Synthetic Benchmarks: Ivy Bridge-E vs. Haswell
Before our full-scale testing of the new Ivy Bridge-E processors we want to check out what performance per CPU core they can offer. We really feel some doubts about Intel’s decision to update its flagship CPU series using not the most advanced microarchitecture. The high number of computing cores may make up for that, yet not all applications can utilize six CPU cores, especially with Hyper-Threading enabled. Moreover, the Ivy Bridge-E series includes a quad-core model, Core i7-4820K, which has a slightly higher clock rate, a larger L3 cache and a quad-channel memory controller in comparison with the Core i7-4770K, the top-end Haswell-based LGA1150 model. Is it enough to justify the higher model number, especially as the LGA1150 processors not only feature a more progressive microarchitecture but also support the FMA/AVX2 instructions?
To answer this question and better understand the highs and lows of the Ivy Bridge-E design, we began by comparing the quad-core processors, Core i7-4820K and Core i7-4770K, at the same clock rate of 4.0 GHz. To make the picture complete, we added the previous-generation LGA2011 processor Core i7-3820.
Synthetic benchmarks are the best way to form your first impression about the peculiarities of CPU designs. We use SiSoftware Sandra 2013 SP6 for that purpose.
Frankly speaking, there’s nothing unexpected about these results. The Ivy Bridge-E design copies the familiar Ivy Bridge, transferring the microarchitecture to another platform. That’s why the Ivy Bridge-E is but slightly superior to its predecessor in terms of per-core performance. We can only see it with integer operations where the newer microarchitecture can improve performance by up to 10%.
It turns out that the LGA2011 platform accelerates more through higher clock rates rather than through internal improvements. The clock rates are not much higher, though, so the new LGA2011 CPUs are going to be just 5 to 10% faster than the older ones.
The Ivy Bridge-E is considerably inferior to the Haswell in terms of per-core performance. Even if we put aside the new AVX2 instructions which can improve performance of many popular algorithms, the quad-core Haswell is always ahead of the quad-core Ivy Bridge-E by 6 to 7%. It turns out that the LGA1150 platform offers more progressive processors and the new LGA2011 CPUs can only be superior by having more computing cores.
The larger L3 cache together with the high-performance 4-channel memory controller seem to be one of the advantages of the Ivy Bridge-E over the Haswell design but there are nuances that make this advantage questionable. First of all, the caching mechanism of the Ivy Bridge-E is slower than in the Haswell. The bandwidth tests make it clear.
The Haswell has higher L1 and L2 cache bandwidth, which shows up in practical applications. The Ivy Bridge-E and Haswell are comparable in terms of cache latency, though.
Secondly, the practical speed of the quad-channel memory controller is not much higher compared to the Haswell’s dual-channel controller:
The results you can see in the diagram refer to the peak bandwidth which is achieved at multithreaded memory access whereas everyday applications access memory in one thread, producing a different picture of practical bandwidth.
The Haswell’s memory controller copes better with typical desktop loads, even though has much lower theoretical bandwidth. The Ivy Bridge-E’s controller is optimized for multithreaded server applications but its four channels are not so good otherwise. This was typical of the Sandy Bridge-E CPUs and the new CPU generation has the same peculiarity.
Now that we know about the highs and lows of the Ivy Bridge-E design, it's time to benchmark it in real-life applications.