 GB total capacity) by the middle of 2008. JEDEC specified DDR3 memory to run at a default voltage of 1.5 V - the default DDR2 voltage is 1.8 V and DDR1 ran at 2.5 V. Many memory vendors over voltage their products within safe boundaries to tweak the latencies - all for the sake of better performance. History will certainly repeat itself with DDR3 memory.
Due to the voltage reduction, DDR3 memory will consume less energy. However, we could not confirm this in our tests, as the test systems with DDR3 memory consumed more energy than they did with DDR2 memory. Intel says that DDR3-1333 memory should have an equal power draw compared to DDR2-800, and there should be a 25% power saving at comparable clock speeds. We will see if and how this claim turns out to be true or not in the future.