My Threadripper 1900X Semi-formal Review - Part 1

Paratus

Lifer
Jun 4, 2004
16,614
13,297
146
Introduction__________________


GY1Ly22.jpg



I built a new machine back around Christmas. I had been looking to replace my current system:
  • CPU: OC’d i7920 (3.5Ghz)
  • RAM: 12GB DDR3 1333 9-9-9
  • MOBO: ASUS P6T Deluxe V2
  • GPU: MSI Lightening 7970 GHZ Ed.
  • Boot: Win10 on a 1TB Samsung 850 Evo
wiith something new. I was looking for a platform that would last me several years, good light office and production performance, decent gaming, and not completely break the bank. (Difficult with DDR4 and GPU prices).


As I’ve mentioned before, back in the day that OC’d i7920 on X58 was at or near the top for all 4 of those categories. Today you get to pick one:

  • Gaming: 8700K
  • Production: 7980XE
  • Overall Bang for the Buck: Ryzen 1700(until the 2X00 series release)
  • Best platform: X399

For my needs I settled on the least expensive AMD Threadripper model - 1900X and the X399 platform.

Since the 1900X wasn’t sampled to the tech press there aren’t a lot of reviews so I thought I could run some benches and show how the top of the line first gen Ryzen based 8C /16T chip performs for you guys. It also might be an interesting comparison with the 2700X as they have the same core counts and the top end frequencies are in the same ballpark.

Testbed Overview____________

  • CPU: AMD Threadripper 1900X (8C/16T, 3.8G, 180W, 20MB - $380-$550)
  • MOBO: MSI X399 Gaming Pro Carbon AC ($300)
  • COOLER: Enermax Liqtech TR4 360 ($130 360mm radiator with full coverage block)
  • RAM: G.Skill Trident Z 32GB 4X8GB DDR4-3200 14-14-14 ($480 - Samsung B-die)
  • GPU: Powercolor Red Devil RX Vega 56 8GB ($690)
  • SSD: Win10 Pro on a 1TB Samsung 960 Evo M.2 NVME ($450)
  • Case: Thermaltake View 31 TH RGB ($95 with 5X 140mm Thermaltake Riing+ RGB fans bought separate)
  • PSU: Corsair RM850i 850W modular ATX power supply ($160)
  • Monitor: Dell U2515H 25in 2560x1440 60hz IPS

The AMD Threadripper 1900X has the following specifications:

  • Cores / Threads: 8/16
  • Base Clock: 3.8GHz
  • Boost Clock: 4.0GHz
  • XFR: upto 4.2GHz
  • TDP: 180W
  • Vcore: 1.35V
  • Max Temp: 68C (stock)
  • PCIE 3.0 Lanes: 64 - ( 48 Lanes for Slots, 12 Lanes for NVME, 4 Lanes to Chipset)
  • RAM: Quad Channel DDR 4
  • Cache: 4MB L2, 16MB L3

Under the enormous heat spreader an AMD Epyc processor would have 4 separate chips made of 2, 4 core CCX’s for 32 total cores. The fully functional Threadripper 1950x has 2 non functional chips and 2 functional chips with 2, 4 core CCXs for a total of 16 cores.

27105334523l.jpg


The 1900X, a derivative of the 1950X has one functional 4 core CCX in each of the two functional chips for a total of 8 cores. This is a small but fundamental difference between the 1900X and the Ryzen R7s (1700, 1700X, 1800X and likely the new 2700, and 2700X). Those R7s are a single chip with two 4 core CCXs.

The benefit of this design for the 1900X over the R7’s is each chip provides 32 PCIE 3.0 Lanes and a dual channel DDR4 memory controller. This gives 64 PCIE Lanes and quad channel RAM across the two chips vs the single dual channel controller and 24PCIE lanes of the R7’s

The downside compared to the R7s is if data needs to be passed between CCXs the R7’s keeps it on the same chip while the 1900X has to go off chip and suffer a latency penalty. In addition some software may have UMA vs NUMA issues with the 1900X. It’s possible lower latency/faster RAM can mitigate this. Hopefully testing might shine a light on this.

The other downside compared to the 1950X and 1920X is missing 20MB of cache due to the loss of the other two CCXs.

Testing Methodology_________

I was curious to see how the 1900X behaves at stock under load and how the CPU responded to changes in memory speed and clock speed changes. To evaluate the 1900X I varied the CPU and RAM Speeds a follows:

CPU
  • 3.8GHz
  • Stock - 3.83GHz - 3.96GHz (varies by benchmark)
  • 4.15GHz
  • 4.2GHz - CPUz only due to stability

RAM
  • Default - 2133 15-15-15-36-1T
  • XMP1 - 3066 14-14-14-34-1T *
  • XMP2 - 3200 14-14-14-34-1T**
  • OC’d - 3333 14-14-14-34-1T
*3066 for Cinbench and CPUz benches only
**Stock - this is the rated speed of the kit.

Against these benchmarks:

  • Radeon Driver 8.2.3
  • MSI BIOS 7B09v17 12-19-2018

To capture hardware data I used:
  • HWINFO64 (5.72 3333) for voltages, clock speeds, power, temperatures, etc
  • Kill-A-Watt as a sanity check for system power against HWINFO64
  • iPhone 6s and dB Meter Pro for acoustics

Results_______________________

System Behavior Under Load at Stock

I ran Prime 95 SmallFFTs and recorded the following system behavior:

mGO1PNy.jpg


During the run, the 1900X ran an average of 3.83GHz. One to two steps above its base clock. VID was requesting voltages that peaked at 1.356V and dropped as low as 1.263 for certain cores. Vcore unsurprisingly closely followed VID:

owsBFID.jpg



Power use peaked at 262W for the entire system (idle is about 131W) and slowly dropped to 246W. This drop in power corresponded to a slow increase in temperature up to 67.8C Tdie and a slow drop in VID/Vcore.


At stock, AMD specifications say it has a max temperature of 68C. It appears that as the system heated up, voltage was reduced to prevent the cores from going over this limit. The cooler fans did ramp up to ~2700 RPM. Acoustics reached ~45dB up from a floor of 36 dB and ~ 39dB idle as measured using dB Meter about two feet from the case exhaust.


What was interesting was while the CPU wasn’t under load Turbo Boost and XFR had several cores up to 4.124GHz with Vcore as high as 1.43V. I’ve even seen XFR take the frequency as high as 4.175GHz.


Now I was hoping the Enermax Liqtech 360mm would be able to keep the 1900X at or below 60C. I may need to play with the fan curves and double check what other similar TR builds are getting at stock load as it’s possible I may need to use more or better TIM.


It also maybe that since the 1900X has the highest base clock at 3.8GHz and at stock runs pretty close to 3.9-4.0GHz that high temperatures and voltage are inevitable.


We know Zen gen 1 hits a voltage and frequency wall around 4.0GHz and other Threadrippers with 12 and 16 cores still have to share the same 180W maximum meaning they run with lower core voltages to stay under the same power cap.


If anyone other Threadripper owners have some thoughts I’d love to hear it.


Single and Multicore Synthetic Bench Scaling


For these tests I varied the RAM speed and timings with the CPU at stock to measure memory scaling. I also varied CPU speed with memory at stock to gauge OC’d performance.


ao6IAs3.jpg



First up memory scaling. For a 55% increase in bandwidth and 44% decrease in latency at best we see a 3.9% increase in the Cinebench R15 Multicore benchmark. CB R15 Single only sees a 1% gain with the CPUZ benches showing no memory scaling. So at least for these synthetics memory doesn’t play a huge part.


Frequency scaling on the other hand shows definite gains. For Cinebench Single Core we see a fairly linear increase of 0.04 points / MHz across the entire range. A 9% OC gives us a 9% increase in score


Cinebench Multicore shows a similar 7.7% increase for a 9% OC. However when we look at the difference in rate between 3.8GHz to 3.96GHz and 3.96GHz to 4.15GHz we something very interesting.

  • 3.8GHz to 3.96GHz - 0.20 points/ MHz
  • 3.96GHz to 4.15GHz -0.52 points/ MHz


The rate more than doubled. This accelerating increase in performance as frequency increases is very interesting. It’s possible it’s just fluke as I only had time to run each bench once. Or it could be an indication of increasing efficiency as clocks increase. It’s really to bad AMD has a such a clock deficit compared to high end Coffelake as a 5GHz Zen2 would probably be neck and neck.


Looking at ATs 2700X review the 1900X Cinebench Multicore Score puts it right behind the 7820 at stock and slightly ahead of the stock 2700X when OC’d.

97213.png



When AT tested the 1950X in Cinebench they had a multithreaded score of 3004 at stock compared my score of 1727. That’s 57.5% of the 1950X. The 1900X can clock higher on all cores since it has the same 180W TDP as the 1950X. ATs 1950X did a slightly higher single core at 167 vs my score of 164 possibly due to ATs CPU having a better bin than mine.


Encoding Benchmark Scaling


For our one non-synthetic benchmark I timed how long it took the 1900X to convert a 1 minute 4K movie clip to 1080P 30 using Handbrakes MKV X.265 preset as detailed in this thread.


KIf9Xeg.jpg


Lower is better


Much like our synthetic benches we see a similar performance change due to RAM and clock speeds.


2133 to 3200 buys us only a 4% reduction in encoding time. With really no measurable change going to 3333. In this test if you bought the top supported 2667 RAM you probably wouldn’t be missing any noticeable performance by not having faster RAM.


Clock speed on the other hand shows a nice 9% time reduction for a 9% clock speed increase just like in our Cinebench testing.


In fact when comparing the performance increase from 3.8GHz to 3.92GHz and 3.92GHz and 4.15GHz we see performance jump 50% as the clock speed increases, from -0.022 Seconds/MHz to -0.033 Seconds/MHz.


Another example of the 1900X gaining more performance the faster it’s OC’d. This performance puts the 1900X at 4.15GHz in the same neighborhood has other R7s at 4.0GHz, an i7 8700 at 4.3GHz, and the i7 5960X at 3.625GHz.


Synthetic Gaming Benchmark Scaling

Our last test is with 3DMark Time Spy and Firestrike. I tested against the over all Firestrike and Time Spy scores along with the Physics and CPU scores respectively.


For these runs the Red Devil Vega RX 56 8GB was used and left at stock.


Powercolor Red Devil Specifications:

  • Core Clock Speed: Averaged 1560MHz
  • HBM Clock Speed: 800MHz
  • TMUs: 224
  • Shader Cores: 3584
  • ROPS: 64
  • HBM Bus Width: 2048bits
  • Bandwidth: 409.6 GB/s


pad5Fi2.jpg



Firestrike

For Firestrike we see an interesting change in the two scores based on memory speed changes. For the actual Firestrike score we see a fairly significant 6.8% increase from 2133 to 3333 memory. Most of that gain comes in the first jump to 3200. Its really important for some of these tests to not leave the memory set at default.


Conversely during the physics test we see only a 1.8% increase going from 2133 to 3333 memory. Once more going from 2133 to 3200 actually shows a 2% loss in performance followed by a 3.9% gain from 3200-3333. These small changes are likely due to simple variations in the runs. It also shows that the physics test really isn’t bound by memory speed or latency.


Looking at a break down in performance as memory speed increases shows again that as long as you have memory running at or above rated speed you won’t be missing out on more than a few percentage points of performance


With a 9% clock speed increase we see a 4.2% increase in total Firestrike score and an 8.7% increase in physics score. The Firestrike score exhibits an increasing return on performance as clock speed increases. From 1.79 pts/MHZ to 2.20 pts/MHZ.


The physics score on the other hand shows a small 1.3% decrease in performance going from 3.8Ghz to stock. Followed by a 10% increase to 4.15GHz.


I don’t currently have a good explanation for the performance drop at stock speed. Overall score increased as expected but physics dropped. It’s possible I had a background process running that affected the score. With more time taking the average of a number of runs at each speed would have probably mitigated that issue.


Time Spy


Time Spy unlike Firestrike shows a much smaller impact to its overall score due to increasing memory speed at 1.9% but a much larger impact to the CPU test at over 15%. In both cases the increase in performance comes from memory speeds between 2133 and 3200 with little improvement from 3200 to 3333.


Time Spy also doesn’t show much change to the overall score due to clock speed increases coming in at just 1% for a 9% increase in CPU clock speed. That shouldn’t be to surprising and suggests that the bottleneck is likely in the GPU. All the increase in score came from OCing to 4.15GHz as the slower two speeds had virtually the same score.


The CPU test showed good scaling with a 6.2% increase for the 9% clock speed increase. The increase was linear over the range tested.


Power and Acoustics_________


Using the Kill a Watt I took power at the wall measurements during each of the stock runs and recorded the peak power usage. I also used HWIFO64 to record the peak temperatures:


Test Power(W) CPU Temp (C)

  • Idle: 131W 32.1C
  • CPUZ: 222W 64.3C
  • CBR15M: 244W 63.9C
  • CBR15S: 137W 52.4C
  • 3DM FS: 447W 64.6C
  • 3DM TS: 433W 64.8C
  • HandB: 257W 75.1C


The temperatures and power usage all seem reasonable with the exception of the 75.1C temp during the Handbrake run. It was single datapoint right at the end of the run when the rest of the temperatures had been in the low to mid 60’s. It’s possible it’s a measurement bug.


Acoustically the PC is very quiet under normal usage including gaming. It’s only under large CPU loads that the 3 120mm radiator fans spin up to their max and become loud. The sound is not unpleasant being simply rushing air, but it is very noticeable.


  • Noise Floor: 36dB
  • Idle: 38.5 dB
  • Full CPU Load: 45 dB


Conclusion_____________________


The 1900X is a thoroughly modern processor and was fairly easy to setup. The biggest problem I had was actually the BIOS. Several of the OC settings have poor UI. For example setting the multiplayer requires you to type the number in or type auto. There is no indication whether you’ve successfully selected the field or not until you start typing.


Performance wise the CPU is not significantly inhibited by being half a 1950X instead of a whole 1800X. The improved turbo clock speed (by being a top 5% bin) and XFR means much like the 2700X there is no real reason to OC. Lightly threaded loads see cores clocked at well over 4.1GHz and under all but the heaviest loads sees the 1900X running all cores around 3.95GHz.


Faster memory can provide some significant improvements (upto 15%) in some benches, especially Samsung b-die over 3000 MHz. However chasing faster memory appears to have diminishing returns. So if you have to save some money go for a quad kit in the 2667-3066 range and don’t sweat the 1-2% loss.


It’s too bad Zen is hitting a wall around 4.0-4.2GHz because in most of these benches performance is still increasing at or above a linear rate as clock speed increases.


Part 2 will be in AT VC&G as I undervolt and OC the Powercolor Red Devil RX Vega 56.


Bonus i7920 vs 1900X Fight!______


So how much faster is a new late 2017 1900X@stock / Vega 56 vs an old 2010 i7920 @ 3.5GHz / 7970GHz Ed?


.....Bench......I7 920....1900X.....X Faster
  • CPUZ M:...1549......4803........3.10X
  • CPUZ S:....293........449.........1.53X
  • CBR15 M:..549........1727.......3.15X
  • CBR15 S:...102........164.........1.61X
  • 3DM TS: ....2497......7023.......2.81X
  • 3DM TSC:..2809.......8780......3.13X
  • 3DM FS:.....7709......17055.....2.21X
  • 3DM FSP:..8647.......19569.....2.26X
  • HandBrk:...400sec....126sec...3.17X


Power and Acoustics:


.....Bench......I7 920..............1900X
  • Idle:...........170W(52dB)....131W(39dB)
  • CBR15 M:.265W(52dB).....257W(45dB)
  • 3DM FS:...459W(52dB).....447W(40dB)


A single 1900X core is 50-60% faster

Multithreaded the 1900Z is over 3X faster

Gaming benches put the Vega 56 at 2.2-2.8X faster.


All at roughly the same power draw, (& only 75% of the idle power). Not bad.


For my use I’m happy with machine so far. If you have any questions let me know and I can try and answer them.
 

Shmee

Memory & Storage, Graphics Cards Mod Elite Member
Super Moderator
Sep 13, 2008
7,382
2,419
146
Cool! While you are at it, you should put a Xeon X5660 in your x58 rig :D can be had for under $50.
 

Paratus

Lifer
Jun 4, 2004
16,614
13,297
146
Cool! While you are at it, you should put a Xeon X5660 in your x58 rig :D can be had for under $50.

I’ve thought about it. The case the X58 is in is really thermally constrained. The 4X 80mm Panaflow 3600RPM fans are noisy (50+dB) and the largest cooler I can fit is limited to 92mm fans.

However as a second system with just occasional use? Might be worthwhile now.
 

scannall

Golden Member
Jan 1, 2012
1,944
1,638
136
I’ve thought about it. The case the X58 is in is really thermally constrained. The 4X 80mm Panaflow 3600RPM fans are noisy (50+dB) and the largest cooler I can fit is limited to 92mm fans.

However as a second system with just occasional use? Might be worthwhile now.
The Westmere Xeon's draw less power than that i7-920. By quite a bit. Go faster, run cooler all at the same time.
 
  • Like
Reactions: Paratus

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
Cool! While you are at it, you should put a Xeon X5660 in your x58 rig :D can be had for under $50.
My only problem is that both of my socket 1366 systems are having problems with the 1709 win 10 update. It makes them unusable (black screen) and won't even boot in safe mode to completion.
 
  • Like
Reactions: Drazick

Shmee

Memory & Storage, Graphics Cards Mod Elite Member
Super Moderator
Sep 13, 2008
7,382
2,419
146
My only problem is that both of my socket 1366 systems are having problems with the 1709 win 10 update. It makes them unusable (black screen) and won't even boot in safe mode to completion.
That is very odd. No luck with a fresh install of latest build? It could be that a windows update or something broke the OS, but a fresh install of latest version might work.
 

scannall

Golden Member
Jan 1, 2012
1,944
1,638
136
My only problem is that both of my socket 1366 systems are having problems with the 1709 win 10 update. It makes them unusable (black screen) and won't even boot in safe mode to completion.
To echo what Schmee said, that is odd. Works fine here as well.