PC setups, configurations, benchmarks and general recommendations

Because the interest is high among new users of SQX (i saw many topics with the same goal), i thought it would be a good idea to concentrate everything hardware related in one topic where new users but also more experienced ones, can easely find in one place what they are looking for. What to buy, what configuration, air or liquid cooled…etc.

This topic is designed to help users in choosing the most apropriate hardware configuration according to their budget, a new computer or an upgrade. Also, posting benchmarks for processors, general raw performance benchmarks as a guide but also examples of settings in SQX and the resulting performance for others to have an idea what expectations they can have from a certain processor.

I am sure the contribution of the more experienced users will be greatly appreciated by everyone.

For starters, in the following days and weeks, i will post different results (pairs and timeframes) on my configuration, an AMD 3700X (3,6Ghz 8 cores) and 32GB RAM.

To be relevant, we can agree on a minimum criteria, for example, each setting must be run at least 24 hours, and we can agree to count the accepted strategies/hour (in databank).

yes, indeed

if i were to make a recommendation for a custom build, i would say the 3900X, 3.8 Ghz 12 cores, is or sits at the sweetest spot as price/performance ratio. Just below the top Threadrippers and on AM4, a very cheap socket, cheap motherboards and air cooling. A brand new desktop with 3900X it would cost aprox 1.100 EUR.

I consider the level of the 3700X and other similar performance to be the minimum. Before i decided, i draw a red line below the 3700X, considering anything below, too weak for generating.

Presently, there is a long list of pairs and timeframes where i cant generate strategies in reasonable time. By reasonable, i mean anything above 30 accepted strategies per hour. If its below, i skip it and move on.

And as we speak, it starts https://www.youtube.com/watch?v=a8apEJ5Zt2s .

for Threadrippers, if i were to make an assessment, today, at current available models and prices, as the new 3rd gen Ryzen 9 3900X is very close to the former 2nd gen TR 2920X and 2950X, i wouldn’t recommend these two because the platform is more expensive and you pay more for the same performance, so 3900X is ideal.

Above 3900X is the Ryzen 9 3950X, still AM4 but with liquid cooling, sits in the middle between the rest and the very top 3rd gen Threadrippers 3960X and 3970X.

Whoever wants the very best and max, the two new 3rd gen TR is the only way. Probably close to 3.000 EUR or a little more (for the entire system). Having huge power consumption, only experienced traders can be profitable by using such systems, one must know SQX inside out and all building blocks by heart and know what to look for.

For me, personally, i would at least in theory, buy but i would need hard proof and extensive benchmarks to be sure that the performance increase is real in SQ on several markets and timeframes.

For example, on USDJPY or EURGBP where i had 1 accepted strategy/hour on H1 or M30, or even less, i also had 10 strategies in 24 hours, what number would a TR reach, when we would need at least 1.000 in databank. I have some suspicions that the performance increase is not proportional in the actual field or “battleground”.

Yes, you are right.  I recently got hooked on these YOUTUBE build videos that I did not know existed.

there is also a third solution, a hybrid between a custom build and a refurbished workstation:

a custom build workstation with every component being new, except for the two processors which are refurbished

https://talcikdemovicova.com/32-thread-build-on-budget/

the article is from 2016, well before the rise of the new AMD Ryzen processors, that configuration is a bit weaker than a 3700X but roughly at the same price

regarding temperatures, my pc sits around 60C on the CPU, 55C on VRM, 30C on chipset, with the fans around 80% speed, very relaxed but…..(with a very big but) it sits in a small unheated warehouse on around 22C. Keeping the unit in a pleasant warm room of 26-27C, for example if one has children, the unit gets stressed and the noise very apparent especially at night, so a warm room and a unit permanently on full load, are incompatible or very uncomfortable.

Analysis of Our Q4 2019 HEDT Processor Reviews

https://www.servethehome.com/analysis-of-our-q4-2019-hedt-processor-reviews/

One interesting note here is that today’s launches are essentially desktop launches of server products. The new AMD Threadripper platform is a de-featured AMD EPYC 7002 Series “Rome” CPU that adds clock speed and a chipset. The LGA2066 socket that the Core i9-10980XE sits in is an evolution of the four-generation LGA2011 Xeon socket. That Intel Core i9 is a Xeon with overclocking support and without ECC RDIMM support. These are both server chips that have seen feature changes to make them HEDT parts to replace prior generations of Xeon workstations.

56C on CPU and VRM, in full 100% load, with 3 fans at 50% in a cold unheated room. If i increase the load on the fans, the temp lowers at around 48C.

One advantage with large 120 or 140 mm fans with PWM control is that you can easily adjust the temp of the system

the unit has no keyboard, mouse or monitor attached, its controlled only by Windows Remote Desktop.

in the last few days, i brought this energy meter:

https://www.peaktech.de/productdetail/kategorie/engergiemessgeraet/produkt/p-9035.html

so now we have exact measurements. To summarize again my configuration (measurements for unit only, no monitor or keyboard, mouse):

Processor: AMD 3700X 3.6 Ghz 8 cores
Motherboard: Gigabyte B450M DS3H
Video card: Zotac GT730 2GB DDR3 Zone Edition (passive cooled)
Memory: Corsair 32GB DDR4 3000Mhz CL16 CMK32GX4M2D3000C16
SSD M.2 NMVe: Samsung 970 EVO Plus 500GB
Power supply: Seasonic M12II EVO Bronze 520W
Case fans: 3 x Arctic F12 PWM with PST 120 mm, 4 pins
Cooler CPU: Arctic Freezer 33

average power: 119 – 120W with max 123W, full load 24h a day.

the monthly consumption in Kwh will follow

again excellent thread. thanks ivan;  I have amd 2950 threadripper 64 gb ram ec2 disk and im satisfied 🙂 your choice is ok

clonex / Ivan Hudec wrote:

again excellent thread. thanks ivan; 

thank you

another useful information for beginers, in cases where work is done with the max 10.000 in databank, on 32GB, the software “chokes” and stops responding because the heap memory reaches full so 64GB RAM is the minimum for a complete system where one can work without any problems.

Indeed, it must be said that 10.000 in databank is reached only in very few cases but nonetheless, a minimum of 64GB is desirable. I will also look to add more ram.

to be more exact, not the generating itself is the problem but the retest phases like different timeframes…etc, at that moment when it gets at about half of the work, it chokes

everything lies in setup, and because of that the user need to know what he is doing – its not neccessary to save to databanks 10000 strategies and work with them after that – better to make smaller packs with 1000 strats and make different packs with different settings and make very quick final filtering.

only this way you will diversify your portfolios

Best to use custom projects.

I have 128 GB RAM because of old SQ3, now my projects are using only 20 GB RAM

hankeys wrote:

everything lies in setup, and because of that the user need to know what he is doing – its not neccessary to save to databanks 10000 strategies and work with them after that – better to make smaller packs with 1000 strats and make different packs with different settings and make very quick final filtering. only this way you will diversify your portfolios Best to use custom projects. I have 128 GB RAM because of old SQ3, now my projects are using only 20 GB RAM

yes, these are very good ideas regarding setup

i think it would be a good idea to centralize all SQ setup advice or good practices into one thread where everyone but especially beginners can find easily setup advice. Currently, all these useful information on setup are scattered because each beginner started a thread on each small question but all refering to setup.

This thread we should keep it for hardware setup only although they are linked. If one uses selected timeframe precision setting, the generation process is very quick and demanding, reaching 10.000 in a few hours on some pairs, but the downside is that on the retest, if one retests on 1 minute precision, all of those 10000 will fail. But the idea is that being so demanding, the software is in risk of “not responding” if it runs only on 32GB RAM

If you use 1 minute tick precision, the generation process is too slow and you need weeks of generating, although in this case, on the retest, many strategies will pass

The monthly power consumption for my configuration in around 85 – 90 kw (based on 5 days measurements). But this is a consumption with several breaks between projects included, not continuous operation. Probably, as well based on assesment, for 30 days continuous 100% load, the consumption would be at about 120-130 kw.

The 3900X has become not only wide available but now it can be brought cheaper from the grey market, sealed box for as low as 380EUR in several european countries.

the 3900X has 3.8 Ghz and 12 cores and 227W consumption (entire system).

The configuration would include an air cooler like Cryorig H5 universal and 2 or 3 case fans of 140 mm.

I attached a benchmark list to see where 3900X sits.

This is right now the best buy and the best choice for mainstream SQ systems.

In a few months, its possible that i will upgrade to 3900X and perform extensive tests and benchmarks which will clarify many unknowns to many, including me.

Regarding the average strategies per hour count, there are a few things. The speed has ups and downs but decreases as the possibilities are exhausted. The complete average is much lower than in the first hour of generating. For example if in the first hour you get 300/hour, by the time you reach 5.000 or 10.000, the average is much less, like 120/hour. I mean if you divide the 5.000 to the total time it took.

Best results, best speed was in GBPJPY and XAUUSD, with 300 – 800/hour with a bit more complex strategies, on both precision types, selected timeframe and 1 minute. On majors like GBPUSD, EURUSD, the speed is much slower, 70 – 100/hour and decent only on specific timeframes, usually larger like H4.

Also, another important aspect, the quality of the databank can vary greatly. In some cases, with 5.000, you are left with many final strategies after all filters and robustness checks, in other cases, with 8.000, by the time you reach the final filters, you are left with 0.

The greatest challenge is on CAD and CHF pairs (EURCAD, USDCAD, EURCHF, USDCHF), nevermind crosses. So far i was unable to obtain working strategies.

Regarding memory, i just upgraded to 64GB from 32GB. In many cases, on 32GB is simply not enough, especially given the fact and most beginners (like me) are holding their test planforms on the same system. At this level, its not efficient to keep two PCs in continuous use, one for generating and one for keeping the platforms. So i strongly recommend to start with a minimum of 64GB.