Evaluating CPU cores for optimisation - page 4
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Above - Tree_Brut_TestPL_F
Here's the second one, in the same configuration:
Same (without F) when running 12 virtual cores:
Expert "with F" when running 12 virtual cores:
I'm going to turn off hypertrading now...
Expert "without F" and without hypertrading with six passes:
Expert "with F" and without hypertrading with six passes:
Very strange result - it turns outTree_Brut_TestPL without hypertrading disabled in bios 6 agents average pass 0:00:33.712, and with it disabled about 1:10 - no confusion?
While forTree_Brut_TestPL_F Expert Advisorwithout hypertradition disabled there are 6 agents" average pass 0:01:10.931"in the bios and "average pass 0:01:10.621" when disabled, i.e. there is no difference.
Next Subject Phenom II X4 960T 3000Mhz
Tree_Brut_TestPL - 4 agents
Tree_Brut_TestPL_F - 4 agents
Testing the FX-8350 again, but it's running at 4200Mhz and in one memory channel and another mother on the 760G chip
Tree_Brut_TestPL - 4 agents
Tree_Brut_TestPL - 8 agents
Tree_Brut_TestPL_F - 4 agents
Tree_Brut_TestPL_F - 8 agents
Very strange result - it turns outTree_Brut_TestPL without switching off hypertrading in bios 6 agents average pass 0:00:33.712, and with switching off about 1:10 - no confusion?
While theTree_Brut_TestPL_F Expert Advisorwithout hypertradition disabled has 6 agents in the bios with" average pass 0:01:10.931" and with it disabled has "average pass 0:01:10.621", i.e. there is no difference.
I also think that hypertrading shouldn't make a difference when the task is purely computational. As I said before, the main benefit of virtual cores in optimization is precisely that we're not accessing memory, everything is in our cache. When I was testing different variants - I remember very well that on "real ticks" for two years there was no profit from hypertrading. And on the contrary, if I work with 1MOHLC I immediately get 10-20% profit. That is, the gain is due to caching.
The first result, on the other hand, surprises me...
The next test subject is a mother with two workstation/server CPUs, E5-2670 v1 processors at 3000Mhz (the mother forcibly boosts them)
Tree_Brut_TestPL - 16 agents
Tree_Brut_TestPL - 32 agents
Tree_Brut_TestPL_F - 16 agents
Tree_Brut_TestPL_F - 32 agents
I'm troubled by the scatter of agents in all variants - best - worst - up to 25 seconds - perhaps it is a payment for the dual-processor architecture, but I'm not sure - if anyone has a similar setup, please test it.
It seems to me, too, that hypertrading shouldn't make a difference when the task is purely computational. I've already said that the main benefit of virtual cores in optimization is that we don't access memory, everything is in the cache. When I was testing different variants - I remember very well that on "real ticks" for two years there was no profit from hypertrading. But if I work with 1MOHLC I immediately get 10-20% profit. That is the profit is due to caching.
But the first result - I'm surprised myself...
I'm not arguing with the theory, maybe it is so: all code with appendage in cache will fit in it and everything will be good and fast, while in the opposite case we will work with memory.
ConcerningTree_Brut_TestPL test- I think you must have made a mistake by switching off hypertrading in BIOS - try it again, the first Expert Advisor is faster than the second, so I assume this is a mistake.
So far, the ranking is based on the average run time in seconds - the two penultimate columns, and the last column shows the number of CPU runs per hour.
The table is filtered by the last column as the heaviest EA option in terms of computational resource consumption.
So far I'm surprised by poor result of E5-2670 processors, while it (2 of them) scored 2000 points in Cinabench r15, i7-8700 got only 1191 points! Something is not right here...
Tree_Brut_TestPL - 2 agents
Tree_Brut_TestPL_F - 2 agents
2990WX test results
Tree_Brut_TestPL - 32 agents
Tree_Brut_TestPL - 64 agents
Tree_Brut_TestPL_F - 32 agents
Tree_Brut_TestPL_F - 64 agents
The results are very shocking and I cannot understand what's wrong - the compiler doesn't understand this architecture or something else - single runs are extremely mediocre and the spread is very large.
Yes, the overall result is the fastest for the processor, but I expected much better results! Who has Ryzen - let's test it!
2990WX test results
Tree_Brut_TestPL - 32 agents
Tree_Brut_TestPL - 64 agents
Tree_Brut_TestPL_F - 32 agents
Tree_Brut_TestPL_F - 64 agents
The results are very shocking and I cannot understand what's wrong - the compiler doesn't understand this architecture or something else - single runs are extremely mediocre and the spread is very large.
Yes, the overall result is the fastest for the processor, but I expected much better results! Who has Ryzen - let's test it!
The 2990WX is different. It consists of four Zeppelin crystals, with 32 processing cores. On the X399 platform, AMD has imposed some restrictions on this processor so that it doesn't hurt sales of EPYC server chips.
Chief among these limitations is the presence of only four memory controllers. Although there are two more Zeppelin chips, AMD calls them compute chips. This means they don't have access to local PCIe or DRAM, for that they have to address I/O components via Infinity Fabric. As there are twice as many crystals, the bandwidth of the Infinity fabric is twice as low, around 25Gb/s if DDR4-3200 memory is used.