A Quick Path to Memory

Our investigation begins with the most visibly changed part of Nehalem's architecture: the memory subsystem. Nehalem implements a very Phenom-like memory hierarchy consisting of small, fast individual L1 and L2 caches for each of its four cores and then a single, larger shared L3 cache feeding the entire chip.

Nehalem's L1 cache, despite being seemingly unchanged from Penryn, does grow in latency; it now takes 4 cycles to access vs. 3. The L2 cache is now only 256KB per core instead of being 24x the size in Penryn and thus can be accessed in only 11 cycles down from 15 (Penryn added an additional clock cycle over Conroe to access L2).

The L3 cache is quite possibly the most impressive, requiring only 39 cycles to access at 2.66GHz. The L3 cache is a very large 8MB cache, 4x the size of Phenom's L3, yet it can be accessed much faster. In our testing we found that Phenom's L3 cache takes a similar 43 cycles to access but at much lower clock speeds (2.0GHz). If we put these numbers into relative terms it takes 21.5 ns to get a request back from Phenom's L3 vs. 14.6 ns with Nehalem's - that's nearly 50% longer in Phenom.

While Intel did a lot of tinkering with Nehalem's caches, the inclusion of a multi-channel on-die DDR3 memory controller was the most apparent change. AMD has been using an integrated memory controller (IMC) since 2003 on its K8 based microprocessors and for years Intel has resisted doing the same, citing complexities in choosing what memory to support among other reasons for why it didn't follow in AMD's footsteps.

With clock speeds increasing and up to 8 cores (including GPUs) making their way into Nehalem based CPUs in the coming year, the time to narrow the memory gap is upon us. You can already tell that Nehalem was designed to mask the distance between the individual CPU cores and main memory with its cache design, and the IMC is a further extension of the philosophy.

The motherboard implementation of our 2.66GHz system needed some work so our memory bandwidth/latency numbers on it were way off (slower than Core 2), luckily we had another platform at our disposal running at 2.93GHz which was working perfectly. We turned to Everest Ultimate 4.50 to give us memory bandwidth and latency numbers from Nehalem.

Note that these figures are from a completely untuned motherboard and are using DDR3-1066 (dual-channel on the Core 2 system and triple-channel on the Nehalem system):

Memory accesses on Conroe/Penryn were quick due to Intel's very aggressive prefetchers, memory accesses on Nehalem are just plain fast. Nehalem takes a little over 2/3 the time to complete a memory request as Penryn, and although we didn't have time to run comparable Phenom numbers I believe Nehalem's DDR3 memory controller is faster than Phenom's DDR2 controller.

Memory bandwidth is obviously greater with three DDR3 channels, Everest measured around a 70% increase in read bandwidth. While we don't have the memory bandwidth figures here, Gary measured a 10% difference in WinRAR performance (a test that's highly influenced by memory bandwidth and latency) between single-channel and triple-channel Nehalem configurations.

While we didn't really expect Intel to somehow do wrong with Nehalem's memory architecture, it's important to point out that it is very well implemented. Intel managed to change the cache structure and introduce an integrated memory controller while making both significantly faster than what AMD managed despite a four-year headstart.

In short: Nehalem can get data out of memory quick like bunnies.