Intel Haswell-EP Xeon 14 Core Review: E5-2695 V3 and E5-2697 V3
by Ian Cutress on November 20, 2014 10:00 AM ESTE5-2695 V3 and E5-2697 V3 Conclusion
Reviewing CPUs that differ only in CPU count or clock speed but are based on similar architectures is predominantly a point and click affair. The CPU with the higher single core frequency does well in response focused benchmarks, and those with more cores with enough memory bandwidth to support them. For Xeons, the price difference for more cores or more frequency never makes much sense if you calculate the difference in terms of power, but becomes more realistic when the time-to-complete for intensive workloads is taken into account. To throw a spanner into the mix, as Johan found in his Haswell-EP testing, Xeons like the E5-2667 v3 exist to take advantage of higher AVX clock speeds due to increased power headroom, accelerating mixed-AVX workloads more than a higher core-count model which has to reduce frequency to compensate. The E5-2667 v3 therefore costs an extra $1000+ over its nearest core counterpart.
The two CPUs we have reviewed today, the E5-2695 V3 and E5-2697 V3, fall directly into Intel’s 2.3-2.6 GHz product line for Xeon E5-2600 V3 CPUs. This comes across as the main segmentation for Intel’s binning process from 8-core to 18-core, with the main differentiator relating to core count and pricing. The difference in 300 MHz comes at the expense of 25W TDP on paper as well as $278 in the back pocket.
An interesting point to note about the 2.3-2.6 GHz stack is the turbo frequencies. Not all SKUs are made similar:
Buying a 2.3 GHz base frequency processor with more cores does several things aside from the price: raise the peak turbo frequency, extend the turbo frequencies across more cores and increase the power consumption. As a result, buying the next CPU up in the stack affords more than just a couple of extra cores – there is both single thread and multithread performance of interest. With this in mind, it might be worth examining the Xeon range from through a MHz lens next generation rather than a core viewpoint.
On the 2.6 GHz graph, the 8-core model starts with the higher turbo frequency and has a more regular decline, while the 10-core and 12-core are evenly matched until the 12-core reduces to a lower all-core turbo. The peak turbo frequency again lies with the parts with more cores, and thus ends up with more TDP and cost more to purchase.
I showed the 2.3 GHz graph to some of the other editors and they pointed out the obvious differentiator: the 16 core CPU has a significant MHz advantage from 1-8 core loading over the 14C variant. For any software with a variable threaded load, the 16 core would push the performance and be on par with the 18 core. The price difference between the 14 core model and the 16 core model is nearly $1000, making budgets and workloads being important factors in this decision. The prevalence of multithreaded code is server and workstation environments make the frequency difference extremely important, especially when it comes to the types of workloads that have frequent memory delays and accesses.
This is apt, as our next element of Xeon coverage will be on the 16-core E5-2698 V3. If we can obtain a sample of the E5-2699 V3 as well, it will complete the set.
Facebook
Twitter
RSS
44 Comments
View All Comments
bill.rookard - Thursday, November 20, 2014 - link
I don't think I'd want one. If you notice, a majority of the benchmarks which a normal person might find useful the i7-4790K came in highest and flat out won 12 of those benchmark tests. And of course it costs about 1/10th the price.MrSpadge - Thursday, November 20, 2014 - link
Since he's reading AT I don't think he's a normal "person" ;)dgingeri - Thursday, November 20, 2014 - link
you're right about that. I'm a systems admin in a test lab. I'd want it for my VM practice system. A Core i7 3930k isn't quite as nice to run a dozen VMs as I had hoped.cjcoats - Thursday, November 20, 2014 - link
About those Linux benchmarks:1) That "NPB, Fluid Dynamics" reports millions of operations _per thread_, which is deceptive as regards actual delivered "in your face" work performed per wall-clock time; the existing chart should be supplemented by one giving [ops per thread per sec]*[number of threads] -- i.e., total ops per sec, which is what I really care about.
2) For Linux benchmarks in general: what compiler and compile-flags? ...and is this "one binary for all the machines"? The performance can be greatly influenced by targeting the actual processor architecture ("-xHost" for Intel compilers, "-march=native -mtune=native" for Gnu. For the codes I use and the (SandyBridge or later) servers I use, targeting the architecture I'm running on typically may give as much as a 70% boost, which is nothing to sneer at when my run-times are measured in hours or even days.
mapesdhs - Friday, November 21, 2014 - link
Re your 2nd point, that's certainly true for th C-ray test. One can do all sorts of optimisationsto show huge performance gains which are not remotely realistic. Wouldn't surprise me if this
affects the other tests too.
Ian.
FriendlyUser - Thursday, November 20, 2014 - link
Will it play Assassin's Creed: Unity?anubis44 - Thursday, November 20, 2014 - link
You forgot to say: "BUT can it run (insert terribly optimized pig-of-a-game here) ?"Cravenmor - Thursday, November 20, 2014 - link
Enjoying my coffee...Laststop311 - Thursday, November 20, 2014 - link
These xeons are just crazy priced. I would be happy with an i7-5960k clocked at 4.2ghz to match my i7-980x clocked at 4.2ghz. But really even after 4 years the performance increase still does not justify spending close to 2000 on cpu + x99 mobo + 4x8GB ddr4. Since broadwell is just a tick broadwell-e probably wont justify the expense either so maybe after 6 years have passed with gulftown skylake-e will finally make the expense worth it. Even if not the extra features like pci-e connected storage, ddr4 and pci-e 4.0 will finally make it worth it. By then the ridiculous ddr4 ram prices should be in line with how ddr3 is priced now too.wallysb01 - Thursday, November 20, 2014 - link
You must not be doing much heavily threaded work then, because for those that are the total system performance delivered with say, a $7K workstation, has gone through the roof compared to westmere.For example, the X5660 was 6 cores at 2.8GHz and the E5-2650v3 is 10 cores at 2.3GHz, but actually operates at 2.5GHz (something westmere didn’t do with the difference in turbo binning), both for a little over $1000 bucks each. Then, if we add 5% performance per generation, that brings the 2.5GHz to 2.9 “westmere equivalent” GHz. Then the 10 core vs 6 core means the E5-2650v3 is giving you 70% more performance than the X5660 did.