Intel’s whole CPU range has undergone a refresh over the last 12 months. We report on the performance of their new processors and their suitability for audio.
Back in SOS July 2015 (http://sosm.ag/cpu-tests-0715), I wrote up the results of my tests of all the latest Intel and AMD CPUs for audio use, and explained a little about why certain processors might make better choices than others. Since then, it’s been a busy year-and-a-bit in the CPU world. Intel’s whole range has undergone a refresh over the last 12 months and there are rumours that AMD will have launched their latest range by the time you read this (though I can’t guarantee that — these things do have a tendency to slip!). In this article, then, I’ll report on the performance of the new Intel range. If you’ve not read last July’s article, I recommend that you do, as it gives lots of useful background that’s relevant, as well as demonstrating the performance of older systems alongside last year’s latest and greatest, and if you’re upgrading from an older system it will give you an idea of the sort of performance increases you can expect.
Before we get into the test results, a little context is helpful — without it, the results might appear confusing. Over the last few years, Intel have ensured that the professional-level solutions are based around established designs: this ‘workstation’ level hardware remains a complete architecture cycle behind the mid-range models. By handling the rollout this way they can ensure that any issues with new implementations of features that may arise within a chipset are caught and dealt with before they find themselves disrupting more critical workstations. Thus, while Skylake is Intel’s newest CPU architecture, at the time of writing, it’s only found in their mid-range CPUs. Broadwell-E and Broadwell-EP are, respectively, the current ‘enthusiast’ and ‘workstation’ products. The Skylake range maintains the same 14nm construction as its Broadwell predecessor (discussed last time), but offers architecture optimisations that deliver greater performance and reduce overall power consumption — remember, lower power consumption means less overheating, which translates to less fan noise.
Once again, I’ve conducted these tests using DAWBench (http://dawbench.com) although this time I’ve used a new beta update, due to some of the new processors outperforming the limits of the older version! I’ve also moved over to Windows 10 as the new testing platform. While that changes things a little compared with the previous test, this is the OS that will be used for most new systems, so it made sense to do this. Again, the test essentially stacks instances of Cockos Reaper’s ReaXComp multiband compressor plug-in, at a range of different buffer settings. All the results are displayed on the main chart, and what follows is much more detail about what you can see there at a glance.
We start the chart with the new Skylake-based processors, and specifically the highest-end i5 processor: the four-core i5 6600K. For those working mostly outside the box, but who find themselves recording and carrying out the minimum of editing and processing inside the system, the i5 should prove sufficient, and it tends to be where we at Scan start looking when we consider what CPU to base a system around.
Following it is the i7 6700HQ, which is this generation’s laptop CPU performance choice, offering four cores with additional hyper-threading, as per the rest of the i7 range. Often, laptop CPUs tend to lag behind their desktop processor counterparts, but the i7 6700HQ puts in a strong performance, and should prove a good choice for anyone working on the road or in multiple locations. Incidentally, the i7 6700HQ results closely mirror the low-power 35W TDP i7 6700TE desktop chip variant, which stands out as one of the highest-performing models for anyone wishing to assemble a totally passive system — that is, one with no moving parts such as CPU fans, PSU fans, hard disk drives and so on.
The consumer-level flagship and the last of the Skylake options is the i7 6700K. This is a popular choice for smaller project studios, as it offer strong enough performance to keep many in-the-box users happy, but keeps the overall system build cost down, by giving the builder access to the cheaper, more consumer-oriented boards and parts.
The i7 5820K and i7 5960X, six- and eight-core models respectively, are from the last generation of enthusiast-level X99 chipset CPUs. Whilst the Haswell E architecture they’re based around is two full generations behind that of the current mid-range Skylake chips already discussed, they certainly remain strong performers. Indeed, the refinements in the Broadwell E models are evident, and the newer chips can be fitted into the older motherboards giving an easy upgrade path, although the costs in transitioning may mean they offer poor value. There may be some debate to be had for power-hungry users looking to improve on the i7 5820K to one of the newer eight- or 10-core products, but I suspect owners of the last-generation i7 5960X will find it a little harder to justify the leap.
The dual-CPU W2620 V4 and single-CPU W2687 V4 listings are our two Broadwell EP Xeon setups. The eight-core W2620 V4 currently costs roughly the same as the i7 6800K, with a pair costing a little less than an i7 6900K, and the performance here showing to be only on par with a single i7 6800K. The W2687 V4, on the other hand, contains 12 cores but lags slightly behind the i7 6900K, which has fewer (eight) but faster cores. Note that when comparing CPUs, you must consider clock speed as well as the number of cores. In this instance, each of the enthusiast-class CPUs run with higher per-core clock speeds, thus allowing for more efficient processing. Whilst this isn’t a benefit for every type of computer processing task, it remains a strong one for real-time audio processing.
The first of our Broadwell E models is the aforementioned six-core i7 6800K, which is the direct replacement for the i7 5820K. Like its predecessor, it’s joined in the range by another six-core CPU, with only marginally faster core speeds but costing around 40-50 percent more. The key difference between the two variants in both cases is that the more costly chip has the same total of 40 PCIe lanes that can be found in the rest of the enthusiast range, whereas the better value option is restricted to only 28 connections.
PCIe lanes are used to give hardware in the system a quick path to the CPU for devices that require a lot of bandwidth. For instance graphics cards, which tend to be designed as 16x devices — the 16x denotes that a device uses 16 PCIe lanes. With only 28 connections to play with, users wishing to fit multiple video cards may find themselves running short of lanes to work with, thus restricting performance. It’s for this reason that video-editing systems using this class of chip tend to start around the i7 6900K CPU level, whereas the less demanding hardware found in audio-only systems means the cheaper option can be used, with very little trade-off in terms of performance. High-speed M.2 or AIC-based storage products, which are 4x in design, can be attached to this bus, but other cards are most likely to be 1x types, so most users are likely to run out of physical slots long before they run out of PCIe resources.
The i7 6900K can be viewed as a direct replacement for the previous generation’s flagship i7 5960X. It enters the market at roughly the same price and is an obvious choice for anyone who was previously considering the older chip.
This leaves us with the new top-of-the range 10-core option, the i7 6950X. This chip has broken Intel’s consumer-chip performance and pricing records. It offers the extra performance for those who require it, but at a cost: it’s priced roughly 55 percent higher than an i7 6900K, so the better value for money choice is clearly the cheaper chip!
For the most part, then, the chart looks largely as you’d expect it would, with the newer, greater core-count CPUs outperforming their older equivalents. However, the Xeon Broadwell EP results tell a whole different story, and it’s one that’s worth delving into deeper.
The ASIO buffer is important, as it’s the point at which data is collected before being passed through to the CPU. The buffer is designed to capture and release the data at set intervals, and the CPU only has until the next cycle is released to carry out all the processing required on the data. If it’s unable to do so, the likely result is audio drop-outs, due to data being discarded. The extra strain that lowering the ASIO buffer (in order to improve latency) places on the CPU is related — the overhead of data management takes its toll as the bursts of data are intensified.
With dual-CPU systems there is an extra level of complication, as each CPU has it’s own selection of memory banks, but can still access the memory space owned by the other CPU. When a CPU requests data that is not being held locally, this causes a tiny delay that wouldn’t be experienced with a single-chip arrangement. For many purposes, this wouldn’t be a problem, as when a CPU can receive a constant flow of data it is perfectly effective: off-line video rendering is an example of how raw processing power, left alone to work through the data in its own time, can carry out the job very efficiently. With real-time audio processing, though, the burst-delivery nature of the data to be processed doesn’t play to the platform’s strengths. That’s not to say Xeon chips can’t offer a lot to the power user who can justify the expense, and the promise of workstation-grade stability continues to be attractive for those managing commercial studios, for instance.
The second Xeon entry on the chart, the single W2687 V4, however, highlights another problem: software support. Through the testing process, it has become apparent that the various different DAW applications place a (seemingly arbitrary) cap on the number of CPU core threads that can be addressed. Depending on your DAW software, this could be as few as 16 threads, although 24 or 32 threads appears more common. When you consider that the high-end i7 6950X has 10 cores (so 20 threads), this doesn’t seem problematic. But for anyone considering dual-Xeon setup and hoping to squeeze every last drop from it, it may be worth consulting with your DAW manufacturer to find out what, if any, core/thread limits apply to their software. (All in all, though, this seems another good reason to aim for fewer cores operating at higher speeds.)
It’s worth noting at this point that in our tests the dual-W2687 V4 setup (not shown on the chart) easily exceeded the i7 6950X scores at each buffer setting by over a 100 instances. However, the processors were being capped in that test at 32 threads (the host DAW software’s limit). Looking at how it handled in testing, it would appear that the dual-Xeon systems should give you rather better performance if you need other software to run alongside your DAW: that could run on the spare cores without denting your DAW performance, which would not be the case on the single-chip i7 CPUs.
Another consideration that has often, in the past, tipped buyers in the direction of Xeons is memory support. Anyone doing soundtrack work with large, complex sound libraries could often find that even a memory limit of 64GB (which the X99 chipset had at launch) may not have proven to be quite enough. Xeons, with their extended memory support, were popular options for this reason and with the most recent updates the current generation of Xeon boards can offer support for up to 1024GB of RAM. Whilst not quite in the same league, a large selection of X99-chipset-based boards have been issued BIOS updates over the past year that increase the board support to 128GB, which I’m sure will continue to keep the vast majority of users happy for quite a while to come.
With Intel’s cards on the table, and no more major releases until Skylake-E and Kabylake around the middle of 2017, the performance increases we’ve seen this time around are certainly welcome, but may not tempt anyone who has purchased in the last couple of years to upgrade just yet. The other main dot of interest on the horizon is AMD’s Zen platform, which is due out at the end of the year. While it’s not expected to topple Intel’s place at the top of the performance table, the architecture change has the potential to challenge for the mid-range title, and so should generate a lot of interest when it eventually launches.