MP 7,1 Ram: Samsung vs Crucial

[Z28McCrory]

Yup ~ My second RAM order from Crucial is on backorder. The email stated that they 'hope' to send out in 7 days.

Yeah they originally told me 7 days (within a few hours of initially placing the order on the 19th), but today customer service told me they didn’t have an ETA.

I wasn’t expecting to get my 7,1 until the 30th, but now it’s looks like the 26th. So I canceled the Crucial order and placed a Newegg order of the SumerMicro strips with overnight shipping

 

[moab1]

I assume these will work fine?

Supermicro (MTA36ASF4G72PZ-2G9E2) 32GB SDRAM ECC Registered DDR4 2933 (PC4 23400) Server Memory Model MEM-DR432L-CL01-ER29 - Newegg.com

Buy Supermicro (MTA36ASF4G72PZ-2G9E2) 32GB SDRAM ECC Registered DDR4 2933 (PC4 23400) Server Memory Model MEM-DR432L-CL01-ER29 with fast shipping and top-rated customer service. Once you know, you Newegg!

www.newegg.com

The Micron original part number for this particular SuperMicro DIMM is MTA36ASF4G72PZ-2G9E2

The Micron original part number is the same as the one you have listed in the chart (and even what Apples uses as OEM), except this one has an "E2" at the end of it.

The specs all appear to be the same though.

I originally placed an order at Crucial, but then received an email saying it was on backorder. I chatted with custom service today and they said they have no ETA on the next stock.

I'm hoping since Micron makes the Crucial memory, this is essentially the same thing?

Did anyone confirm if these work for sure?

 

[Z28McCrory]

Did anyone confirm if these work for sure?

I had HOPED to be able to test them yesterday, but Newegg dropped the ball and sent them ground, despite me paying for 1 day.

I've been waiting on them all day today, and just got a delivery attempt failed notification from UPS (even though I'm in the room next to my front door... I have no idea how UPS left the card on my door without my dog freaking out).

Anyway (sorry for the rant, lol), I'll have them in my hands on Monday hopefully, and will be sure to report back.

 

[Z28McCrory]

Did anyone confirm if these work for sure?

Just finished installing this RAM. Happy to report that it works fine. Supermicro (MTA36ASF4G72PZ-2G9E2)

It shows up at Micron

 

[barnyard]

Would either of these work? (not my auctions!)

64GB DDR4-2933Mhz RDIMM Memory TSV MEM-DR464L-HL02-ER29 MEM-DR464L-CL01-ER29 | eBay
These are TSV RDIMMs They cannot be mixed with regular RDIMMs! (They mix well with 128GB RDIMMs which are by default TSV). 64GB 1x64GB DDR4-2933, PC4-23400R 3DS (TSV) (RDIMM) Load Reduced, 1.2V, 288-pin.
www.ebay.com

64GB DDR4-2933Mhz RDIMM Memory TSV Intel Skylake Xeon Scalable, AMD EPYC CPU | eBay

These are TSV RDIMMs They cannot be mixed with regular RDIMMs! (They mix well with 128GB RDIMMs which are by default TSV). 64GB 1x64GB DDR4-2933, PC4-23400R 3DS (TSV) (RDIMM) Load Reduced, 1.2V, 288-pin.
www.ebay.com

64GB DIMMs for $175

 

[aliengirl]

I got these Samsung sticks and they are working great so far. If you ask for a coupon code via chat or email that will knock a bit off the reg $148 price. They shipped fast.

https://memory.net/product/m393a4k4...r4-2933-rdimm-pc4-23466r-dual-rank-x4-module/

 

[mward333]

I got these Samsung sticks and they are working great so far. If you ask for a coupon code via chat or email that will knock a bit off the reg $148 price. They shipped fast.

https://memory.net/product/m393a4k4...r4-2933-rdimm-pc4-23466r-dual-rank-x4-module/

@aliengirl : I had not shopped at memory.net in the past, but I saw your post and wanted to add a note about my successful purchase there too. I bought 12 of these DIMM's (for a total of 1.5 TB of RAM) and installed them in my Mac Pro 2019. They arrived promptly, and they work great. They are all made by Samsung (which is very clear from the memory.net website).
https://memory.net/product/m386aag4...4-2933-lrdimm-pc4-23466l-quad-rank-x4-module/
Hope that helps anybody who is still trying to decide what to buy, and from which seller!
I didn't know about memory.net until I made this purchase.... but they seem to be reputable.

 

[OkiRun]

Those in numbers 2 and 11 are from Crucial and are x8. All the rest are from NEMIX and are x4. Everything working fine so far.

 

[bxs]

Those in numbers 2 and 11 are from Crucial and are x8. All the rest are from NEMIX and are x4. Everything working fine so far.

View attachment 889636

What does x8 and x4 mean ?

 

[curmudgeonette]

What does x8 and x4 mean ?

The width of each DRAM chip. A rank of x8 will have eight, um nine counting ECC, chips. The x4 will need eighteen chips. If the RAMs are made at the same technology level, they'll store the same number of bits. This means the x4's have twice as many addresses, i.e. the rank is twice as big. At the same time, the internal structure might not differ much - each row has the same number of bits. This means that with x4's, the memory controller has to generate a wider column address.

 

[rylim]

I'm thinking to upgrade my MP to 384GB, and wonder about the number of sticks I should get for the best performance.

Should I get 6 x 64GB or 12 x 32GB?

 

[erroneous]

I'm thinking to upgrade my MP to 384GB, and wonder about the number of sticks I should get for the best performance.

Should I get 6 x 64GB or 12 x 32GB?

The Lenovo XeonSP memory scaling document states that memory bandwidth between a balanced 6 dimm setup and balanced 12 supposedly increases about 3%. Only you could know if your workload is sufficiently bandwidth-limited for this to make a material difference.

Balanced Memory Configurations with Second-Generation Intel Xeon Scalable Processors

Configuring a server with balanced memory is important for maximizing its memory bandwidth and overall performance. Lenovo ThinkSystem servers running second-generation Intel Xeon Scalable processors have six memory channels per processor and up to two DIMMs per channel, so it is important to...

lenovopress.com

I would worry about cost (1x64GB sticks I’ve seen are usually substantially more than x2 the cost of 2x32GB) v future expansion (will you ever want to go to 768GB?).

 

[rylim]

The Lenovo XeonSP memory scaling document states that memory bandwidth between a balanced 6 dimm setup and balanced 12 supposedly increases about 3%. Only you could know if your workload is sufficiently bandwidth-limited for this to make a material difference.

I would worry about cost (1x64GB sticks I’ve seen are usually substantially more than x2 the cost of 2x32GB) v future expansion (will you ever want to go to 768GB?).

Thanks for the info.

I afraid I need more RAM down the road, perhaps in another 5 - 6 years. I really missed the deal of Nemix 2 x 32GB RAM for below $200, and now the price is really around $300

 

[AidenShaw]

Should I get 6 x 64GB or 12 x 32GB?

I'd definitely go for 6 x 64 GiB. The extra cost now will mean that you won't have to discard the 32 GiB DIMMs down the road.

It will be almost impossible to measure a speed difference between six and twelve DIMMs on actual application.

 

[bxs]

At this time 4x32GB NEMIX (R-DIMMs)will cost you $585. So 12x32GB will cost you $1755.

Amazon.com: Buying Choices: NEMIX RAM 128GB 4x32GB DDR4-2933 PC4-23400 2Rx4 ECC Registered Memory

At this time 4x 64GB NEMIX (LD-DIMMs) will cost you $1,357. So 6x 64GB will cost you $2,036.

256GB (4x64GB) DDR4 2933 (PC4 23400) 4Rx4 288-Pin 1.2V ECC Load Reduced Memory by NEMIX RAM

256GB (4x64GB) DDR4 2933 (PC4 23400) 4Rx4 288-Pin 1.2V ECC Load Reduced Memory by NEMIX RAM

www.amazon.com

The 12x will provide 100% memory bandwidth, and the 6x will provide you with 97% of max bandwidth.

For the extra $281 it makes sense to go for the 6x 64GB LD-DIMMs that allows you to grow the RAM later on.

 

[AidenShaw]

The 12x will provide 100% memory bandwidth, and the 6x will provide you with 97% of max bandwidth.

You will never be able to measure the 3% loss of bandwidth in almost all applications - these CPUs have very large caches, so you'll only see that 3% on the small percentage of memory reads that are cache misses. A small percent on a small percent is unimportant. (The "memory bandwidth" tests in benchmarks are written to effectively disable caches - whereas applications are typically written to optimize cache use.)

In practical terms, however, filling six DIMM slots and leaving six empty means that adding memory in the future is super easy. Using twelve smaller DIMMs for the same capacity, means that you'll have to discard the small DIMMs to add memory. Penny-wise, pound-foolish.

 

[bxs]

Fluid Dynamic code knows all about L3 cache and pull in just enough to fit in the cache as it proceeds to solve a problem, and yes, so long as the needed data is in the cache the main memory performance is not an issue. However, once the data in the cache is no longer needed (has been used) and needs to be replaced with the next chunk of RAM data the memory performance becomes important. If this cache 'refresh' is done 1000s of times over several days (can be over 2 weeks) then memory speed/access times become very important. So I agree with you, except that cache misses can be very significant if there are 1000s happening all the time over several days for a long running multi-threaded code.

Yes, having to go beyond 12x 32GB will more expensive than going from 6x 64GB, but that may or may not be necessary for several years. The savings with having the former can be put to good use.

 

[AidenShaw]

If this cache 'refresh' is done 1000s of times over several days (can be over 2 weeks) then memory speed/access times become very important. So I agree with you, except that cache misses can be very significant if there are 1000s happening all the time over several days for a long running multi-threaded code.

Are you seriously claiming that a cache miss that happens every 15 minutes and is 3% slower is "very important".

Please build a system with six DIMMs and one with twelve DIMMs, and show us the timing for your two week jobs.

And, to put some data on the table - look at https://forums.macrumors.com/thread...bservations-with-various-mem-configs.1704700/ which looks at the performance of an MP6,1 with one to four DIMMs. (Spoiler - for most applications one DIMM and four DIMMs were not significantly different. Reason - cache.)

 

[bxs]

Are you seriously claiming that a cache miss that happens every 15 minutes and is 3% slower is "very important".

Please build a system with six DIMMs and one with twelve DIMMs, and show us the timing for your two week jobs.

And, to put some data on the table - look at https://forums.macrumors.com/thread...bservations-with-various-mem-configs.1704700/ which looks at the performance of an MP6,1 with one to four DIMMs. (Spoiler - for most applications one DIMM and four DIMMs were not significantly different. Reason - cache.)

Well, without going into pages and page of explanation let me simply respond with this...

The careful coding for Fluid Dynamic (FD) problems and/or Computational Fluid Dynamics (CFD) problems as an example, is needed to make full use of a processor cache to minimize memory fetches for very long running executions.

The compiler can help, but sometimes the coder has to make some changes to help the compiler to maximize the use of cache and to minimize memory fetches. Reducing memory fetches is critical for CFD code performance, and if not done carefully to be minimized will result in far too many memory fetches that are subject to the system's memory bandwidth.

My point I'm trying to make is that accessing main memory (RAM) is a very costly operation compared to say, L3 cache. Costly to the tune of close to an order of magnitude. If many costly memory fetches have to be made due to poor coding or improper compiler optimization the time for ultimate problem solution takes longer. Thus, faster memory will always be better even if one memory architecture delivers just the slightest improvement over another. For long running computations that may take weeks to complete the difference in memory performance can add to extra hours or even more days for the code to deliver its results. You will simply have to take my word for this, or else start reading some of the material listed below for confirmation.

The MP7,1 processor has a healthy L3 cache size for sure and care to keep and reuse data in the cache to avoid memory fetches is paramount for engineering/scientific codes so they can solve problems as quickly as possible. The memory bandwidth is a large factor in allowing these type codes to execute fast.

Go read some of these articles (if you're so inclined - you will need some computer science skills and a touch of High Performance Computing techniques)

https://conservancy.umn.edu/bitstream/handle/11299/158483/1/Jayaraj_umn_0130E_14121.pdf

How about L3 cache? -- CFD Online Discussion Forums

Hey, So I have tried to figure out why my (overclocked) 7600k is slower than the 8700k, at similar core-counts (also, the 8700k does not improve when

www.cfd-online.com

Building a "home workstation" for ANSYS Fluent -- CFD Online Discussion Forums

Hello. A collegue asked me to help him build a new home PC next month. Its sole purpose is to run flow simulations with large meshes using mainly

www.cfd-online.com

Which is important in the computational modeling and analyzing...

I mean in the general CAE programs such as ANSYS, ABACUS, Matlab and etc the more cores, the faster computing or the more maximum frequency of each core, the faster computing.

www.researchgate.net

https://pdfs.semanticscholar.org/ac02/b7a0443a9db95b0fe013976754d472a8e083.pdf

https://books.google.com/books?id=u...id Dynamics code using computer cache&f=false (goto Chapter 5)

 

[rylim]

I'd definitely go for 6 x 64 GiB. The extra cost now will mean that you won't have to discard the 32 GiB DIMMs down the road.

It will be almost impossible to measure a speed difference between six and twelve DIMMs on actual application.

That's my thought as well, and it costs more as the resale value for 12 x32 GB won't be the same as the present

At this time 4x32GB NEMIX (R-DIMMs)will cost you $585. So 12x32GB will cost you $1755.

Amazon.com: Buying Choices: NEMIX RAM 128GB 4x32GB DDR4-2933 PC4-23400 2Rx4 ECC Registered Memory

At this time 4x 64GB NEMIX (LD-DIMMs) will cost you $1,357. So 6x 64GB will cost you $2,036.

256GB (4x64GB) DDR4 2933 (PC4 23400) 4Rx4 288-Pin 1.2V ECC Load Reduced Memory by NEMIX RAM

256GB (4x64GB) DDR4 2933 (PC4 23400) 4Rx4 288-Pin 1.2V ECC Load Reduced Memory by NEMIX RAM

www.amazon.com

The 12x will provide 100% memory bandwidth, and the 6x will provide you with 97% of max bandwidth.

For the extra $281 it makes sense to go for the 6x 64GB LD-DIMMs that allows you to grow the RAM later on.

Yeah, that makes sense. Is R-DIMM and LD-DIMM something that I need to worry about?

You will never be able to measure the 3% loss of bandwidth in almost all applications - these CPUs have very large caches, so you'll only see that 3% on the small percentage of memory reads that are cache misses. A small percent on a small percent is unimportant. (The "memory bandwidth" tests in benchmarks are written to effectively disable caches - whereas applications are typically written to optimize cache use.)

In practical terms, however, filling six DIMM slots and leaving six empty means that adding memory in the future is super easy. Using twelve smaller DIMMs for the same capacity, means that you'll have to discard the small DIMMs to add memory. Penny-wise, pound-foolish.

Same thought here

Fluid Dynamic code knows all about L3 cache and pull in just enough to fit in the cache as it proceeds to solve a problem, and yes, so long as the needed data is in the cache the main memory performance is not an issue. However, once the data in the cache is no longer needed (has been used) and needs to be replaced with the next chunk of RAM data the memory performance becomes important. If this cache 'refresh' is done 1000s of times over several days (can be over 2 weeks) then memory speed/access times become very important. So I agree with you, except that cache misses can be very significant if there are 1000s happening all the time over several days for a long running multi-threaded code.

Yes, having to go beyond 12x 32GB will more expensive than going from 6x 64GB, but that may or may not be necessary for several years. The savings with having the former can be put to good use.

Thanks for the info!

 

[majus]

Yeah, that makes sense. Is R-DIMM and LD-DIMM something that I need to worry about?

Don't mix them.

 

[AidenShaw]

The careful coding for Fluid Dynamic (FD) problems and/or Computational Fluid Dynamics (CFD) problems as an example, is needed to make full use of a processor cache to minimize memory fetches for very long running executions.

You've proven that CFD is sensitive to memory bandwidth, and CFD is important *for you*.

Most applications, however, are not that sensitive. People could spend money for the "perfect" memory layout - when their workflows won't see any difference.

BTW, I always use balanced sets on all channels.