Comparing Westmere-EX and Nehalem-EX

Yesterday, Intel announced that their flagship server processor, the Xeon Nehalem-EX, is being succeeded by the Xeon Westmere-EX, a process-shrinking " tick" in Intel's terminology. By shrinking Intel's largest Xeon to 32nm, the best Westmere-EX Xeon is now clocked 6% higher (2.4GHz versus 2.26GHz), gets two extra cores (10 versus 8) and has a 30MB L3 (instead of 24MB).

As is typical for a tick, the core improvements are rather subtle. The only tangible improvement should be the improved memory controller that is capable of extracting up to 20% more bandwidth out of the same DIMMs. The Nehalem-EX was the first quad-socket Xeon that was not starved by memory bandwidth, and we expect that the Westmere-EX will perform very well in bandwidth limited HPC applications.

With the launch of Westmere-EX (and Sandy Bridge on the consumer side before it), it appears Intel is finally ready to admit that their BMW-inspired naming system doesn't make any sense at all. They've promised a new, "more logical" system that will be used for the coming years. The details of the new Xeon naming system are presented in the image below.

There is some BMW-ness left (e.g. the product line 3-5-7), but the numbers make more sense now. You can directly derive from the model number the maximum number of sockets, the type of the socket, and whether the CPU is low-end, midrange, or high-end. Intel also has the "L" suffix present for low power models.

Westmere-EX SKUs and Performance

Intel has three general classifications for the new 10-core processors. The E7-x870 runs at 2.4GHz and has a 30MB L3, while the E7-x860 runs at 2.26GHz and the E7-x850 runs at 2.13GHz; the lower clocked models are also downgraded to a 24MB L3. As there are versions for 8-, 4-, and 2-socket servers, we get nine new SKUs. Pricing on the highest end 8-socket models is very high, but that's not too surprising for the target market: the E7-8870 costs $4616, the E7-8860 goes for $4061, and the E7-8850 comes with a "bargain" price of $3059. For quad-socket chips, you can knock that down about $225, and dual-socket chips will reduce prices another $175 or so.

What this means is pricing for the fastest Xeons has increased by about $1000, a clear sign that Intel does not expect any competition in the small but profitable high-end x86 market. And indeed, the current AMD "Magny-cours" is targeted at the midrange server market, not the high-end. Bulldozer will be AMD's attempt to take on these new Xeon models, and it probably won't attack this high-end market before the end of the year.

Also present is a low power E7-8867L, which runs 10 cores at 2.13GHz but needs 105W instead of the 130W of the E7-8850. For the 25W drop in TDP, the price jumps up to $4172.

The E7-8837 is another interestin processor. It has eight cores and forgoes Hyper-Threading, so it's like the Nehalem-EX but with a 2.67GHz clock it runs 18% higher than the fastest Nehalem-EX CPUs. Thus, it looks like the 45nm to 32nm transition either delivers two extra cores and a 6% clockspeed increase, or it improves clocks by 18% with the same number of cores. Below is the complete table of new SKUs.

Taken together, the two extra cores (up to 25% more performance) and a 6% clock increase, along with some subtle core improvements, mean the Westmere-EX parts should deliver 20-40% more performance within the same TDP. It's not enough to revolutionize the server space, but more performance per watt is nothing to sneeze at.

Our Test Server

We are currently testing the new Intel QSSC-S4R server. Intel claims that even though the TDP of the best quad-socket Xeons is still 130W, the total platform power consumption is quite a bit lower. We are definitely interested to see if this is the case.

The test system comes with eight memory daughterboards, each with two memory buffers and four DIMMs per memory buffer. With 8GB DIMMs, you can easily fit 512GB of memory into the system.

The four CPUs are cooled by massive 3U high heatsinks.

We are performing some virtualization benchmarking (VMware vSphere 4.1 update 1) as we speak. SQL Server 2008 and power numbers will be added too. At a later date we should be able to give some Linux KVM benchmarking numbers. For now, we don't have any results to report, but stay tuned....