Last weekend I was reading a few blogs on Intel’s recent Broadwell chip. The current offering is the EP variation. I regularly read & enjoy articles at The Register, The Four Hundred, AnandTech and The Next Platform. Working with Enterprise platforms for most of my career, I sometimes get critical (maybe sensitive) of x86 technologies. After reading a few I was motivated to put together a table comparing the features and performance from Nehalem (Gainestown release) through the latest Broadwell-EP E5-26xx v4 chips. Essentially the 2 socket systems.
I was just enjoying some political tweets when I saw a tweet by @TheRegister promoting
an article on The Next Platform by Timothy Prickett Morgan titled “Intel does the math on Broadwell server upgrades“. Tim writes “It all comes down to the math …”. He is right except the premise behind this statement is there is *more* value going with Broadwell-EP (ie E5-26** v4) chips vs previous models. I am not saying the author is saying this beyond the information as he understand it or was given by Intel.
This prompted me to stop my political tweeting since Trump & Cruz are in my home state today to write this blog. There has been a “Great Charade” played by Intel duping customers for years. Setting aside any comparison of Intel server chips to processors from IBM or Oracle, I will just focus on Intel vs Intel over the last 8 years.
The row labeled with “1” shows the Relative Performance score used from this The Next Platform article. My methodology is to divide the Rel Perf score by the # max number of cores available with that processor generation to obtain its per core score labeled by the row with a “2”. The row labeled with “3” shows the % of increase in cores from the previous generation. The row labeled with “4” is where it gets interesting. This shows what the score would be if using the Rel Perf per core score for the original Nehalem processor 0f .29 by the number of cores available in the current generation. So, Ivy Bridge is 0.29 x 12 to obtain 3.48. The The data in row labeled with “1” comes from Intel even thought I obtained it from The Next Platform article. Why is this important? The actual score for the Ivy Bridge processor of 3.73 is greater than the extrapolated score of 3.48. Having a higher score for row 1 over row 4 is better for Intel performance. However, when you look at the Haswell & Broadwell-EP processors, the actual score is below the Nehalem extrapolated score indicating there is a decline in per core performance.
Now that you understand the methodology, lets look at the results for each tick-tock. The Nehalem processor was released 8 years ago with 4 cores and a Relative Performance rating of 1.16 or .29 per core. The next release was the Westmere-EP processor with 6 cores having a chip score of 1.98 or .33 per core. Westmere was followed by Sandy Bridge-EP with a Rel Perf score of 2.55 for its 8 cores or .32 per core. After Sandy Bridge was the long awaited and much hyped Ivy Bridge delivering 12 cores for the EP model. Its Rel Perf score of 3.73 translates to .31 per core. Notice the trend? After IB-EP was Haswell-EP with 18 cores delivering a 5.20 Rel Perf score or .29 per core leading us to the latest and greatest Intel offering; Broadwell-EP with 22 cores & a 6.34 Rel Perf score or .29 per core.
What does this mean? With each Tick-Tock or successive release, Intel touts magnificent performance yet many of the improvements and performance benefits tout the total socket capacity vs its per core capabilities. There is nothing wrong with socket totals if this is whats required or your software is priced in this fashion. However, many enterprise ISV’s charge license & maintenance fee’s based on cores. Some are based on the total number of cores in the server such as Oracle and others simply base it on the number of cores required. Either way, the stronger the core the better.
Yet the data shows that per core performance peaked with Westmere-EP followed by Sandy Bridge-EP. At best you could argue performance has been flat over the last 8 years with little hope for the two anticipated successors in SkyLake and CannonLake. I extrapolated the data using the same methodology showing performance will be flat to regressing. This of course is consistent with the Intel Exec VP & GM William Holt who said “The best pure technology improvements we can make will bring improvements in power consumption but will reduce speed“. Intel began moving away from Moore’s Law to an economic and financial model to remain on a tick-tock schedule rather than taking the approach to build an improved chip; not just one with more capacity. Intel marketing is forced to hype the capacity increases in the latest chips as performance when it was mainly due to the addition of cores. Yes, there were micro-architecture improvements but that benefits the internal plumbing to maintain coherence & data flow through increased cores. Core counts increased by 50%, then 66% and now it is slowing down with Broadwell-EP as they run into another problem which Intel’s Holt points to as well…..maintaining performance while maintaining TDP.
All of this as Intel moved from PCIe2 to PCIe3, DDR3 to DDR4 memory, more memory channels per controller and enhancing the QPI bus. What is more telling with Intel is that clock speeds have not just stalled but decreased with the top end chips because Intel is unable to deliver high core counts with high frequencies. All models still deliver 2 threads per core, per core L3 cache has modestly increased with Nehalem at 2 MB to 2.5 MB.
The data speaks for itself. Intel has perpetrated this Great Charade convincing customers they have receive increased performance with each successive processor release when in fact they are buying flat performance, subsidizing Intel while increasing software licensing / maintenance costs.