The launch of Threadripper marks a move closer to AMD’s starting point for the Zen architecture. Contrary to popular belief, AMD did not start its plans with desktop Ryzen and then glue modules together until Epyc was created; no, instead, the company started with an MCM CPU more similar to Epyc, then worked its way down to Ryzen desktop CPUs. Threadripper is the fruition of this MCM design on the HEDT side, and benefits from months of maturation for both the platform and AMD’s support teams. Ryzen was rushed in its weeks leading to launch, which showed in both communication clarity and platform support in the early days. Finally, as things smoothed-over and AMD resolved many of its communication and platform issues, Threadripper became advantaged in its receipt of these improvements.
“Everything we learned with AM4 went into Threadripper,” one of AMD’s representatives told us, and that became clear as we continued to work on the platform. During the test process for Threadripper, work felt considerably more streamlined and remarkably free of the validation issues that had once plagued Ryzen. The fact that we were able to instantly boot to 3200MHz (and 3600MHz) memory gave hope that Threadripper would, in fact, be the benefactor of Ryzen’s learning pains.
Threadripper will ship in three immediate SKUs:
- 16C/32T 1950X
- 12C/24T 1920X
- 8C/16T 1900X
Respectively, these units are targeted at price-points of $1000, $800, and $550, making them direct competitors to Intel’s new Skylake-X family of CPUs. The i9-7900X would be the flagship – for now, anyway – that’s being more heavily challenged by AMD’s Threadripper HEDT CPUs. Today's review looks at the AMD Threadripper 1950X and 1920X CPUs in livestreaming benchmarks, Blender, Premiere, power consumption, temperatures, gaming, and more.
A Note on Overclocking & Memory Access
AMD’s done well with XFR on the Threadripper CPUs. Because of this multi-die configuration, Threadripper is able to engage XFR on up to four cores simultaneously, boosting +200MHz over base. The result is a possible 4.2GHz speed on 4C for Threadripper, which consequently proves itself most beneficial in lightly threaded applications. A game might be a good example of such an application. Because we were unable to overclock beyond 4.2GHz XFR (maxing out at 4.0GHz all-core OC), some overclocking numbers will be lower than the stock configuration. This is true for any application which better makes use of XFR than our overridden all-core clock configuration.
Per-core overclocking is not available on Threadripper, and won’t be. Overclocking can be done in core pairs, with OC expectations similar to Ryzen. AMD noted that the Threadripper CPUs consist of the top 5% of presorted Ryzen dies (though later changed this number to 2%, depending on who was talking). The performance headroom is leveraged already in Threadripper’s relatively high frequency, particularly considering core count, though there’s some room for play for overclockers. An XOC team at the AMD event was able to push the 1950X to 5.2GHz on 1.6V, using LN2 as the cooling solution. Expect in the 4.0 to 4.2GHz range for more traditional modes of cooling.
There are also two memory access modes, here: Distributed and local, also known as UMA (Uniform Memory Access) and NUMA (Non-Uniform Memory Access). Distributed is the default mode and tends to work best for content creation tasks, as UMA isn’t restricted to one die and allows the scheduler to do whatever it wants. NUMA tends to be better in some games, but restricts a workload to 1 die and the memory attached to that one die. Creative applications, meanwhile, minimize thread sharing and synchronization between cores, so they’re not that latency intensive. Gaming applications care a lot about latency and synchronization; you might be synchronizing the physics and game logic threads with AI, for instances, and so NUMA tends to be better in these use cases. Threadripper hosts two separate memory controllers, working to support quad-channel memory on the board. For reference, we were able to get 3200MHz working with no effort, and managed 3600MHz with some trivial amount of effort (both GSkill kits). The 3600MHz kit was used for stream benchmarking on both the 7900X and 1950X.
As an aside, cores 0-15 are located on the first die of the CPU, with cores 16-31 on the second die. Each Threadripper CPU consists of four total pieces of silicon under the IHS, two of which are active. AMD tells us that the same two dies are active each time in Threadripper, something that we later looked into with thermal imaging of our 1950X & 1920X. We do not have a large enough sample size to conclude that AMD’s statement is accurate, but hopefully some other media members have also checked their CPUs. For our units, we determined that, orange pull tab oriented down, the top-left and bottom-right corners of the CPU are the active dies. The other two units appear to be deactivated or dead, depending. AMD claims that these are silicon substrate interposers that are instituted for mechanical mounting reasons, though it is still entirely possible that they’re actually just Epyc CPUs which have been pushed as Threadripper.
Learn more about our stream testing here (part 1) and here (part 2). The rest is discussed on the next page, where we define differences for today.
Game settings were manually controlled for the DUT. All games were run at presets defined in their respective charts. All other game settings are defined in respective game benchmarks, which we publish separately from GPU and CPU reviews.
Average FPS, 1% low, and 0.1% low times are measured. We do not report maximum or minimum FPS results as we consider these numbers to be pure outliers. Instead, we take an average of the lowest 1% of results (1% low) to show real-world, noticeable dips; we then take an average of the lowest 0.1% of results for severe spikes. GN originally coined the phrases “1% LOWs” and “0.1% LOWs” for these metrics.
Note: fan and pump settings are configured on a per-test basis.
- ASUS Zenith Extreme X399
- 32GB 3200MHz GSkill RGB Memory
- EXCEPT: Streaming benchmarks conducted with 3600MHz GSkill Trident Z Black memory (this is also used on the X299 platform for streaming benchmarks)
- Gigabyte Gaming 9 X299
- 32GB 3200MHz GSkill Trident Z Black Memory
- EXCEPT: Streaming benchmarks conducted with 3600MHz GSkill Trident Z Black
EVGA Supernova 750 G2L 80+ Gold
HyperX Savage 32GB 2400MHz (4x8GB)
Corsair Force LE 240GB SSD
Open Air Test Bench
EVGA GeForce GTX 1080 FTW
970 (RD9x0) Platform:
-ASUS 970 PRO GAMING/AURA
-Phenom II X6 1055T (125W TDP)
-Phenom II X6 1090T
Core Components (Unchanging)
- NZXT 1200W Hale90v2
- For DDR4 platforms: Corsair Vengeance LPX 32GB 3200MHz
- For DDR3 platforms: HyperX Savage 32GB 2400MHz (note: only 2133MHz was supported on our SNB platform)
- Intel 730 480GB SSD
- Open Air Test Bench
- Cooler #1 (Air): Be Quiet! Dark Rock 3
- Cooler #2 (Cheap liquid): Asetek 570LC w/ Gentle Typhoon fan (this is the one we used for this particular article)
- Cooler #3 (High-end): Kraken X62
- Video Card: EVGA GTX 1080 FTW1
- Note: fan and pump settings are configured on a per-test basis.
- MSI Gaming Pro Carbon
- i7-7700K (x2) samples from motherboard vendors
- i5-7600K purchased by GN
- MSI Gaming M7
- i7-6700K retail
- Gigabyte Z97X G1 WIFI-BK
- MSI GD65 Z77
- i5 3570K
- ASUS Crosshair VI
Dx12 games are benchmarked using PresentMon onPresent, with further data analysis from GN-made tools.
- ASUS Deluxe Pro X58
BIOS 0305 was used on the Zenith Extreme.