Please recommend the configuration of the computer workstation

  • 291 Views
  • Last Post 12 June 2018
  • Topic Is Solved
h395523899 posted this 24 May 2018

I need to buy a computer workstation for fluid and static simulation.

Is there a good configuration recommendation?

Please tell me what factors you need to consider when configuring your computer workstation. I will answer them one by one. Thank you

Order By: Standard | Newest | Votes
peteroznewman posted this 24 May 2018

What is your budget?  There are different recommendations if you are a poor student versus having some else pay for an awesome workstation.

Which license are you on: Student (limited size) or a Research/Corporate (unlimited size) license?  You don't need as much RAM if you can't run large models.

Does it need to be mobile or can you get a desktop?

The most important thing is to get a graphics card from the list of tested cards.  That will ensure a good user experience with the product.

Get at least 8GB of RAM.  Get more if you can afford it.  Find out the maximum RAM capability of the motherboard, even if you don't install the maximum initially.  This allows for a future RAM upgrade, say when you change from a Student license to a Research license.

Get the fastest processor with 4 or more cores if you can afford it. However, don't trade off clock speed to get more cores. Some vendors offer well engineered overclocked processors such as BOXX.

Get SSD storage instead of HDD if you can afford it.  They are much faster on large problems when you run out of RAM. I have an SSD for the C: drive so programs load fast, a much larger SSD where the projects I am working on are solved. Use a HDD for archive storage that is really large.

Get a supported GPU if you can afford it.  But my testing showed GPUs were not accelerating my Static Structural models. They do help on FLUENT models.

 

  • Liked by
  • raul.raghav
  • h395523899
h395523899 posted this 26 May 2018

1. My budget is less than $7,000.

2. How to check the license type? Student (limited size) or a Research/Corporate (unlimited size). I do not understand this

3. Desktop workstation

4. I am going to use ssd and hdd in combination.

5.The number of cores and the size of RAM made me entangled. The site suggests e5 2600 v4 cpu, I do not know how the performance.

The maximum size of the fluent model I have established can reach 100m*100m*50m, and the model will also include the 0.01mm class. So the dimensions in the model range from 100m to 0.01mm. So my grid should reach millions or even millions.

The structural mechanics model size is about 100m*5m*5m

 

This is the current computer configuration

CPU: Core i7-4790

RAM: 4G

HDD: 1T

Graphics: NVIDIA GeForce GT 720

When I draw structural grid of more than 100,000, the computer becomes very slow

peteroznewman posted this 26 May 2018

 1. A respectable amount to get a useful computer.

2. If you downloaded the free ANSYS Student software from the link at the top of this page, you are on a Student license. But you can't run models the size you mention on this license type. The Student license will not solve structural models when nodes and element exceed 32k or FLUENT models when nodes and cells exceed 512k. You must be on a Research license.

3. You get more compute power for your money on a desktop.

4. Good choice. One issue is that the SSD will be smaller than the HDD, but you want to solve on the SSD to pick up the speed advantage. However, on a really big transient model, the disk space requirements can be huge and you don't want to run out of disk space before the solution completes.

5. See my lengthy answers below. I had to break them into smaller posts because of a 3000 character limit on each post.

 

  • Liked by
  • h395523899
peteroznewman posted this 26 May 2018

CLOCK SPEED AND CORE COUNT

Your current computer has a Core i7-4790 processor, which has 4 cores at 3.6 GHz.  If you purchase a computer with a 1.8 GHz clock, your new computer will take twice as long to do most things than your current computer, which would be disappointing even if it had 14 cores and solved CFD models faster.

"The site suggests e5 2600 v4 cpu" Look at this reference.

There are 29 products that are e5-2600-v4, out of the 44 products in the e5-v4 family. Below I plotted all 44 processor with the Clock frequency (GHz) versus the number of Cores. The ones in blue are in the 2600 family. I have labeled two processors. Pretend those two processors cost the same and you have to choose which one to purchase. One has 6 cores running at 3.7 GHz while the other has 12 cores running at 1.9 GHz.

The clock speed determines how fast a given program will run. If you have a 1.8 GHz computer and a model runs in 2 hours, if you move it to a 3.6 GHz computer, it would run in 1 hour (they both have 6 cores). You know the 3.6 GHz computer is more expensive than the 1.8 GHz computer.  Now you could choose a third computer that has 12 cores running at 1.8 GHz that costs the same as the 6-core, 3.6 GHz computer. Do you go for more cores or a faster clock? Read the next post to learn that structural models don't scale perfectly with cores, and most of the programs running on your computer don't know how to use multiple cores, so I always go for clock speed over core count.

Some manufacturers, like BOXX, will use high-performance cooling on the processor and overclock the processor beyond what Intel specifies for ordinary cooling. A BOXX Apex S-class is plotted on the graph and you see how much higher the clock is set compared with the standard products. The BOXX Apex will feel 33% faster for any program that you are currently using on your computer. A pdf file from BOXX of a computer with an ANSYS-tested graphics card that costs < $7000 is attached for your reference.

Attached Files

  • Liked by
  • h395523899
peteroznewman posted this 26 May 2018

 ANSYS PERFORMANCE SCALES WITH CORES

If you solve a structural model on 2 cores and it takes 19 hours, it might take only 9 hours on 4 cores, 5 hours on 8 cores and 3 hours on 16 cores. Do you see the diminishing returns of adding more cores? Doubling the number of cores doesn't halve the time.

Those numbers are one illustration and different models will have tremendous variation away from the ideal of halving the time when doubling the cores. The solver even estimates how well the model is balanced and warns you if it is not well balanced. FLUENT models respond closer to the ideal of halving the solve time when the number of cores is doubled.

Here is a link for how to setup Mechanical to use more cores. I hope you have done this on your 4-core computer. On Fluent, you check Parallel and type a number when you start building the model. 

ANSYS PERFORMANCE IS BEST WHEN SOLVING IN MEMORY

When making a purchase decision, you want the model to run entirely in RAM if possible. The model that ran for 5 hours on 8 cores may have needed 64 GB of RAM to run entirely in memory. If that model ran on a computer with only 32 GB of RAM, it might take 11 hours. ANSYS uses the term "in-core" to mean running in memory (RAM), but "in-core" has nothing to do with the number of cores.

RAM IS BETTER THAN CORES

For this model you will be better off with 4 cores and 64 GB of RAM rather than 8 cores and 32 GB of RAM. Doubling RAM is much better for the solution time than adding cores, but this only applies if the solver needs a large amount of RAM.  If the solver only needs 16 GB of RAM, it won't run any faster on a computer with 64 GB of RAM. RAM is cheaper than cores so my advice is to install the maximum amount of RAM that the motherboard supports.

  • Liked by
  • h395523899
peteroznewman posted this 26 May 2018

UPGRADE PATH - ADDING A GPU

If you have two open PCI slots in the computer, one of which is a PCI Express 3.0 x16 slot, you can plug in a GPU at a later date to reduce the solve time on FLUENT models. The NVIDIA QUADRO GP100, which supports GPU solver acceleration on both FLUENT and Mechanical models, costs nearly $7000 just for the card. That card adds 3,584 Cuda cores, which is why it is useful on large CFD models. But before you buy a computer with a plan to add a GPU later, make sure that the power supply has the capacity to power the GPU. That GP100 card needs 235 W so you have to know that the power supply in the computer has that unused capacity. If you purchase a computer with a 300 W power supply and add up how much is used by all the components it came with, there may only be 100 W of unused capacity, so this GPU would not work in such a computer. The BOXX has a 650 W power supply and is engineered to support multiple GPUs.

UPGRADE PATH - ADDING A SECOND PROCESSOR

The BOXX computer I show can’t hold more than 64 GB of RAM, so you wouldn’t be able to go to 128 GB on that motherboard. A different model of motherboard must be chosen if you want the option to expand the RAM to that level. Some motherboards support two processors. You can buy a computer initially with one 6-core processor, and at a later date, purchase a second 6-core processor, but the cooling system has to be able to support two processors. Some motherboards have the capacity to hold more than 64 GB of RAM, but only if you have two processors, so you have to read the fine print.

  • Liked by
  • h395523899
h395523899 posted this 27 May 2018

Thank you so much for answering my question with such a length

Through your answer I probably understand the following points

1. At present, the configuration of my calculation is still available, but it is only necessary to increase the RAM, which can increase the speed of calculation.

2. The memory limits the number of grids, the CPU limits the calculation speed, and at the same time, the memory speed may affect the calculation speed.

 

 

The workstation I want to buy is two E5 2600 V4 CPUs.

12 or 16 cores.

RAM: 32-64g.

2T HDD, 512G SSD

peteroznewman posted this 27 May 2018

If you increase RAM on your current computer, the elapsed time to wait for the solution will be reduced on large models, but will have no effect on small models. On a structural model, under Analysis Settings, Solver Controls, set the Solver Type to Direct. You can easily tell if the solver is running entirely in memory by looking at the solution output. Click on the Solution Information folder in the Outline and the Worksheet will show the Solver Output. Click on the text and type ctrl-F to  search for allocated.

You can see that the model above was allocated nearly 14 GB of available RAM and ran entirely in that memory block.

A much larger model will show this note.

You can see that the model above needed 283 GB to run entirely in RAM and that was not available, so the solver will use the storage to hold parts of the matrix while it is computing the solution.

2. The memory limits the number of grids, the CPU limits the calculation speed, and at the same time, the memory speed may affect the calculation speed.

The example above shows that the memory didn't prevent the solver from running, it just affects how it runs. Running out of disk space will prevent the solver from running.

Your proposed computer configuration shows only one SSD. You want to configure the solver to use that drive while solving. The simplest way to do that is to store your model there. But if that is the C: drive, the installed programs use up some of that space leaving less available for the solver. You don't ever want to run out of disk space during a solve. Another issue with having only one SSD is the Windows OS needs to write files to the disk at random times, therefore, it is optimal if Windows is writing to a different drive than ANSYS. That is why I recommended two SSDs. One for the C: drive, which could be 256 GB, and a second drive, 512 GB, just for ANSYS models to solve on, while the third HDD is for moving old models from the SSD to the HDD so you don't run out of disk space on the SSD during a solve.

MECHANICAL SOLVERS

If you leave Solver Type to Program Controlled, ANSYS may choose the Iterative solver instead of the direct sparse solver. The iterative PCG solver might solve the same model in more or less time. If you are going to be solving almost the same model many times, say in a parameter study, you may want to solve the exact same model twice, once with Direct and once with Iterative. Scroll to the end of the Solution Output and note the Elapsed Time for each solution, then you can select the solver that took less time for the rest of your study.

  • Liked by
  • h395523899
peteroznewman posted this 27 May 2018

When I draw structural grid of more than 100,000, the computer becomes very slow

Do you mean updating the display during view operations like pan, zoom and rotate becomes sluggish?
Or do you mean just generally waiting a long time for the mesh to be computed, or the model to save?

A better graphics card can help the first problem. Your current GeForce card has only 1 or 2 GB of graphics RAM and only 192 Cuda cores to compute display rotation. A better graphics card such as the NVIDIA QUADRO P2000 has 5 GB of graphics RAM and 1024 Cuda cores to compute display rotations.

The QUADRO series of cards cost more than an equivalently featured GeForce card, but the difference is in the drivers. The GeForce drivers were tested on games and any problems were fixed so the games ran well. The QUADRO drivers were tested on CAD and CAE programs like ANSYS and any problems were fixed so those professional programs ran well.  You can see several posts on this site about ANSYS not properly displaying things on graphics cards that are not on the tested list.

The second problem of waiting a long time for the mesh to build is somewhat unavoidable. But if Windows is running low on RAM during the meshing process, having more RAM can improve that problem.

You don't say what the Clock frequency in GHz is for the two processors you are considering. If you get a Clock less than your current computer's 3.6 GHz frequency, the mesh will take longer to build (assuming RAM was not a limiting factor).

You can show your appreciation by clicking Like below the posts that are helpful.

  • Liked by
  • h395523899
  • raul.raghav
h395523899 posted this 12 June 2018

Hello, can you recommend a dell or hp desktop workstation? Or own assembly parts

The boxx workstation you recommend is not available in our country.

If I am running ansys, I still need to run matlab. Is it possible to pursue high clock time?

Similar recommended boxx configuration 4.8 Hz, six cores. A cpu can satisfy it?

I have two plans now

1. A cpu above 4.0 Hz. More than six cores. This may only be selected Core Series. For example: i7-8700k

2. Both cpu over 3.5 Hz. More than four cores each. This can choose Xeon series. E.g:

Xeon Gold 5122

Xeon E5-2637 v4

Which is better?

peteroznewman posted this 12 June 2018

Overclocking is not done by Dell or HP, they run with the specification that Intel provides, which is okay. Smaller companies might build overclocked machines for Gamers, but I worry about the reliability of their solution, I don't want the CPU to overheat and shutdown and lose tens of hours of computation. So forget overclocking.

i7-8700k  - 8th Generation (Coffee lake) is limited to 128 GB of RAM but is much cheaper than the Xeon processors.

Xeon Gold 5122 - 6th Generation (Skylake)  is limited to 768 GB of RAM

Xeon E5-2637 v4 - 5th Generation (Broadwell) is limited to 1.54 TB of RAM.

 

My HP research showed I can't stay under your budget on a Xeon Gold 5122 with 2 processors of 4 cores each.

But on the i7 processor, here is a sweet $6,697 + tax computer that hits all the right places.

HP Z4 G4 Workstation

Intel® Core™ i7-7820X Processor
(3.6 GHz, up to 4.3 GHz w/Turbo Boost, 11 MB cache, 8 core)

Z4 G4 1000 W Chassis for Core®-X Processors

128 GB (8x16 GB) DDR4-2666 Memory

512 GB HP Z Turbo Drive TLC M.2 NVMe SSD

1 TB HP Z Turbo Drive Quad Pro SSD

6 TB 7200 RPM SATA Enterprise 3.5" 1st HDD

NVIDIA® Quadro® P4000 (8 GB GDDR5, 4 x Displayport 1.4) Graphics

 

If including the tax puts you over the top of your budget, you can scale back the NVIDIA card to a P2000 and save $365. You can scale back the RAM to 64 GB and save $1850. You can even buy it with 4 slots open and fill the open RAM slots later when you have more money. This workstation can also support the NVIDIA GP100 GPU if you have an extra $7k to spend.

h395523899 posted this 12 June 2018

Thank you, for your configuration, I look at our domestic prices.

Regarding the issue of hyperband, I really did not consider it.

How does the i7-8700 compare to the i7-7820X?

 

About Gold 5122

Dell 7920 chassis, two Gold 5122 cpu, 1T SSD, 512G SSD, 64G RAM, 2T HDD, P2000

  $7652.9019 including tax

How do you think about this configuration

peteroznewman posted this 12 June 2018

Benchmarks vary comparing i7-7820X to the i7-8700K.

One site showed the 7820X faster than the 8700K for Ray Tracing.

Another site showed the same advantage for multicore processing, and shows the 7820X with 8 cores vs 6 cores on the 8700K.

The Dell 7920 is a good choice, and a good price. I can't navigate the Dell website to configure the hard drives, it keeps flagging problems with the configuration. The HP website is much easier to configure a custom system.

h395523899 posted this 12 June 2018

Xeon w-2125 w-2145 w-2155 i9-7900x i9-7940x

How about these cpu? I see the w series is also very cheap

peteroznewman posted this 12 June 2018

These are all good processors, don't worry too much about that now. You are in the right "ballpark". Fast clock, 8 cores, plenty of RAM, SSD storage and the right graphic card, with room for expansion to a GPU when you have more cash. Focus now on the quality of the support. How easy is it to find updated drivers on the vendor's website? How long is the warranty? How loud are the cooling fans?

If you get a computer with 8 slots for RAM but you only purchase 64 GB, do you leave 4 slots empty for a later upgrade or fill all 8 slots? Memory access is faster when all the slots are filled, but that means you throw out your 8 x 8 GB SIMS and purchase 8 x 16 GB SIMS if you upgrade from 64 GB to 128 GB.

Is the power stable and reliable in your area?  Should you purchase a UPS to keep your computer going when the power goes out?  How long do you want that to keep it running, 10 minutes or 10 hours?

Close