The Bottom Line
Introduction, Specifications, and Pricing
The time has finally come, and at this early hour of the morning (6 AM if you are on the west coast of the USA) you will get to see what the hype behind AMD's Threadripper CPUs is all about. AMD's recent revelations, such as those seen from leaked CINEBENCH scores and early sanctioned benchmarks from Alienware Area 51 systems touting the new CPU, don't reveal much beyond synthetics, and many of you have waited for this moment, and it's finally here.
Today we put a ridiculous 16 core and 32 thread CPU through the ringer along with its 12 core and 24 thread variant. These CPUs utilize the same Zen core micro architecture as the Ryzen 3, 5, and 7 series consumer CPUs and AMD's EYPC lineup of server CPUs.
Without further hesitation, let's see what AMD's Ryzen Threadripper 1950X and 1920X can do!
Specifications
AMD's Threadripper 1950X, 1920X, and other Thread Rippler CPUs are made up of two CPU Core Complexes (CCX). Each of the CCXs inside the 1950X is configured with 4+4 cores, while the 1920X is 3+3. In the end, you get two silicon dies in a configuration such as (4+4) + (4+4). Each of the dies is configured horizontally from the other, and two dud dies sit next to the working dies to balance out pressure from heat sink coolers.
Both CPUs have a 180W TDP and 32MB of total L3 cache. L2 cache is 8MB in the 1950X and 6MB in the 1920X. The 1950X has a base clock of 3.4GHz while the 1920X has a base of 3.5GHz. Both CPUs have an all core boost frequency of 3.7GHz, a four core boost up to 4.0GHz, and a 4.2GHz XFR boost on four cores if thermal headroom is here.
Both CPUs also have 64 PCI-E lanes, four of which have to go to the X399 chipset, and the rest can go to a total of 7 devices, typically four GPUs and three M.2 slots. While hardware RAID of NVMe drives (typically M.2) isn't currently supported, there is software RAID. AMD really needs to get NVMe RAID working on X399, and I think they might do it in the future, as RAID on a chipset means that total bandwidth is bottlenecked at the CPU to chipset connection; however, that wouldn't be the case with Threadripper since the M.2 slots are routed directly to the CPU.
Pricing
The Ryzen Threadripper 1950X will cost $999, and the Ryzen Threadripper 1920X will cost $799.
Threadripper Internal Configurations
The good news here is that AMD claims that Threadripper CPUs are at the top 5% of all Ryzen CPU dies, and that should mean that they have the best silicon in the batch and should clock higher. Threadripper's internal uArch is based on the same Zen micro architecture as is used in everything from Ryzen 3 CPUs to EPYC server CPUs, with a few exceptions when it comes to encryption units and the like.
However, the difference lies on the outside of the dies, and now the two dies are connected to each other and system components such as DRAM. AMD has incorporated different operating modes into the Ryzen Master software, which allows users to decide whether or not to put the CPU into a gaming mode, change how the dies interact with DRAM, and even offers a legacy mode to make the CPU like their Ryzen 7 counterparts but with quad-channel memory.
As I stated before, although the CPU has four silicon dies, only two are actually active or contain transistors (according to AMD they are not rejected EPYC parts), and these each has two CPU Core Complexes (CCX). Inside each die, the two CCX are connected by AMD's Infinity Fabric, and then each die is connected to each other through Infinity Fabric. AMD's Infinity Fabric operates at 102.22GB/s bi-directionally die to die, with near memory latency at 78ns and 133ns for far memory. AMD has made it possible to access memory differently to take advantage of bandwidth or latency.
You can choose within Ryzen Master to allow the CPU to operate in UMA mode (Distributed in the application) or NUMA mode (Local mode in the application). Distributed mode is the default operating mode, and it actually allows you to engage another mode called "Legacy Mode," which disables one die and allows one die to use quad channel memory. When in UMA mode, you get very wide DRAM access, but latency is increased while bandwidth is increased. In NUMA mode, DRAM accessed is close to each die, greatly reducing latency.
AMD's own benchmarking is quite close to the results I have seen, with UMA mode providing high bandwidth and NUMA providing better latency but slightly lower bandwidth.
The CPU and Test Setup
The CPU
You have probably already seen the packaging and CPU and how to install the CPU, whether it be from my live stream on Facebook last week or through many other publications, but I decided to do a quick section on the package.
Our sample arrived in a Pelican case with a dud CPU encased in glass (not included in the final package), as well as two CPUs in their retail packaging with LEDs under them to illuminate the packaging. The normal packaging uses foam on the outside, and the CPU and its accessories sit inside.
Accessories include a case badge, a Ryzen sticker, an Asetek mounting bracket, and a hex tool to open and close the socket to recommended weights. The tool allows you to screw the socket shut with 13.3In-lb, 15Kg-cm, or 1.5Nm.
The CPU itself sits in a secure plastic box, and you have to unscrew it to unleash the power within. The CPU then sites in a secure area, and it does take some muscle to release the metal clip that keeps the plastic retention lid secured to the CPU holder.
Both the 1950X and 1920X have identical packaging, and the CPUs look identical except for the actual markings. The CPU comes in a plastic tray, and it's highly recommended that you never remove the CPU from the tray as it helps with correct mounting of the CPU. There are so many pins, and since they fit into a smaller area, they are a bit thinner than Intel's LGA sockets, so it's easier to damage them.
Test Setup
- CPU: AMD Ryzen Threadripper 1950X and 1920X
- Motherboard: ASRock X399 Taichi (BIOS 1.3) and ASUS X399 Zenith Extreme (BIOS 401)
- Cooler: AMD Wraith Stealth
- Memory: G.Skill TridentZ RGB 3200MHz 4x8GB
- Video Card: NVIDIA GeForce GTX 1080 Ti Founder's Edition - Buy from Amazon / Read our review
- Storage - Boot Drive: Samsung 950 Pro and 960 Pro
- Storage - USB Drive: Corsair Voyager GS 64GB - Buy from Amazon / Read our review
- Case: Corsair Obsidian 900D - Buy from Amazon / Read our review
- OS: Microsoft Windows 10 - Buy from Amazon
- Monitor: ASUS PA328 ProArt 32" 4K - Buy from Amazon
- Keyboard: Corsair K70 LUX - Buy from Amazon
- Mouse: Corsair M65 PRO RGB - Buy from Amazon / Read our review
- Headset: Corsair VOID RGB Wireless - Buy from Amazon / Read our review
ASRock's X399 Taichi motherboard was used in this review and performed admirably; it also produced a 4.1Ghz overclock on the 1920X. Our first X399 motherboard review will be on the X399 Taichi from ASRock, and you will see us use it for our GPU and Gaming articles from Anthony.
ASUS's X399 Zenith Extreme was the first motherboard we got, it was included in the press review kit, and it was used for a lot of the benchmarking as well.
Out of the Box Performance: CINEBENCH, wPrime, and AIDA64
All CPUs are tested with Out of the Box Performance, which means that no settings were altered, so basically default CPU and RAM speeds. AMD's default mode is UMA/Distributed/Creator operation mode, where the memory is wide, so I also tested the 1950X in Gamer's mode, which makes the memory local and disables one of the dies, so it makes it basically a 1800X (8c/16t) with quad channel memory.
I also tested both the 1920X and 1950X with "AMD Optimized," which means memory speed of 2666MHz as stated on AMD's official CPU specifications.
CINEBENCH R11.5 had to be removed from our benchmark results as it seems to be bugged with the new CPUs; the last set of tiles takes forever to render, but R15 has no issues. CINEBENCH is not very memory reactive, but we do see slightly bumps with higher memory speeds. The 1950X is almost a 1800X in Gamer Mode, which disables one die, so it runs 8 cores/16 threads with quad channel memory. We do see single thread performance is slightly better than the 1800X (perhaps because of the higher XFR), but it's still behind Intel's.
CINEBENCH and wPrime results display how powerful the new Threadripper CPUs are if you need more cores and threads, Threadripper is what you want. FPU tests show the same trend as before, and AIDA64 greatly reduces Skylake-X's highly inflated scores, so they are now more comparable to AMD's offerings.
AMD offers excellent 64-bit Integer IOPS, while Intel still holds onto the SP-FLOPs title. Memory bandwidth is excellent when we increase memory speeds on Thread Ripper to 2666MHz, as it also increases fabric speed. We can see how Gamer Mode greatly reduces memory latency from 117.9ns to 88.8ns on the 1920X, but we also see that memory bandwidth takes a little hit.
Out of the Box Performance: Handbrake Video Transcoding, ScienceMark, and SuperPI
All CPUs are tested with Out of the Box Performance, which means that no settings were altered, so basically default CPU and RAM speeds. AMD's default mode is UMA/Distributed/Creator operation mode, where the memory is wide, so I also tested the 1950X in Gamer's mode, which makes the memory local and disables one of the dies, so it makes it basically a 1800X (8c/16t) with quad channel memory.
I also tested both the 1920X and 1950X with "AMD Optimized," which means memory speed of 2666MHz as stated on AMD's official CPU specifications.
We would expect the 1950X and 1920X to greatly trump the 7900X since they have many more cores and threads in applications such as HandBrake. The truth is that it doesn't beat the 7900X when rendering changes a 720P movie's codec to something more compressed, but the 1950X does beat the 7900 when taking a 4K video to 1080P.
The reason is that HandBrake didn't utilize all the cores and threads, as each different HandBrake profile utilizes cores and threads differently, some take advantage of all cores and threads while some don't. ScienceMark surprises and Threadripper will beat Intel's similar offerings with ease in many cases for overall scores.
While Intel leads in cryptography and molecular dynamics, Threadripper has a big advantage in memory and other things such as BAS/FLOPs, putting it ahead in final score. SuperPI 32M shows how the latency penalty from multiple infinity fabric jumps effects single core performance when it's heavily dependent on memory speeds. We can see that AMD's Gamer Mode, which greatly improved memory latency, does help it do a lot better, as SuperPI is extremely latency dependent.
Out of the Box Synthetic Gaming Performance: UNIGINE and 3DMark
All CPUs are tested with Out of the Box Performance, which means that no settings were altered, so basically default CPU and RAM speeds. AMD's default mode is UMA/Distributed/Creator operation mode, where the memory is wide, so I also tested the 1950X in Gamer's mode, which makes the memory local and disables one of the dies, so it makes it basically a 1800X (8c/16t) with quad channel memory.
I also tested both the 1920X and 1950X with "AMD Optimized," which means memory speed of 2666MHz as stated on AMD's official CPU specifications.
I think this is what everyone is waiting to see. Does AMD's Threadripper somehow overcome Intel's offerings in regards to gaming? The truth is Threadripper can't beat out Ryzen 7 in gaming, even with the Gaming mode, where the 1950X or 1920X can emulate Ryzen 7 CPUs. It does maintain decent FPS, but it doesn't go above and beyond what we have seen from Ryzen 7 or competitor offerings.
However, there is a lot of hope, especially in synthetic benchmarks, and that hope comes in the form of memory overclocking and optimizations. When we increase memory speeds from 2133MHz (stock) to 2666Mhz (AMD Optimized Stock), we see huge jumps in certain games, allowing the 1950X to compete head to head against Intel's offerings.
CloudGate is an exception to this, AMD's offerings pretty much stay steady. UNIGINE shows bigger increases with memory speed increases, but Intel is still quite strong. UNIGINE seems to benefit from Gamer mode, and so do the synthetics to certain extents. However, these are just synthetic benchmarks, and now we will see real gaming performance.
Out of the Box Gaming Performance: Resident Evil, Tomb Raider, GTA:V, Ashes of the Singularity
All CPUs are tested with Out of the Box Performance, which means that no settings were altered, so basically default CPU and RAM speeds. AMD's default mode is UMA/Distributed/Creator operation mode, where the memory is wide, so I also tested the 1950X in Gamer's mode, which makes the memory local and disables one of the dies, so it makes it basically a 1800X (8c/16t) with quad channel memory.
I also tested both the 1920X and 1950X with "AMD Optimized," which means memory speed of 2666MHz as stated on AMD's official CPU specifications.
So here is where Threadripper's weakness becomes apparent. Many games still can only take advantage of many fewer cores than Threadripper offers, and they like high frequency and high per core performance, so it was no surprise with the added latency of the Infinity Fabric between two core clusters, that we would see lower performance than Ryzen 7 CPUs.
AMD's Gamer Mode does seem to help bring back gaming performance by disabling one complex and routing all memory to just one. Another thing that helps a lot is memory speed, as this affects the Infinity Fabric speed as well, and we see big increases in gaming performance with just memory speed boosts, and ours in these threads are actually quite minor (2133 to 2666MHz).
AMD has yet to optimize many of the titles such as Ashes of the Singularity (a fun game actually), but I assume we will see some of these numbers go up in the coming months as we did with the first round of Ryzen CPUs. If you are curious about the zeros on the 1950X Gamer Mode results for AoS, it's because the game produced an error in legacy mode.
Overclocking and Power Consumption
CPU Overclocking
Wow, much easier and better than I expected. For the first time I was able to get a Zen based CPU to 4.1Ghz on the first try, and while you might think those temperatures are high, the "Tdie" is a more meaningful measure of temperature, as the "Tctl" uses an offset. Both motherboards overclocked my memory to 3200MHz with ease; they did it so easily I was super impressed. I was able to get the 1950X to 4GHz on all cores while the 1920X could hit 4.1GHz.
I used 1.35v set for both CPUs, as the motherboards will auto set your VCore to 1.36v when you overclock, and I thought that was a bit high regarding thermals. Overclocking Threadripper is easier than more fun than overclocking Ryzen 3, 5, or 7 and I didn't expect that. The overclock on the 1950X resulted in one of the highest 4K rendering FPS I have seen, almost breaking 100FPS.
For the ASRock motherboard I set CPU Frequency and Voltage change to manual and set the CPU frequency and VCore, then I just enabled XMP, and I had a stable overclock.
On the ASUS motherboard, I set overclock mode to D.O.C.P. and then CPU Core ratio and then I manually set VCore. I was able to boot at 4.1GHz, but the VCore required to remain stable was too high, I assume it's the luck of the draw.
Power Consumption
The Ryzen Threadripper 1950X's power draw is roughly 220W from the CPU power connectors, meaning that the CPU is using roughly 210W (if not less) itself after VRM switching losses. It's below that of the Skylake-X parts I have tested, but very close.
The 1920X is the same story, but in Gamer Mode, it's significantly less, but not close to that of the 1800X, perhaps because of the internal IO and Infinity Fabric.
What's Hot, What's Not & Final Thoughts
This is where you can fast forward to the final section of the review, and get a quick recap and points on the AMD's Ryzen Threadripper 1950X and 1920X.
What's Hot
It Rips Threads for Dinner: When you need thread ripping power, look towards AMD's Threadripper CPUs. At $1000, you get 16 cores and 32 threads of processing power, while the competition only offers 10 cores and 20 threads. For a few hundred less, you lose 4 cores and 8 threads, but still get a whopping 12 cores and 24 threads. When applications can use all threads, and use them to their maximum, Threadripper is almost unstoppable. The same can be said if you are running multiple applications, each of which takes up a lot of cores/threads, but we haven't had enough time to test that as we didn't get samples until a little over a week ago.
Top 5% of Ryzen Dies and Overclocking: AMD claims that Threadripper CPUs use the top 5% of Ryzen dies, and I believe them. Typically, higher core count CPUs don't overclock high, and typically have lower clocks than lower core count counterparts. Not only do the 1950X and 1920X have some of the highest boost frequencies among AMD's Zen core lineup (includes Ryzen 3, 5, and 7), but they also overclock to 4GHz and above with ease. Memory overclocking is also super easy, and I am not sure if this is because of BIOS optimizations from X370, or the fact that maybe AMD also looks at IMC quality to support high-density modules, and it doesn't matter because it overclocks with ease.
Lower Level Customization: AMD realized that in a fixed mode, the Threadripper CPU could be greatly limited, and so in their Ryzen Master software, AMD added in two options. One of them allows for memory to be in distributed or localized modes to the cores, the former better for content creators while the latter should be better for gaming. Another mode allows for legacy operation, where the CPU turns off one die and routes all four memory channels to a single die, cutting core count in half, but making it easy for games to take advantage of the cores in a good way. AMD has a Creator Mode set as default, and a Gaming Mode next to it, so you can easily switch between the two.
AMD's Optimization Potential: We have seen first-hand how AMD has greatly improved their performance numbers in applications through optimizations. We have also seen AMD improve performance and overclocking potential through BIOS updates, and I am sure we will see the same with Threadripper.
What's Not
Gaming Performance: I have had friends actually think AMD has marketed Threadripper as a gaming CPU, and that is just not the case. Threadripper is not a gaming CPU; it's made for the HEDT market, which mainly consists of content creators or those who multitask without concern for system resources. While Threadripper can play games, and with enough FPS for smooth game play in most games, it still could use some optimizations.
Final Thoughts
What's there to say? The benchmarks have said it all; Threadripper is excellent when you have many threads that need to be processed. Whether or not Threadripper is for you depends on the applications and use-cases you will use the CPU most for. Do you have a program that uses many threads and uses them to their maximum? Then there is probably no better option at this time.
Do you play a lot of games and do little else? Then perhaps you should look elsewhere. Do you mix a lot of single threaded, multithreaded, and applications that like frequency? Then perhaps you should do more studying, and it might be 50/50. We need more time to explore this use case. Do you multitask or single task? All these questions should be in a flow chart that I hope to produce in the coming weeks ahead.
Threadripper looks like it would be excellent for heavy multitasking, where you might be playing a game at the same time you stream the gameplay and do other things in the background, but with limited test time, we weren't able to explore that option (we will soon). Threadripper also has a ton of potential if AMD can do for Threadripper what they did for Ryzen. AMD also offers you options to improve Threadripper's gaming performance, and while some of those are built into software, you can always get a free boost by grabbing a 3200MHz memory kit and enabling auto memory overclocking to 3200MHz.
I was impressed with the way AMD can drop so many cores and threads on the market at relatively affordable rates. If you need threads, then you need The Thread Ripper.