Branded vs Assembled PC

[rsjsouza]

A well built DIY PC can certainly last. I built mine 8 years ago and it's approaching 60,000 operating hours according to the main SSD, only major component replaced was the GPU in order to add 4K compatibility.

I just opened your photos: Mike, why do you have a massive computer grade capacitor inside?

[SilverSolder]

I still remember the days (90's) when running Photoshop could really tax a PC to its limits (and beyond!).  These days, even a low end PC is fine.  Progress! 

Maybe it is before the time many people around here, but Photoshop (version 6, IIRC) was the first software that made great use of the then newfangled MMX extensions of the latest models of the Pentium I generation. I remember Intel promoting their compiler which made use of this technology. It was a very large leap for the time.

Even running the highest end Apple kit at the time, stuffed full of RAM, it was a real chore running Photoshop -  but it was the future, right?   And here we are...   

[tkamiya]

What surprises me is how many software don't use multiple cores very efficiently, if at all.  Yet, number of cores are big selling point on CPUs.  Also, memory usage is not that efficient on Windows.  Photoshop is very weird in this regard.  I used to run it on dual cpu system with 6 cores/12 threads each.  One process will absolutely peg.  Few processes show some activities and none on rest.  Yet, system itself is struggling.  Also, if the system has 8 gigs of memory, it will start to swap.  I added 36 gigs to it.  Now, it won't even use 8 gigs....   

How long have we had multi-core cpus now?  I'd thought we'd be making use of them more by now.

[bd139]

It's because not all algorithms make sense on multiple cores.

Consider a simple one such as the Fibonacci Sequence. Say we need the first 1 million numbers from that sequence.

If we use the basic formula which is "the results from the last two values added together" then we have a dependency on the previous state of the algorithm. That means that they have to be done in sequence to get the next state. This can't really be shared across multiple cores with any efficiency gain because they have to be done in order so regardless of how many cores you have it'll still only be able to use one at a time.

However if someone comes along and produces a formula where you trivially plug the number in and out pops the result without any knowledge of previous state then if you have 4 cores you can stick 250,000 of those calculations through each core in parallel, then sort the results as they turn up. win! Well no actually...

If you go and look up n-th term formulas for fibonacci you'll see what a pain in the ass it is moving from one to the other and how difficult it is. Sometimes it's easier to just wait and that's what a lot of software does. Some problems aren't even possible to solve in this way (yet).

Really multiple cores are great for when multiple separate things at once, a few very specialist loads where one thing is done in parallel and for keeping the house warm

[tkamiya]

My background is enterprise level applications on Linux/Unix.  I spent 20+ years at not-to-be-named big database company as a programmer.  Other than those software where multiple copies of same thing runs simultaneously, I really don't see a point in high core count CPUs.  What you said exactly my point.

Getting back to the original topic, one of my suggestion would be choosing the right CPU.  Spending $$$ on top end core count CPU may not be the best course of action on desktop PCs.  Raw clock speed may be more important.  Personally, I don't see a point in going more than 4 cores. 

So far, other than databases, web servers, or concurrent anything, I haven't seen good use of lots of cores.  This is where comparing raw benchmark numbers will fail you.  In case of Photoshop, applying filters where processes are repeated and few others, jobs appears to run on a single core.

[bd139]

Games use a lot of cores. My desktop is mostly used for running up entire application stacks with tens of containers so that tends to spank it a bit. Also video transcoding.

[olkipukki]

Personally, I don't see a point in going more than 4 cores. 

Well, if you are using (let's say) Premiere Pro instead of Photoshop, you just wasting your time then...

Photoshop well-know a single threaded dependency, but it doesn't mean that people will stop to buy 16/32 cores if just single one outperform everything. Again, a time is more valuable than one-off $$$

[SilverSolder]

Some Photoshop filters will use multiple cores, if I remember correctly (will have to look next time I use it!).

Some software is very good at using multiple cores, e.g. ray tracing graphics software will typically take everything you give it, and make the coffee boil!

Another good reason to have many cores and a lot of RAM is if you use VMs on your PC to run many things concurrently, even if each task is single threaded.

[olkipukki]

Branded:

There is something you cannot DIY or buy off shelf. No, I don't mean Apple 

AMD Threadripper Pro 39nnWX

https://www.lenovo.com/us/en/thinkstation-p620

[bd139]

You can buy it if you have a disty contract. I can get one. However the price is “phone and ask” which I’m not going to do 

I may buy a Lenovo workstation next time actually. Prices are pretty good and Lenovo support is actually ok here. If you buy anything off their shop it’s fulfilled by digital river who accept PayPal and don’t even bother to respond to complaints and refunds so all the stuff is free 

[olkipukki]

Let's assume you phoned and asked  ... and a price has satisfied 

What you will do then?
There is a special socket for this chip in my understanding 

[Mechatrommer]

So far, other than databases, web servers, or concurrent anything, I haven't seen good use of lots of cores.

how parallel processing can be a good use to database or web server applications? when the bottleneck is disk access and data throughput?

This is where comparing raw benchmark numbers will fail you.

it will fail you, not most of us... benchmark is made for commonly used applications by majority of people, such as 3d graphics, single disk speed access and some other SW that can take advantage of parallel processing.

In case of Photoshop, applying filters where processes are repeated and few others, jobs appears to run on a single core.

photoshop is a simple process bottlenecked by user interaction, much like when people typing on Ms Word or forum here. it will not worth an effort to complicate the code for multithreaded process. batched based SW can take advantage of this. i built a simple tool, much simple than photoshop that uses all cores to maximum to process many pictures in short time, but then mostly my hand is a lot slower at editing than a single core processing a picture that is sent to batch processor. other heavy tools such as simulation (emsolver, fem, physics etc), CAD, videos and processes that are measured in minutes, hours or days will surely maximize the parallel processing as much as possible. we love high processing clock, but we cant have or wait 6 or 10GHz cpu anytime soon.

https://www.pugetsystems.com/labs/articles/Adobe-Photoshop-CC-Multi-Core-Performance-625/

[bd139]

Let's assume you phoned and asked  ... and a price has satisfied 

What you will do then?
There is a special socket for this chip in my understanding 

Apparently it's orderable. I didn't check board. Think it needs liquid cooling as well. They will sell you this shit even if it's useless 

[NiHaoMike]

I just opened your photos: Mike, why do you have a massive computer grade capacitor inside?

If you haven't noticed it, it has both AC and DC power inputs. The huge capacitor keeps the voltage steady when handing off operation between the two converters. It was designed with lead acid in mind since that was the most economical at the time, but with just a little reprogramming of the PIC handling the power management, it now runs on a LiFePO4 pack.

Ideally, I would have a circuit to very quickly drop the CPU and GPU frequencies to minimum when operating on backup power, but since the only way to do that would involve modifying the motherboard, I settled for a slightly slower software implementation instead.

[rsjsouza]

Thanks Mike.