New on LowEndTalk? Please Register and read our Community Rules.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Dedicated with 16+ threads, power/price ratio
Alienighted
Member
I am looking for dedicated server with best power/price ratio.
Requirements: 8+ cores, 4 GB RAM, max 300USD/server.
Due to way of how CPUs are used by my computing algorithm, following formula should be used:
threads x clock / price
So far i found:
2x5650 24x3.06=73 / $55 = 1.32 wholesaleinternet instant
2x2670 32x3.3=99 / $79 = 1.25 wholesaleinternet custom
Comments
As a genuine advice, get their passmark score and use that instead of threads x clock
Clinging onto Westmere EP in 2018...
FORMULA INCLUDED!? That's elementary school maths at best.
No offense but its bad advice.
https://www.cpubenchmark.net/cpu.php?cpu=Intel+Xeon+E3-1285+v6+@+4.10GHz&id=3158
https://www.cpubenchmark.net/cpu.php?cpu=Intel+Xeon+E5-2670+@+2.60GHz&id=1220
Just check this, according to passmark these 2 cpus should have similar performance. Thats why passmark is total joke, its not accurate, you cant predict CPU performance in real life scenarios when you take a look on passmark score (unless you are gamer).
Not following your formula. See https://www.lowendtalk.com/discussion/comment/2700029/#Comment_2700029
This laughing is completly pointless. I am right with my formula in my case.
I am performing a computing operations which has nothing to how Passmark result is calculated. I have algorithm which is strict based on the combined cores power.
Points = result of algorithm.
I tested over 12 known Xeon and i7 CPUs, and result is always same = GHz ratio give more equal results compared to the Passmark.
Thank you, but seems like this thread is outdated, it is from march and there are single machines, i need atleast over 10 of them in one provider.
@Alienighted You're brilliant - I really appreciate the formula!!
Ask them. IOFlood servers seem to be in stock.
You should get a job at NASA
If reading with understanding is difficult for most of humans, that is right - i should.
It really depends on the workload @Alienighted.
If your app is mostly single core, than multi core will not benefit you a lot, but rather something that can be bursted to more GHz a.k.a TurboBoost. Of course a multi threaded app will benefit from more cores if it's written correctly, but may not need high clock rate.
This all depends on the workload you're doing, so the CPU benchmark score is not the final number you should look, but the single core and multi core score in order to determine what you may need.
Your calculation is not bad in general, but not all clock rates are the same it also depends on the CPU architecture, instructions that your CPU has, and much more. Compare some older AMD and Intel CPUs with similar specs, and you will see differences in benchmark but may not see a difference when doing your calculation.
Also a lot of apps depend on the disk speed as well, so it really depends, you can have a lightning fast CPU with a lot of cores, but your bottle neck will be a HDD for example.
Quick! Everyone dust off your Athlon X2s and Curo2Duos!
Take for example:
Xeon X3370 vs Xeon E3-1220v5 (http://cpu.userbenchmark.com/Compare/Intel-Xeon-X3370-vs-Intel-Xeon-E3-1220-v5/m17169vsm139116)
Both run at about 3GHz, but the single code performance of the E3 is almost the double of the X3370, and why is that?
Read on:
https://en.wikipedia.org/wiki/Instructions_per_cycle
And don't they have the same performance? (I'm genuinely interested)
I'll add the geekbench also in here:
https://browser.geekbench.com/processors/2097
https://browser.geekbench.com/processors/896
No they dont have similar performance in server scenarios. Geekbench is also crap.
I have some Xeon 1230v3/v5 servers and dual Xeon 2660 now, ive also had dual Xeon 2620v4, 5650, 5639 and others in the past and i compared performance in heavy cpu tasks.
I know crypto mining is disliked here, but hashrate gives some info about CPU real power. Take a look for example here:
https://forums.servethehome.com/index.php?threads/aeon-mining-performance.16983/
There are also benchmarks like for example cinebench. Its not perfect, and it dont scale up good with cores count, but it gives much more accurate info about cpu performance. Take a look also at corona benchmark, cray etc.
Just to give you an example:
https://www.cpubenchmark.net/cpu.php?cpu=Intel+Xeon+X5650+@+2.67GHz&id=1304&cpuCount=2
Thats dual Xeon 5650, very old CPU. In real scenario this cpu is able to handle 2-2.5 times bigger workload than Xeon 1230v3-v5. For sure it can handle 2 times bigger workload than single 12XXv6
I would say single Xeon 2670 is able to handle nearly 2 times bigger workload than 12XXv6, single 2670v2/2650v3/2630v4 is able to handle 2 times bigger workload than single 12XXv6 for sure.
One more thing:
https://www.cpubenchmark.net/cpu.php?cpu=AMD+EPYC+7501&id=3153
According to cpubenchmark Xeon 12XXv6 have 50% of AMD EPYC 7501. We are comparing here 4cores/8threads to 32cores/64threads. Just lmao...
Take a look here what kind of CPUs this epyc kick in ass:
https://www.servethehome.com/amd-epyc-7501-benchmarks-and-review-32-cores-per-socket/2/
Just few worlds about geekbench. Here you have geekbench of AMD Epyc 7601 - https://browser.geekbench.com/processors/2060 , according to this Epyc 7601 is only ~35% faster in multicore scenarios than Xeon 12XXv6. Just LMAO again.
No offense but cpubenchmark, geekbench and probably made by idiots. These benchmarks just dont show real cpu performance in multithread scenarios.
I find passmark predicts real workload performance pretty well, much better than geekbench.
You have points. But I think CPU performance depends on tasks and cache. The benchmark is designed to perform various different tasks to asset the general performance.
For Aeon mining, it uses pure CPU power to calculate, no cache hungry. Eg: Gold 6134 is just 25% more powerful than Atom C3955. It's just meaningless to asset CPU perf. If it's true, noone will buy Gold 6134 because it's the same generation and 5 times more expensive.
For Geekbench and your EPYC 7601 case, I think Geekbench is not (yet) designed to test AMD architecture. The results are also very few so the error margin would be high. Eg:
This test give 94k: https://browser.geekbench.com/v4/cpu/6818590
But this test https://browser.geekbench.com/v4/cpu/4502830 gives only 13k.
I also don't rely on geekbench score because when they perform the test for multiple threads, the cpu load is about 30%.
Moreover, the more core counts, the more context switching will occur and it affects the performance a lot. That's why you will see higher performance in less core respectively.
In my case, I have performed some tests that I designed myself. It surf a lot of websites simultaneously. I achieve the same performance ration with passmark though more cores seems to give better performance.
I don't find it's nearly 2x. Maybe 1.5x tops, which is what you'd expect by frequency scaling. Even that is sort of doubtful.
Benchmarks perform a wide range of tests to the CPU, while mining only performs one.
As I understood, some benchmark tools are not able to give a fair score to high core count or dual CPUs. Mining scales better with core count, and therefore "shows" the difference in performance.
We just need to find a better way to benchmark, because mining software is not a way to benchmark CPUs. CPUs with a certain instuction set or more cache (depending on the algorithm) will have higher hashrates, while not being necessarily faster in other tasks.
If mining scales with pure core count * frequency, then Scaleway would be a mining haven. I don't think it is though.
Are you actually talking about scaleway.com ?
Best what they offer is ARMv8 (i may not be even A57), but even tho, 64c x 1.9=121 for $328. I can get 4x Dual E5-2670 4x32x3.3=420 which cost me $280.
Obviously such comparistion here is a bit pointles, but i just wonder what you meant.
If you're only after cores the 64c instance at 280 euro is silly since you can get two of the 32c's for 70 euro each = 140 euro. Or for that matter 16 of the 4c for 48 euro total. Maybe even 33 euro total if you shut off the public IP addresses which are 1 euro each on 15 of them. You can still reach the non-public ones from the internal network so you'd get a public ip for just 1 of them, and connect to the rest through it.
They also have the C1 which is a 4-core dedicated server at something like 1.2ghz, at 3 euro (2 euro without the public ip). That's a little more expensive per core-cycle but you'd have those cores completely to yourself.
How do you get those numbers? The E5-2670 is an 8 core cpu (you can't count the hyperthreads as cores) and its base clock rate is 2.6 ghz, not 3.3. So it would be 4x16x2.6=166.4. Maybe you can add 10%-20% for the hyperthreads depending on your app's memory access pattern. You're still below 200.
I was looking for a low key funny thread on the front page. Guys, I finally found it.
It's just a summer thing. Kids with too much time will do anything. And I do mean anything.
Shhh, don't tell him that. If htop shows 16 bars, it means it has 16 cores goddammit!
For 300 USD I can offer - 2 x 2678v3 ( total 24 cores, 48 threads x 2.5 Ghz ( Turbo 3.1Ghz ), 16GB RAM, 1 x 250GB SSD, 1Gbps @ 50TB
https://www.cpubenchmark.net/cpu.php?cpu=Intel+Xeon+E5-2678+v3+@+2.50GHz&id=2584
Order page: https://www.bacloud.com/cart.php?a=add&pid=126
24c 2.5 $300?
I am ready to rent machines 16C 3.3, 64GB each for 1/2 of your price, gonna be rich soon still!
So you want to tell me that HT do not give advantage over ARM?
Sure i do, as long as i am getting more money by result of my calculations and tests than based on pure passmark one sir.
What we've been trying to tell you all along is:
1) cores aren't directly comparable across architectures even when scaled by frequency. they're sort of correlated but that's about it.
2) HT usually gives advantage over non-HT but it's not anything like 2x advantage. It's 1.1x or maybe 1.2x if you're lucky, and it can actually be less than 1x, depending on what you're doing. For explanation, watch the video that vovler posted.
Your "32c" E5 server is actually 16c+HT which is about like 18c or 19c without HT, at best. What are you actually running on it and can you post some benchmarks instead of acting all mysterious?
1) yes i know it, depend of purpose, and it is just best to determinate with performing a tests
By 32c i mean two machines, but still.
I am mostly running direct assembler instructions out of regular os.
With google im getting 48c HT Skylake under $550, but it not allow me to do what im doing with some of the dedicated servers.
No i am not crypto miner like someone previously assumed.