AMD Threadripper 3970X and 3960X Review: High-End Domination

In an incredible alignment of the silicon stars, AMD's third-gen Threadripper 3970X and 3960X launch today alongside Intel's new Cascade Lake-X Core i9-10980XE, giving us a look at the future of the high end desktop all in one go. 

Intel surely has its work cut out for it. Earlier this month, AMD showcased the advantages of its Zen 2 microarchitecture and 7nm process by cramming the 16-core/32-thread Ryzen 9 3950X into its mainstream platforms, doubling the core count of Intel's competing chips. That created a new tier of threaded performance for mainstream motherboards, one Intel simply can't match, and upset Intel's pricing structure for its high-end desktop lineup in the process. 

But that wasn't the full extent of the damage: Intel preemptively slashed the gen-on-gen pricing of its forthcoming Cascade Lake-X chips in half as it braced for the impact of AMD's Threadripper 3000. 

Now they're here: AMD's 32-core 64-thread Threadripper 3970X and 24-core 48-thread 3960X have landed in our labs with the same core counts as their predecessors, but instead of a headline achievement like a doubling of core counts, Threadripper's new architecture serves as the star of the show. The 'Castle Peak' design brings a new design that does away with many of the previous-gen Threadrippers' eccentricities, which equates to massive performance gains across the board. Add in support for PCIe 4.0, and the Threadripper 3000 series looks like a winner.

Of course, with much more real estate available in the massive sTRX4 socket on TRX40 motherboards (), we also fully expect AMD to bring the density advantages of its 7nm process to bear, and soon it will: The company confirmed it would bring a beastly 64-core 128-thread Threadripper 3990X with an amazing 288MB of total cache and 280W TDP to market in 2020. 

But while we await that ludicrously-appointed chip, we have plenty of excitement with the Threadripper 3970X and 3960X. These chips promise a new level of performance for prosumers and creators, and they deliver. 

We've been busy in our labs testing the processors through our standard test suite of games and applications, but also added an expansive set of workstation tests that expose the benefits of the TRX40 platform, like AMD's leading support of PCIe 4.0. We also have Intel's new Core i9-10980XE on the roster, along with the Ryzen 9 3950X, to give you a view of the entire high-end battlefield. 

Threadripper 3970X and 3960X Specifications and Pricing

AMD's Threadripper 3000 processors debut with much higher pricing than the previous-gen models: At $1,999, the 32-core 3970X debuts for $200 more than its predecessor, and AMD tacked on an extra $100 for the $1,399 24-core 3960X. AMD attributes the increased pricing to the new level of performance and access to Threadripper's leading-edge support for the PCIe 4.0 interface. 

AMD gears the Threadripper 3000 processors for content creators and prosumers, like 3D artists, filmmakers, and software developers. For professionals, any extra time spent completing a task equates to lost revenue, making the higher price a justifiable expense if the premium matches the performance.

You'll also need a new TRX40 motherboard to support the chips. We've tested a couple around the launch of these chips, and our current favorite is ASRock's TRX40 Taichi. Due to the higher pricing of Threadripper 3000 models and the requirement for a new motherboard, AMD is leaving the existing second-gen Threadripper processors on the market for the foreseeable future as a 'value HEDT' offering for content creators. You can regularly find the 32-core Threadripper 2990WX for  roughly $1,700 and the 24-core 2970WX for about $925, and we expect those prices will fall soon. 

The Threadripper 3970X weighs in with 32 cores and 64 threads that operate at a 3.7 GHz base and 4.5 GHz boost paired with 144 MB of total cache (128MB L3), while the Threadripper 3960X has 24 cores and 48 threads that run at a 3.8 GHz base and 4.5 GHz boost with 140 MB of total cache. 

Both chips expose 88 PCIe 4.0 lanes, but the TRX40 chipset consumes some of those lanes, leaving 64 exposed to the user. The PCIe 4.0 interface is a notable advantage of AMD's Ryzen desktop lineup, but it is more important in the HEDT space where high-speed storage and networking devices are more likely to find their way into powerful systems. 

Both processors support four channels of DDR4-3200 memory, but data transfer rates differ based on DIMM population. The chips support up to eight DIMMs of 32GB memory, totalling 256GB. The processors also support ECC memory, but implementation varies by motherboard. The Threadripper processors feature the same memory controllers as the Ryzen 3000 chips, so memory overclocking is a rather straightforward affair.

Intel's new Cascade Lake-X represents the company's flagship competitor on the high-end desktop, but it is only $979, which is less than half the price of the previous-gen Core i9-9980XE. That leaves AMD's Threadripper 3000 largely uncontested in the traditional HEDT space, though Intel does offer its Xeon W processors with the server-derived LGA 3647 socket. 

The Xeon W chips slot into a higher workstation tier, top out at 28 cores, and don't feature unlocked multipliers, meaning they aren't overclockable like HEDT processors. These processors require expensive motherboards and coolers, and also feature six-channel memory controllers. Given their workstation-class features, the Xeon W chips that compete directly (based on core count) against Threadripper 3000 retail for $3,349 and $4,499 for a 28-core chip, so they aren't suitable competitors. 

Intel also has its Xeon W-3175X, the lone overclockable processor on the LGA 3647 platform, for $3,000. This 28-core 56-thread processor drops into exotic motherboards that carry heart-stopping price tags, so it really isn't a direct competitor to Threadripper 3000, either.

AMD doesn't specifically market the Threadripper processors for workstation use, and the significantly lower pricing means Intel has largely ceded the top of the HEDT market to AMD, at least for now.

Third-gen Threadripper Architecture

The Threadripper 3000 series chips come with the TSMC 7nm process, which has density advantages that manifest as higher performance, better power efficiency, more cores, and more cache packed into a smaller die area than the first- and second-gen Threadripper models. 

Like the mainstream Ryzen parts, Threadripper 3000 comes packing AMD's Zen 2 microarchitecture that brings a notable IPC improvement, but AMD spreads the design across four eight-core 'core chiplet die' (CCD), as opposed to two with the mainstream chips. AMD ties the compute chiplets together via the Infinity Fabric to a large central 12nm I/O die that houses two 32x PCIe Gen4 controllers and two dual-channel DDR4 memory controllers. 

Each 7nm compute chiplet features ~3.9 billion transistors, while the 12nm I/O die has ~8.34 billion transistors, yielding a total of ~23.94 billion transistors in the Threadripper 3970X and 3960X spread over 712 square millimeters of silicon.

As we can see in the last image in the album below, the new design stands in contrast to the distributed design in the first- and second-gen Threadripper models, which had up to four die connected directly to each other. Unfortunately, AMD fused off the PCIe and memory controllers on two of those die, which led to increased latency when the die needed to access far memory banks and I/O devices. 

Image 1 of 4

Threadripper 3000's I/O die provides the compute chiplets with uniform access to the PCIe and memory controllers, thus reducing a layer of latency and ensuring a smooth quality of service. AMD also reduced the Infinity Fabric On-Package's (IFOP) power consumption by 27%, which the SoC now allocates to provide extra compute horsepower. 

Like all modern processors, AMD's 3000-series processors and dies come with a mix of faster and slower cores. However, as we discovered, AMD programs maximum boost frequencies on a per-core basis. That stands in contrast to the previous industry status-quo of delivering a uniform boost capability across all cores, but the technique extracts the maximum performance and power efficiency out of each die.

As a result, each core and die has different boost capabilities, so AMD chooses a premium die and places it in the CCD4 position inside the SoC. The company then uses an innovative mix of the Windows scheduler, drivers, and motherboard firmware to target the fastest two cores with lightly-threaded workloads, which allows it to execute at the highest possible frequency. 

The Windows scheduler, which interfaces with CPPC2 (aka 'preferred core), rotates the lightly-threaded work between the two fastest cores on CCD4 to reduce current and thermal density, which fosters more frequent and longer-duration boosts. AMD has weathered plenty of criticism for its early missteps when implementing the CPPC2 functionality, but has fixed the lion's share of the issues. As we'll show on the following page, our Threadripper chips had no issues reaching their rated boost speeds, and even exceeded the rating on rare occasions. 

TRX40 Motherboards

AMD's new TRX40 platform features a new sTRX4 socket that isn't backward compatible with existing Threadripper processors, and the 3000 series isn't backward compatible with X399 boards, either. The sockets are mechanically compatible, so you use a cooler designed for previous-gen Threadripper chips with the 3000-series models, and AMD says that any cooling solution capable of handling 1000- and 2000-series processors should suffice for the 3970X and 3960X. 

The sTRX4 socket features the same 4094 pins as the previous-gen sTR4 socket, but AMD changed the electrical layout of the new socket to accommodate the PCIe 4.0 interface. The company hasn't committed to compatibility with future Threadripper processors, but says the socket is designed with scalability in mind. That will obviously come in handy when the 64-core Threadripper 3990X lands next year. 

Image 1 of 5

The TRX40 chipset features 16 GB/s of throughput between the processor and the chipset, which comes courtesy of 8 PCIe 4.0 lanes. In contrast, Intel supports 4 GB/s of throughput over its DMI link through the PCIe 3.0 interface. TRX40's extra bandwidth will benefit multiple devices that hang off the TRX40 chipset, like large PCIe SSD arrays.

Threadripper 3000 also offers up to 133GB/s of concurrent bandwidth to connected devices, while Intel only offers 52 GB/s, which is another significant advantage. 

MSI, Gigabyte, ASUS and ASRock have all announced leading-edge TRX40 motherboards, and we expect more to follow in the coming months. 

MORE: Best CPUs

MORE: Intel & AMD Processor Hierarchy

MORE: All CPUs Content