jaquith
Glorious
- Feb 8, 2009
- 14,235
- 27
- 49,290
Q - Does the simulation have a large amount of data?
First both your CPU (Xeon, Opteron and many AMD CPU's) and MOBO must support ECC RAM (RDIMM, UDIMM) per supported type(s).
My take on ECC is simple, if the end product is for final production i.e. commercial use then always choose ECC. Some of my friends want even their home PC to have server/workstation components. There's no downside.
Sure you can have an error or errors after 6-10 hours, ECC is important if error-free is important.
Example 100% of my SQL and all servers and workstations are all running Xeon + ECC. I do Enterprise SQL data including CRM, REO, TAX, GIS, IDX, etc data processing with huge amounts of data where just by the sheer volume there would be errors 'guaranteed' i.e. corruption of time with non-ECC.
First both your CPU (Xeon, Opteron and many AMD CPU's) and MOBO must support ECC RAM (RDIMM, UDIMM) per supported type(s).
My take on ECC is simple, if the end product is for final production i.e. commercial use then always choose ECC. Some of my friends want even their home PC to have server/workstation components. There's no downside.
Sure you can have an error or errors after 6-10 hours, ECC is important if error-free is important.
Example 100% of my SQL and all servers and workstations are all running Xeon + ECC. I do Enterprise SQL data including CRM, REO, TAX, GIS, IDX, etc data processing with huge amounts of data where just by the sheer volume there would be errors 'guaranteed' i.e. corruption of time with non-ECC.
ECC often comes hand in hand with registered memory. If you want to run a system with more than 64GB of memory you almost certainly will need to use RDIMMs and thus have ECC by design.
Statistically 10% of all memory modules will experience a memory error at least once per year. Of those modules that do experience errors 90% of them will experience more than one. The error rate compounds with the number of DIMMs installed in a system so holding all things equal a system with a single DIMM has a 1=% chance of experiencing a memory error. A system with two DIMMs has a 19% chance, a system with 8 has a 57% chance. A system with 16 DIMMs has an 81.5% chance of having a memory error.
The more memory there is installed in a system the more important ECC becomes. ECC can detect and correct single bit errors as well as detect but not correct multi-bit errors. Since there's about a 90% chance of an error reoccuring within a relatively short timeframe its very important that the bad module be replaced ASAP. The ability to detect and report module errors is a huge help. Imagine how many hours some people spend trying to find the source of persistent BSODs and multiply that by a datacenter full of modules.
As far as 6-10 hour simulations go it all depends on the consequences of experiencing a memory error during the simulation. If your simulation is error tolerant and/or self correcting it may not be necessary. If there's no consequences to your simulation blowing up, it's not necessary at all. From a purely professional standpoint ECC RDIMM memory should be used because it's honestly not that much more expensive than regular non-ECC UDIMMs
Statistically 10% of all memory modules will experience a memory error at least once per year. Of those modules that do experience errors 90% of them will experience more than one. The error rate compounds with the number of DIMMs installed in a system so holding all things equal a system with a single DIMM has a 1=% chance of experiencing a memory error. A system with two DIMMs has a 19% chance, a system with 8 has a 57% chance. A system with 16 DIMMs has an 81.5% chance of having a memory error.
The more memory there is installed in a system the more important ECC becomes. ECC can detect and correct single bit errors as well as detect but not correct multi-bit errors. Since there's about a 90% chance of an error reoccuring within a relatively short timeframe its very important that the bad module be replaced ASAP. The ability to detect and report module errors is a huge help. Imagine how many hours some people spend trying to find the source of persistent BSODs and multiply that by a datacenter full of modules.
As far as 6-10 hour simulations go it all depends on the consequences of experiencing a memory error during the simulation. If your simulation is error tolerant and/or self correcting it may not be necessary. If there's no consequences to your simulation blowing up, it's not necessary at all. From a purely professional standpoint ECC RDIMM memory should be used because it's honestly not that much more expensive than regular non-ECC UDIMMs
Is there any way of putting in ECC RAM at an affordable price? I mean, because from what I understand you can't use an i7 processor, you have to use a xeon. Are there any other options for a rig that will cost around 3000 USD (should include a 500 gb ssd, 1tb hard disk and an nvidia quadro 4000 at the minimum)?
jaquith
Glorious
- Feb 8, 2009
- 14,235
- 27
- 49,290
That may seem like a lot, but the 'small' testing SQL I run uses 20GB+ memory and the (textual) data is 6GB+ with 100's of thousands of records with 25~100 fields per record.
A Xeon Workstation can be done with your budget, but as in any build some sacrifices must be made. The most important choices from the beginning are both the CPU and the GPU. The cost of a 500GB SSD isn't helping your budget.
Workstation:
$786 Kingston SSDNow KC100 Series 480GB 2.5" SATA III Enterprise SSD SKC100S3/480
$705 PNY VCQ4000-PB Quadro 4000 2GB Workstation Video Card
$110/ea Western Digital RE4 WD1003FBYX 7200/64MB 1TB Enterprise Hard Drive (recommend RAID 1, 5 or 10; 2~4 drives)
======
$1,601 sub-total
Problem - there's a huge cost difference between the E5-2650 8-core vs E5-2620 6-core, yep the 8-core is worth it but kills your budget all to hell. PassMark 12927 vs 8983 Then benches like Cinebench depends on the CPU choice, single or MP and then which CPUs.
$1,100 Intel Xeon E5-2650 8-core
vs
$415 Intel Xeon E5-2620 6-core
$330 SUPERMICRO X9SRA Single Socket R 256GB (RDIMM), 64GB (UDIMM), 2 (x16) PCI-E 3.0
or MP:
$500 SUPERMICRO X9DAi Extended 512GB (RDIMM), 128GB (UDIMM), 3 x PCI-E 3.0 x16 slots
RAM (depends on the MOBO) in general, single CPU 4x16GB RDIMM 1.50v or 1.35v which costs ~$700 per 64GB otherwise 4x8GB costs $300 per 32GB.
+ PSU
+ Case
+ OS
+ ODD, etc
======
$????? Depends on the choices.
Storage Options:
SSD Consumer:
$710 Intel 520 Series Cherryville SSDSC2CW480A310 2.5" 480GB SSD
$700 SAMSUNG 830 Series MZ-7PC512B/WW 2.5" 512GB SSD
HDD quality alternatives:
$300/ea Western Digital VelociRaptor WD1000DHTZ 1TB 10,000/64MB (recommend RAID 1, 5 or 10; 2~4 drives)
$110/ea Samsung Spinpoint F3 HD103SJ 1TB 7200/32MB (recommend RAID 1, 5 or 10; 2~4 drives)
The i7 and comparable Xeon processors are priced similarly, usually within a few dollars of eachother. The really expensive processors are the dual and quad socket capable processors (DP/MP), and octal core processors. Xeon processors that support only a single socket (UP) are marginally more expensive than their desktop counterparts. The same is true for ECC RDIMMs, similar capacities are similarly priced but higher capacities which aren't available in UDIMM are much more expensive.
http://ark.intel.com/compare/64601,63697
Wow, so why would someone buy the 3930k then? Are there some compatibility or building issues with having an e5-1650 on a server board vs an i7 3930k on an intel extreme atx board?
Also, earlier I was talking about the i7-3770k (3730k was a typo), do you know what the Xeon equivalent for that is?
The E5-1650 can only be used in a server or workstation motherboard that supports Xeon processors. The 3930k can be used in almost any socket 2011 motherboard. Additionally the 3930k has a bit more cache and can be overclocked as Xeon processors typically have locked multipliers.
The IvyBridge based xeons are similarly named as the SandyBridge (non-E) based xeons but have an added v2 at the end. The analog for a 3770 (non-k) would be Xeon E3-1270v2 or E3-1275v2 if you also want the IGP. The Xeon E3-1270 and E3-1275 are the analogs for the i7-2600
jaquith
Glorious
- Feb 8, 2009
- 14,235
- 27
- 49,290
ECC.
Further, since the entire discussion is ECC and I mentioned RDIMM (Registered memory) which is only ECC I am confused that you're confused. The build posted was specifically meant as a Workstation (Xeon, enterprise SSD/HDD, Pro GPU).
The problem is your budget for the components you listed even using consumer grade components (SB-E i7-3930K, SSD, HDD, GPU, etc).
- Status
TRENDING THREADS
-
-
Question Tried moving SSD from old PC to new one, now it won't boot from either
- Started by scyer327
- Replies: 8
-
Question why my pc dont have sound after install new Nvidia GPU Card?
- Started by sudsaviour
- Replies: 2
-
News Windows 11 will reportedly display a watermark if your PC does not support AI requirements- Started by Admin
- Replies: 10
-
Facebook
Twitter
Instagram