Nehalem was a huge step forward, due to the simultaneous improvement in both the cores and the system architecture. Westmere has neither the luxury of improving system architecture, nor changing the underlying microarchitecture. Within those constraints though, Westmere is a solid improvement for multithreaded workloads. Two extra cores seem to yield roughly 20-40% performance improvements at the same power level as the previous generation. For embarrassingly parallel workloads, such as those that GPUs excel at, the gains are likely to be higher – approaching 50%.
AES performance increases due to dedicated hardware within each core, exposed through several new instructions. While we did not test any security workloads, Intel has claimed that AES will run approximately 9X faster for select applications of AES (such as full disk encryption), and initial tests seem to confirm this. In other applications of AES with smaller chunks of data, such as serving SSL protected content, dependencies between different chunks of data to be encrypted or decrypted will slow down the algorithm and the performance gains may be much smaller, perhaps 2-3X.
Competitively speaking, AMD is now in a rather tricky situation. AMD’s 6-core Istanbul is priced substantially below Nehalem. A high-end 22xx AMD part runs roughly $1150 according to the official price list, while a high-end 55xx from Intel is $1600. Direct comparisons between the two products generally tend to show Nehalem leading Istanbul modestly; and the relative pricing definitely reinforces that perspective.
With Westmere, the performance gap between Intel and AMD will widen even further. Whereas before AMD had two extra cores to offer, now the two products are at parity from a core count perspective. However, Intel’s processors still retain several distinct advantages: simultaneous multi-threading, a third channel of DDR3 memory, a more efficient cache hierarchy, power gating and turbo boost. From that perspective, Westmere vs. Istanbul looks similar to a rematch of Nehalem vs. Barcelona – with the exception that Westmere and Istanbul will likely have much closer frequencies.
The wild card from the AMD vs. Intel perspective is the pricing and impact of Magny-Cours and Nehalem-EX when targeted at 2-socket servers. While Istanbul will lag Westmere, an appropriately priced Magny-Cours (or Nehalem-EX) could be very tempting indeed for extremely parallel workloads if the overall system price is reasonable.
Overall the performance improvements in Westmere are a welcome addition to Intel’s product line up. The gains are not small, but neither are they substantial enough to motivate upgrading from Nehalem to Westmere – only in select cases will the gains be in the 40-50% range. Similarly, the power efficiency improvements are nice, but not profound compared to the prior generation. However, for anyone using Harpertown or older platforms, an upgrade to Westmere is dead obvious. Nehalem is a huge leap over all of Intel’s prior products, and Westmere only compounds the advantages by throwing in two extra cores and the new security instructions for the same price and power.