AMD Ryzen Threadripper 3rd Gen Overclocking Deep Dive, feat. ASUS ROG Zenith II Extreme 48

AMD Ryzen Threadripper 3rd Gen Overclocking Deep Dive, feat. ASUS ROG Zenith II Extreme

(48 User Comments) »

Introduction

AMD Logo

In this article our resident AMD Ryzen memory and overclocking guru Yuri "1usmus" Bubliy will dive deep into the dozens of new settings introduced with third-generation Ryzen Threadripper processors, so they become accessible both for platform beginners and enthusiast overclockers wanting to climb benchmark leaderboards. Special attention is given to memory settings, one of the most important dials to maximize overclocking performance on AMD's new third-generation Threadrippers.

After comprehensively beating Intel in the desktop space with its third-generation Ryzen processor family, AMD turned its attention to the high-end desktop (HEDT) market with the third-generation Ryzen Threadripper 3000 "Castle Peak" processors. Designed for the new Socket sTRX4 and AMD TRX40 chipsets, these processors come in core counts ranging from 24 to 64 and are essentially client-segment implementations of the company's "Rome" EPYC multi-chip module (MCM).

Besides the performance uplift from the "Zen 2" microarchitecture and increased clock-speeds because of the 7 nm silicon fabrication process, third-generation Threadrippers benefit from the centralization of I/O to the new 12 nm I/O controller die. This results in a new monolithic quad-channel memory interface in which each CPU core has equal access to all available memory bandwidth. For previous-generation Threadripper WX processors, half the cores would have indirect memory access, resulting in severe performance bottlenecks. The I/O controller also nucleates the processor's PCI-Express gen 4.0 root complex, and other platform I/O. AMD also quadrupled the chipset-bus bandwidth by implementing PCI-Express 4.0 x8 between the sTRX4 processor and TRX40 chipset.



HEDT platforms are targeted at the gray area between desktops and workstations. The ideal HEDT users game in their free time, but use their PC for serious money-making through content creation. Due to the increased complexity of the silicon and the platform itself, HEDT processors typically have several more performance-tuning options than mainstream desktop chips, such as the Ryzen 9 3950X.

All testing in this article features a Ryzen Threadripper 3960X processor and an ASUS ROG Zenith II Extreme motherboard. The Zenith II comes with a treasure chest of BIOS settings, including direct access to AMD CBS. It also has 70 A power stages that provide plenty of power for overclocking our Threadripper chip, along with full readiness for the upcoming 64-core 3990X processor.

Heat Dissipation and Cooling

I was extremely surprised by many of the launch-day reviews, which had reviewers pair all-in-one watercoolers or the worst custom watercoolers with the new Threadrippers—these cover only 65% of the hot die surface, which won't be enough to manage the 280 W TDP. In many cases, these reviewers commented that "the product came out quite hot and there was no sense in overclocking." This is probably the main reason why I wanted to share my findings with you, which weren't hacked together in a few stressful days leading up to AMD's review embargo.

Chip manufacturers are reducing the fabrication process size in order to increase the number of chips they can harvest from a single wafer and reduce the energy consumption of the final chip. A smaller process node also gives the manufacturer the opportunity to increase the operating frequency of the chip, while leaving its dimensions unchanged.

For a long time, this tendency (to reduce process size) remained a solid approach, but companies have now begun to delay or even stop developing new silicon nodes. This is partly due to the huge cost of investment and the higher defect rate for these new processes.


Comparing AMD's Zen+ with Zen 2, we got a doubling of transistor density with the same (or slightly lower) power consumption. In numbers, it looks like this:
  • The Zen+ CCD has an area of 213 mm² and a TDP of 145 W; in terms of area, we get 0.68 W/mm².
  • The Zen 2 CCD has an area of 74 mm² (per 8-core chiplet) and a TDP of 95 W (average); in terms of area, we get 1.28 W/mm².
The next nuance is electron tunneling—when the gate becomes too thin and electrons can pass through it. This leads to a loss of charge accumulated on the gate of the transistor, which requires restoration. The result is a transistor that consumes more current, which in turn leads to higher heat dissipation. Individual transistors have almost immeasurable amounts of current loss and temperature increase, but when several billion transistors are placed on one piece of silicon, working together, the effect accumulates and becomes a serious problem. Also, the current does not just flow out of the gate, it can tunnel from source to drain if they are in close proximity, which may impede the transistor's ability to control the current.

With increasing voltage, the leakage current increases linearly or even worse. The effect of temperature on leakage is relatively weak. The formation of leakage current, as a rule, is associated with the imperfection of the manufacturing technology, so do not be surprised about news that your favorite processor has received a new stepping or another suffix with small process technology improvements.

The last important nuance is SIDD (static IDD). Processors can be divided into two camps: high leakage current (high SIDD) and low leakage current (low SIDD). As always, there are differences between samples within the same leakage rate, which means there are good and bad samples in both the low and high leakage categories. A high SIDD sample with a margin of P0 VID will certainly be able to reach X MHz at a much lower voltage than a low SIDD sample. High leakage chips require significantly lower voltage than low leakage models, but consume more current at the same time and heat up much faster than low leakage instances. Also, their breakdown voltage is lower.

As for the third-generation Threadripper line, we have one CCD (or at least one CCX) with high SIDD and three CCDs with low SIDD. This allows AMD to maximize boost clock and use dies that cannot overclock too well, but are not defective, either. This is how AMD killed two birds with one stone. More on that later.

Cooling the Threadripper


Let's start with the heart of the setup—the water block. In this review, I will use the EK-Velocity sTR4 RGB, a full nickel water block from EKWB. There are rumors about an RGB version of that block coming out soon; maybe it will even increase overclocking potential, too.


Its main advantage is the huge chamber area (about 29.7 cm²), while the inner part has 91 micro-fins separated by a wide recess in the central part. This maximizes the cooling for both CCDs and the huge IO die.


The microchannels cover all of the five dies of the Ryzen Threadripper processor, which ensures uniform heat dissipation and better IHS (Integrated Heat Spreader) performance. Above, you see a coverage comparison with the EK-Velocity on the left and a typical AIO waterblock on the right.


Both internal and external surfaces of the main plate are made out of pure electrolytic copper with a nickel coating to prevent corrosion and extend the life of the water block. The weight of the water block in this case reaches a monstrous 980 g. It's also worth mentioning that in addition to AMD Ryzen Threadripper processors of all generations, which are made in the Socket TR4 / sTRX4 package, the EK-Velocity sTR4 water block is also compatible with the identical Socket SP3r2 and Socket SP3r3 designed for AMD EPYC server processors. Another advantage of this beauty is the ease of installation because you do not need to remove the motherboard from the case or use any additional tools.


I'm using two EK-CoolStream CE 420 radiators. The CE 420 is a three-section copper (H90) radiator with a fin density of 16.

I chose those radiators to drop the water temperature as close to the room temperature as possible. The thermal performance of this solution allows you to dissipate up to 1.2 kW of heat, at a noise level of 37-38 dB.


As for the pump, nothing special was chosen; it's a classic time-tested D5 from EKWB, which is based on the Xylem D5 PWM water pump, with a maximum flow rate of up to 1,500 l/h.


I'm using EK-Vardar EVO 140ER fans as they have some of the best static pressure. This eliminates the need for a push-pull build, saving space while maximizing heat dissipation.


The ASUS GeForce RTX 2080 STRIX graphics card is cooled with an EK-Quantum Vector STRIX RTX 2080 D-RGB - Nickel + Acetal. It's an amazing solution that keeps the temperature of the graphics card at 37°C while overclocked to a TDP of 270 W.
Our Patreon Silver Supporters can read articles in single-page format.
Discuss(48 User Comments)