Workstation for scientific computing (any Threadripper owners out there?)

[Luis_V]

Budget (including currency): 8.000 CAD (it depends on how effective the PC is for our purposes, I will have to check with my supervisor)

Country: Canada

Games, programs or workloads that it will be used for: Mainly used for Maxquant software (I explain in detail below)

The build I am considering: THINKSTATION P620 WORKSTATION (found this workstation in a video from LTT)

AMD Ryzen™ Threadripper™ Pro 3955WX Processor (3.90 GHz, up to 4.30 GHz Max Boost, 16 Cores, 32 Threads, 8 MB Cache)

8x 16 GB DDR4 3200MHz RDIMM ECC (total of 128 GB)

NVIDIA® Quadro® P620 2GB

256 GB M.2 Gen 4 PCIe SSD, OPAL

Tower WRX80 92% Power 1000W

Windows 10 Pro 64

$6,330.12 CAD according to Lenovo website

 

Other details: I am a PhD student in a lab where we use a technique called proteomics. In summary, every function in your body depends of a wide variety of proteins. A couple of years ago we would look at just a few of them at a time. Now, with proteomics, we analyze the samples in an equipment called mass spectrometer, which has a high sensitivity and is capable of quantifying thousands of proteins at the same time (super cool eh?). The problem is that this equipment generates a bunch of measurements/graphs that need to be translated back to proteins, and this is what we will do in the workstation. There are a variety of softwares capable of doing this, the one we use in our lab is called Maxquant. This programs depends on the number of cores and the memory in a PC. More specifically, 4 GB of RAM per thread. Here are the requirements according to the developer's website: http://coxdocs.org/doku.php?id=maxquant:common:download_and_installation. My supervisor also said that having a SSD makes a huge impact, probably due to the writing speeds. Currently, we are using an Intel i7-6700, 16 GB of ram, and an "SK hynix sc311 sata 512gb" (this a Dell system). With this build, the analysis can go from 2 to 4 days (or even more), depending on the design of the experiment.

 

We are looking for a new build, and I suggested my supervisor that we buy a Threadripper system with lots more threads and memory to speed up our pipeline. The biggest problem is that this will not be a cheap computer, and I am afraid it will not translate to a processing time that is noticeably faster. Therefore, here are some questions:

- I am afraid the SSD may be a bottle neck for the system. Is that something I should worry about?

- I don't know much about ECC and non-ECC memory, will it have a major impact in the performance?

- Most importantly, is there any way that I can test before buying one? Any Threadripper owners out there willing to test it for us?

 

Thank you in advance!

[brob]

The machine looks quite capable. My only thought is that the NVMe (m.2) drive seems a bit small. The PCIe 4.0 interface is the fastest generally available for this class of machine.

 

ECC (error correction code) memory does run slower than non ecc. However it is able to correct single bit errors without halting the system and detect multi-bit errors, preventing data corruption. For the work described ecc memory is a good idea. The slightly lower raw memory performance is generally worth the added data integrity.

 

The very limited requirements documentation suggests that the program can make use of most if not all the  threads available. Without knowing more about the software there is no way of knowing how efficiently the application scales. Presuming a high degree of parallellism and data sets that can use that parallelism, I think it is safe to assume that a fourfold increase in processing threads should offer a significant performance improvement. 

[SupaKomputa]

$6000 for 16 cores and a $300 quadro gpu doesn't add up for me.

Yes ecc will have performance toll a bit.

Below is a much better config for the same price with 8gb quadro card and 32 cores, aside from the ECC memory.

 

PCPartPicker Part List

CPU: AMD Threadripper 3970X 3.7 GHz 32-Core Processor  ($2699.00 @ Canada Computers) 
Motherboard: Gigabyte TRX40 AORUS PRO WIFI ATX sTRX4 Motherboard  ($635.80 @ Vuugo) 
Memory: G.Skill Trident Z Royal 128 GB (8 x 16 GB) DDR4-3200 CL16 Memory  ($1109.98 @ Newegg Canada) 
Storage: Samsung 980 500 GB M.2-2280 NVME Solid State Drive  ($95.99 @ PC-Canada) 
Storage: Seagate BarraCuda 4 TB 3.5" 5400RPM Internal Hard Drive  ($94.99 @ Memory Express) 
Video Card: PNY Quadro RTX 4000 8 GB Video Card  ($1399.00 @ Canada Computers) 
Case: Lian Li PC-O11 Dynamic ATX Full Tower Case  ($184.99 @ Memory Express) 
Power Supply: Corsair RMx (2021) 1000 W 80+ Gold Certified Fully Modular ATX Power Supply  ($259.99 @ Best Buy Canada) 
Total: $6479.74
Prices include shipping, taxes, and discounts when available
Generated by PCPartPicker 2021-05-22 23:23 EDT-0400

[Luis_V]

On 5/22/2021 at 8:47 PM, brob said:

The machine looks quite capable. My only thought is that the NVMe (m.2) drive seems a bit small. The PCIe 4.0 interface is the fastest generally available for this class of machine.

 

ECC (error correction code) memory does run slower than non ecc. However it is able to correct single bit errors without halting the system and detect multi-bit errors, preventing data corruption. For the work described ecc memory is a good idea. The slightly lower raw memory performance is generally worth the added data integrity.

 

The very limited requirements documentation suggests that the program can make use of most if not all the  threads available. Without knowing more about the software there is no way of knowing how efficiently the application scales. Presuming a high degree of parallellism and data sets that can use that parallelism, I think it is safe to assume that a fourfold increase in processing threads should offer a significant performance improvement. 

Thank you! The SSD is quite small indeed. We generally save every file in a backup drive, but doubling that SSD is a good idea for sure.

 

Sometimes the program does crash for no clear reason, maybe it is a problem that ECC memory could correct? Just guessing though, and it wouldn't matter much as these crashes are not too often.

 

I understand it is hard to predict the amount of improvement. I really needed someone with such build to test if the difference in performance would be significant. I think it will be a bit hard to convince my supervisor to spend this much money when he could spend the same amount on 3 PC's and run 3 experiments at the same time... 

 

On 5/22/2021 at 9:25 PM, SupaKomputa said:

$6000 for 16 cores and a $300 quadro gpu doesn't add up for me.

Yes ecc will have performance toll a bit.

Below is a much better config for the same price with 8gb quadro card and 32 cores, aside from the ECC memory.

 

PCPartPicker Part List

CPU: AMD Threadripper 3970X 3.7 GHz 32-Core Processor  ($2699.00 @ Canada Computers) 
Motherboard: Gigabyte TRX40 AORUS PRO WIFI ATX sTRX4 Motherboard  ($635.80 @ Vuugo) 
Memory: G.Skill Trident Z Royal 128 GB (8 x 16 GB) DDR4-3200 CL16 Memory  ($1109.98 @ Newegg Canada) 
Storage: Samsung 980 500 GB M.2-2280 NVME Solid State Drive  ($95.99 @ PC-Canada) 
Storage: Seagate BarraCuda 4 TB 3.5" 5400RPM Internal Hard Drive  ($94.99 @ Memory Express) 
Video Card: PNY Quadro RTX 4000 8 GB Video Card  ($1399.00 @ Canada Computers) 
Case: Lian Li PC-O11 Dynamic ATX Full Tower Case  ($184.99 @ Memory Express) 
Power Supply: Corsair RMx (2021) 1000 W 80+ Gold Certified Fully Modular ATX Power Supply  ($259.99 @ Best Buy Canada) 
Total: $6479.74
Prices include shipping, taxes, and discounts when available
Generated by PCPartPicker 2021-05-22 23:23 EDT-0400

Thank you for taking the time to build this list, I appreciate it.

 

I don't know how much we want to spend on a graphics card since it doesn't seem to impact our pipeline. Currently, our PC does not have a dedicated GPU, so it is very likely that, for now, we will be fine with a cheap GPU. However, I do get your point that this GPU specifically is expensive for the model.

 

I really like you build, although to get the full power for my application I would have to double the RAM (4 GB per thread) or get the 16-Core processor. I will also check with my supervisor if we can buy the parts separately or if the PC has to come ready to run. This money is stored on a University account, so they may require to buy pre-build computers to avoid possible frauds. Anyway, thank you very much for your input!

[SupaKomputa]

OK that's lovely, loose the GPU so you can have 32gb ram for 256gb total.

You can build this with custom prebuild service such as

https://www.ibuypower.com/Store/AMD-Ryzen-Threadripper-3rd-Gen-Configurator

[brob]

7 hours ago, Luiz.Almeida said:

Sometimes the program does crash for no clear reason, maybe it is a problem that ECC memory could correct? Just guessing though, and it wouldn't matter much as these crashes are not too often.

 

In the PC world the causenof a crash can be difficult to determine. Memory errors are one source, but there are other possibilities.

 

Multi-bit memory errors may not be detected with non ecc memory. This can lead to errors in data.