Intel Core i9-7960X review: It beats Threadripper, but for a price

Mark Walton

Threadripper platform, one thing is clear: Intel once again has the fastest slice of silicon on the market. With 16 cores and 32 threads, matching AMD’s flagship Threadripper 1950X, the i9-7960X is an unashamedly over-the-top processor that breaks benchmarking records and powers through heavy production tasks.

But a processor is more than its raw number crunching prowess. Threadripper raised the bar for HEDT with the rich, consumer-friendly X399 platform, which offers a full set of features without spurious lockouts. More importantly, AMD doesn’t charge through the nose for it. The Threadripper 1950X features 16C/32T and costs £950/$1,000. Intel’s Core i9-7900X offers just 10C/20T for the same price. With the exception of gaming, the 1950X is a much more powerful processor.

Unfortunately, despite the strong competition, Intel isn’t yet willing to compete on price. The i9-7960X costs a whopping $1,700/£1,700—and while it might be faster, it certainly isn’t £700 faster. That’s not to mention that Intel continues to use a weak thermal material to mount its CPU heat spreaders, instead of the superior solder that AMD uses. It makes the i9-7960X a bear of a chip to overclock and noisy at stock without suffering serious thermal issues.

Ultimately, the i9-7960X raises the same question as the i9-7900X: Are you willing to pay for the best performing silicon on the market? Or is Threadripper, which offers most of the performance at a fraction of the price, good enough?

Skylake-X Redux

Intel’s X299 platform, to which the i9-7960X belongs, launched with more of a whimper than a bang. The CPU range, which starts with the questionably useful £240, quad-core i5-7640X and is topped by the $2000, 18C/36T i9-7980XE, is a confused mess of different CPU architectures and platform features.

The cheapest quad-core i5-7640X and i7-7740X chips (the latter of which does at least include hyper-threading), only feature 16 PCIe lanes and dual-channel memory, thanks to being based on the the same mainstream Kaby Lake architecture as the 7700K and 7600K. Even more expensive eight-core chips like the i7-7820X only feature 28 PCIe lanes.

Specs at a glance Intel Core i9-7980XE Intel Core i9-7960X Intel Core i9-7940X Intel Core i9-7920X Intel Core i9-7900X Intel Core i7-7820X Intel Core i7-7800X Intel Core i7-7740X Intel Core i5-7640X
Architecture Skylake-X Skylake-X Skylake-X Skylake-X Skylake-X Skylake-X Skylake-X Kaby Lake-X Kaby Lake-X
Cores/Threads 18/36 16/32 14/28 12/24 10/20 8/16 6/12 4/8 4/4
Base Clock 2.6GHz 2.8GHz 3.1GHz 2.9GHz 3.3GHz 3.6GHz 3.5GHz 4.3GHz 4.0GHz
Turbo Boost 2.0 4.2GHz 4.2GHz 4.3GHz 4.3GHz 4.3GHz 4.3GHz 4.0GHz 4.5GHz 4.2GHz
Turbo Boost 3.0 4.4GHz 4.4GHz 4.4GHz 4.4GHz 4.5GHz 4.5GHz N/A N/A N/A
All Core Clock Speed 3.4GHz 3.6GHz 3.8GHz 3.8GHz 4.0GHz 4.0GHz 4.0GHz 4.5GHz 4.0GHz
L3 Cache 24.75MB 22MB 19.25MB 16.5MB 13.75MB 11MB 8.25MB 8MB 6MB
PCIe Lanes 44 44 44 44 44 28 28 16 16
Memory Support Quad Channel DDR4-2666 Quad Channel DDR4-2666 Quad Channel DDR4-2666 Quad Channel DDR4-2666 Quad Channel DDR4-2666 Quad Channel DDR4-2666 Quad Channel DDR4-2666 Dual Channel DDR4-2666 Dual Channel DDR4-2666
TDP 165W 165W 165W 140W 140W 140W 140W 112W 112W
US Price $2000 $1700 $1400 $1200 $1000 $599 $389 $339 $242
UK Price ~£1950 ~£1650 ~£1350 £1078 £900 £550 £360 £315 £240

This isn’t a problem for the i9-7960X, which features the full complement of 44 PCIe lanes (and can thus take advantage of the X299 platform’s multitude of memory slots, PCIe slots, and I/O). But even that can’t match the 60 offered by AMD’s Threadripper, which allows for some serious three- and four-way graphics card setups with room to spare for PCIe NVMe storage. You even get the full complement of 60 PCIe lanes on the £500/$550 8C/16T 1900X.

Still, 44 PCIe lanes remains a substantial uplift over the paltry 16 offered by Kaby Lake. There’s also official support for quad-channel DDR4-2666 memory (with most motherboards offering support for higher speeds up to 4000MHz), and naturally every X299 CPU is unlocked for overclocking. Unlike Threadripper, however, ECC support remains a Xeon-only feature.

Where the i9-7960X differs from Intel’s cheaper X299 chips is with its hefty 165W TDP, and lower clock speeds. That additional 20W over the 10C/20T i9-7900X might not sound like much, but it makes a substantial difference to power consumption and heat, particularly when overclocking.

Left to right: Broadwell-E, Skylake-X, Kaby Lake-X, Skylake.
Enlarge / Left to right: Broadwell-E, Skylake-X, Kaby Lake-X, Skylake.
Mark Walton

Meanwhile, i9-7960X’s base clock of 2.8GHz is lower than other X299 chips (bar the i9-7980XE), as is the all-core boost, which reaches just 3.6GHz. AMD’s Threadripper struggles to get much past that when all 16 cores are under load, too, but it’s notable that Intel doesn’t have a clock speed advantage over AMD, at least without overclocking.

Intel Turbo Boost Max 3.0 returns to help pick up the pace when fewer cores are needed. Turbo Boost Max 3.0 picks out two favoured cores—those deemed to have the best thermal and voltage characteristics—and uses them to speed up single-threaded workloads by around 100MHz to 200MHz (a similar approach is taken by AMD’s XFR boost). Unlike with Broadwell-E, support for Intel Turbo Boost Max 3.0 is baked into Windows 10, negating the need to manually download drivers.

Intel charges for certain VROC features.
Enlarge / Intel charges for certain VROC features.
Mark Walton

It’s also worth noting that Intel’s Virtual Raid On CPU (VROC) feature—which allows you to link several M.2 NVMe SSDs in a bootable virtual RAID either via on-board M.2 slots, or via an expansion card—remains locked to RAID 0 out of the box. Those wanting to run other types of RAID need to purcahse a small VROC dongle for around £100/$100, which sits in its own slot on the motherboard.

While the market for such a setup is small, that you have to pay extra for it on an already expensive platform is price gouging of the highest order. Threadripper didn’t support bootable NVMe RAID at launch, but an update that enables RAID 0, 1, and 10 for up to 10 drives is now availiable. Best of all, it’s free.

MCM or monolithic?

Like the i9-7900X before it, the i9-7960X is based on the same 14nm FinFET Skylake-SP architecture of Intel’s Xeon server and workstation chips, which feature a new AVX-512 instruction set (up from 256-bit-wide AVX) and a new cache hierarchy. Intel has also dramatically redesigned the way each core exchanges data with another, introducing a mesh topology.

I won’t get into all the details of the tech here (check out the i9-7900X review for a deeper dive), but compared with the old ring bus method of exchanging data between each CPU core, the mesh topology promises to be much more efficient. The side effect is that some multithreaded software optimised for a ring bus may perform slightly slower on Skylake-X, but these applications are few and far between.

Threadripper's MCM design results in a huge CPU.
Enlarge / Threadripper’s MCM design results in a huge CPU.
Mark Walton

What’s more interesting is how Intel has used this mesh topology to construct its multicore CPUs compared to AMD. Intel uses its mesh topology to create a single, monolithic die that contains all 16 cores. Theoretically, since all the cores are on the same die—and because Intel can run the mesh at a solid clock speed, regardless of memory timings—data exchanged between each core is quick and consistent.

The drawback is scalability. The more cores Intel crams into a single die, the larger it becomes and the more difficult it can be to produce at scale. By contrast, AMD has opted for a multichip module (MCM) design for Threadripper. Theadripper is essentially two eight-core Ryzen dies (which are actually just two four-core dies) thrust together onto the same CPU package, and linked together via AMD’s Infinity Fabric tech.

While Infinity Fabric does rely on fast DDR4 memory clocks to function at its best, the benefit of the MCM design is that it’s easy for AMD to take an existing architecture (Zen) and scale it up. Zen even powers Epyc, the server CPU with 32 cores. While Threadripper’s MCM design doesn’t make it run any cooler, AMD’s use of solder helps transfer heat away from the die more efficiently. By contrast, the i7-7820X takes heat, and power consumption, to worrying new heights.

Listing image by Mark Walton

Overclocking: Not for the faint hearted

Once again, Intel has opted to us a standard thermal interface material to join the CPU die to the metal heat spreader, instead of a more efficient indium-tin solder. Broadwell-E, the predecessor to Skylake-X, as well as AMD’s Threadripper and Ryzen processors all use solder. Intel hasn’t explained why it made the switch, although there’s plenty of speculation surrounding the premium attached to using solder, as well as solder cracking the die during extreme temperature changes (although the latter is very rare indeed).

What this means, when paired with a huge die and a TDP of 165W like in the i9-7960X, is users are saddled with CPU thermals that severely limit overclocking and require expensive cooling solutions at stock. Given the ability to overclock is much of the reason to purchase an X-series processors over a comparable Xeon, that’s quite a problem. Intel recommends liquid cooling for Skylake-X, even for operation at stock speeds, but I’d go as far as saying you need 280mm liquid cooler at a minimum, with a 360mm liquid cooler being ideal (custom loops not withstanding).

If all you want to do is run the i9-7960X leaving everything on “Auto,” CPU temperatures hover around 67°C under full load when paired with a 360mm liquid cooler like the BeQuiet Silent Loop. I measured a total system power draw of 520W running Rog RealBench with a GTX 1080 Ti. 67°C isn’t a bad temperature, but it’s higher than a stock Threadripper 1950X, which hits 64°C with a full system power draw of 451W.

Part of the problem is the X299 platform is a little over-zealous with voltage, even at stock speeds (VID is at 1.030v at stock). By how much varies from motherboard to motherboard, but you can claw back some power draw and temperature by reducing the amount of voltage going to the CPU, instead of increasing it like when overclocking.

It is possible to run the i9-7960X using “Auto” settings at 4.1GHz by simply increasing the multiplier. However, the motherboard opts to pump 1.206v into the CPU, resulting in temperatures peaking at 97°C and total system power draw hitting a whopping 687W. That’s far from a useable, everyday overclock.

I found the lowest stable voltage at 4.1GHz to be 1.116v. Again, this will vary from motherboard to motherboard and chip to chip, but the result is a more reasonable 83°C peak and a total system power draw of 600W.

Ultimately, 4.1GHz at 1.116v (with a -500MHz offset for AVX) is what I’d consider a reasonable overclock for most people, without getting into custom loops or delidding, which is the complex process of removing the heat spreader and replacing the thermal material. That’s a shame, because I’ve no doubt that had Intel used solder, higher overclocks would be possible (I got as high as 4.4GHz without crashes, but quickly encountered thermal throttling).

Performance

Threadripper System Specs
CPU AMD Threadripper 1950X/1920X
RAM 32GB quad-channel G-Skill RGB @ 3200MHz
HDD 512GB Samsung 960 Pro M.2 PCI-e 3.0 SSD
Motherboard Asus ROG Zenith Extreme
Power Supply Be Quiet! Dark Power Pro 11
Cooling Thermaltake Floe Ring RGB 360mm liquid cooler
GPU Nvidia GTX 1080 Ti

Still, 4.1GHz across all 16 cores is not be sniffed at, particularly when Threadripper 1950X struggles to even hit 4.0GHz (I could only push the Ars sample up 3.9GHz). However, it’s not a whitewash for the i9-7960X, even with the overclock—far from it.

Like Threadripper, to get a better idea of just how well certain tasks and real-world applications scale across multiple cores, I tested across a variety of synthetic benchmarks, production applications, and games.

This includes Blenchmark, a rendering benchmark for Blender 3D, the free 3D pipeline—modelling, rigging, animation, simulation, rendering, compositing, and motion tracking application. The benchmark renders a 3D image of a BMW across as many CPU (or GPU) cores as you give it access to. Coupled with the POV-Ray benchmark, this gives a good indication of how well the i9-7960X handles heavy production tasks.

X299 System Specs
CPU Intel Core i9-7820X
RAM Corsair Vengeance RGB DDR4 @ 3200MHz
HDD Corsair MP500 480GB M.2 SSD
Motherboard Asus ROG Strix X299 Gaming-E
Power Supply Corsair HX1200i
Cooling BeQuiet Silent Loop
GPU Nvidia GTX 1080 Ti

The Chromium compile test also makes a return. It measures how long the computer takes to build a single instance of Google’s Chromium Web browser using Google’s standard method, which uses GN and Ninja to perform the build. Naturally, there are certain flags and optimisations you can use to reduce the overall build time, but for the purposes of benchmarking the standard build highlights the potential performance benefits of scaling all the way up to 16 cores in real-world tasks.

Alongside the i9-7960X are benchmarking results for the 10C/20T i9-7900X, along with older generation HEDT CPUs like the 10C/20T i7-6950X and i7-5930K. Naturally, both Threadripper 1950X and 1920X (a 12C/24T CPU) are included, as well as AMD’s best mainstream CPU, the 1800X, and Intel’s i7-7700K (the latter remains the best option for pure gaming builds, thanks to its impressive single-core performance and high clock speeds).

Production and synthetics

In Cinebench R15—the go-to willy-waver for CPU makers—at stock speeds the i9-7960X turns in a score of 3170, which is just shy of nine percent more than the 1950X. Considering the i9-7960X costs nearly 80 percent more, that’s a poor result. Overclocking brings the i9-7960X score up to 3579, but even that is only a 12 percent increase over the 1950X’s overclocked score of 3180.

POV Ray, which performs complex ray-tracing algorithms to generate realistic lighting in a 3D scene, is another example of diminishing returns. The i9-7960X renders the test scene in 41.26 seconds, the 1950X in 44.75 seconds, an eight percent increase. With both chips overclocked, the increase shrinks to just two percent.

Across the board, it’s hard to find a particular production task where the i9-7960X makes sense over the 1950X. Blenchmark is 18 percent quicker at stock, while the Chromium compile is 13 percent quicker, thanks in part to Intel’s superior IPC performance over Threadripper. Even in cases where time really is money, it’s hard to see the appeal of a chip that costs so much more, though.

If all you care about is getting things done as quickly as possible (or you’re into breaking benchmarks), cost be damned, then sure, the i9-7960X hits the mark. The only thing faster than the i9-7960X is the i9-7980XE, and that’s a £1,950 CPU.

Gaming

Buying a 16C/32T CPU just to play games isn’t the wisest way to spend your money. However, if you’re an enthusiast that simply wants the best silicon on the market, or you’re one of Intel’s so-called “Megataskers,” then gaming performance remains an important metric. Plus, the days where extra cores made little difference to frame rate are almost at an end.

The differences are subtle, though, and mostly revolve around minimum frame rates. The benchmarks in the graphs above show the 99th percentile score; that is, the minimum frame rate you can expect to see 99 percent of the time with games run at high or ultra settings. Compared to the i7-7700K, there’s a nice uplift to the minimum rate of around nine percent at 1080p in Hitman under DX12, and around 12 percent on average. There are higher gains in Rise of the Tomb Raider at 1080p, where there’s a 21 percent increase to the minimum frame rate.

These aren’t Earth-shattering numbers by any means, but if you’re interested in high-frame-rate gaming without compromising on visual fidelity, the extra cores make a difference. Increase the resolution or switch over to DX11, however, and the gains either decrease, or disappear completely. The 7700K’s superior clock speed continues to be an advantage, particularly in older games like Grand Theft Auto V.

Like Ryzen before it, Threadripper suffers when it comes to 1080p gaming, although the differences between it and the i9-7960X are slight compared to the mainstream chips. The i9-7960X boasts around a three to five percent increase in frame rate over the 1950X, depending on the game. By comparison, there’s a much bigger difference between the i9-7900X and 1950X, which goes to show just how important raw clock speed is when gaming.

The clock speed advantage that Intel usually enjoys isn’t there with the 16-core chip. Even its superior IPC isn’t making up all of the difference. And, thanks to thermal problems, there’s a limit to how much you can make up the difference with overclocking, without going as far as delidding.

Verdict

That Intel finds itself in a place where its flagship processors just barely scrape past the competition is an astonishing turn of events. It might well continue to offer the best gaming performance in the mainstream, where gaming continues to be the key driver for upgrades, but in the HEDT market it’s hard to see why anyone should pick up an Intel system over an AMD one.

Perhaps the reason for this mess is down to physics. Intel continues to pursue Moore’s Law, but since the release of Skylake, IPC improvements and transistor densities in Intel CPUs have stalled. AMD hasn’t fared much better either (compared to Intel at least) with Ryzen only matching Intel’s older Broadwell-E architecture. The difference is that AMD bet on a massively multicore future for consumers and prosumers, designing an architecture with dramatic scaling in mind, when Intel was still umming and ahhing whether to bring six-core chips to the mainstream. Intel got complacent, and the confused X299 line-up is the result.

Ultimately, the i9-7960X is a powerful CPU made for those those willing to custom water cool, or brave enough to de-lid. It’s a CPU for enthusiasts and content creators with more money than sense chasing Reddit bragging rights, even when better value options exist. It’s a CPU that, despite its strengths, I struggle to recommend to anyone but the most hardened of Intel fans.

The good

  • The fastest slice of silicon going
  • Good gaming performance
  • Improved memory support
  • Modern complement of I/O

The bad

  • Not that much faster than Threadripper 1950X
  • Fewer PCIe lanes than the competition
  • Power hungry
  • Continual use of TIM instead of solder
  • High clock speeds and quiet systems out of the question without custom liquid cooling or delidding

The ugly

  • There’s little justification for the price tag. Buy a Threadripper 1950X instead.

This post originated on Ars Technica UK

Powered by WPeMatico