Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
MP 7,1 12 x 32GB for 384GB or 6 x 64GB for 384GB on 2019 Mac Pro?
- Thread starter paulchiu
- Start date
- Sort by reaction score
I'd say 6*64 GiB is far better - because when your projects grow and you need more RAM, you won't have to throw DIMMs into the eWaste bin.
In the long term, the most valuable things in a workstation are DIMM slots and storage bays. Don't fill them with small items - half-fill them with larger items.
And don't listen to people who say that six DIMMs give you 97% of bandwidth, and twelve DIMMs gives you 100%. You'll never see that in applications, because the big caches are far more important than tiny differences in performance on synthetic memory benchmarks.
Thanks AidenShaw.
I did not know what that 97% means when I read it. How did they even compute that?
So far since the Mac Pro came out, prices on the 64GB sticks have increased more than the 32GB ones.
I just want fast!
Thanks again.
Memory benchmarks are "bandwidth viruses" that are deliberately written to make processor caches irrelevant. This is usually done by accessing memory regions that are much larger than the caches, so that the caches don't help - if you don't have "cache hits" the caches don't help.
Whether the "97%" is tested via "bandwidth virus" benchmarks, or mathematically calculated, is not the point.
Applications (and compilers) are optimized to attempt to get the most efficient use of the caches. This reduces the number of times the application has to wait for the memory system, and makes the difference between 97% and 100% disappear for the majority of applications.
For a longer discussion of testing done with a 6,1 with one, two, three, and four DIMMs - see https://forums.macrumors.com/thread...bservations-with-various-mem-configs.1704700/ . The takeaway is that most components of a benchmark suite were basically unaffected by the number of memory channels - only one test was significantly slower with fewer channels.
It is unlikely that you would be able to measure the difference between 97% and 100%.
The speed issue aside, when I bought the memory I paid about $1200 for 12x32GB. It would have cost $1800+ for 6x64GB. Both of those prices have gone up. I voted with my wallet.
Prices are very much in flux at this time. The vendors are obviously aware of the RAM (DIMMs) demands arising from the MP7.1 owners wanting to help with the MP7,1 cost.
I was lucky, as I got out early and bought 8x 32GB R-DIMMs for at around $900. The same source is now charging some $1200 for the same kit.
I also vote with my wallet, which is why I pick the option that doesn't involve eWasting $1200 of DIMMs in order to expand over 384 GiB.
Short term economies often cost more in the long term.
I am not sure what waste you're talking about. If and when you need more, why wouldn't you just sell the DIMMs to someone else? You make it sound like the only option is throwing them in the trash.
according to these benchmarks and depending on various configurations, differences in memory bandwith on the Mac Pro 2019 can vary significantly ->
2019 Mac Pro: Memory Bandwidth
macperformanceguide.com
Hey ! I'm one of those guys ! 🤨 Someone did a lot of research with Cascade Lake Xeon memory bandwidth on the PC side . And now we're trying to verify it with the MP 7, 1 . There should be no difference since the memory controller is in silicon . And yes , there is a various amount of bandwidth available depending on installed memory configuration , ranging from 35 to 100 % . It should not be dependent on caches . Aiden , you should know how easy it is to overwhelm caches in a demanding workflow situations - the whole reason someone would plunk down some serious coin on a MP7,1 . This is not a situation that is just measured with a quick one or two minute long synthetic benchmark .
I was hoping someone by now would run some sustained bandwidth tests , because I'm not fully prepared to do it at the moment . And given the paucity of test software on the macOS side , it'll have to be found on the Windows side . So , I'll need to install Win 10 Pro Workstation on my MP7,1 .
Have you ever run SiSoftware Sandra 20/20 ? The publishers claim it measures sustained memory bandwidth , is able to be run with Cascade Lake Xeons and can be used in a configuration up to a total of 1TB . When I can schedule it in , I'll do this and run some tests on the few modules I have available at this time . I currently have a matching set of four 8GB modules and another matching set of six 16GB modules . I was hoping to wait to run these tests until I had a complete set of twelve matching modules of any size .
On paper, the 6 banks should perform better than the 12. On paper, the 6 banks might also run hotter than the 12 bank solution. The question is what really is the delta of the two latter when it comes to performance. I would argue they are similar enough that it becomes a matter of financial concern. One gives you empty banks for later and the other is cheaper now and you would have to give up some banks that are occupied in order to increase RAM...possibly sell the 32 gig RAM at a lower price.
I admit that if it were me, I would go with the 6x 64 with the notion that I may bet more 64 memory pairs.
I admit that if it were me, I would go with the 6x 64 with the notion that I may bet more 64 memory pairs.
eWasting them is a lot simpler than listing them on an auction site - and hoping that you'll get a reasonable bid, when you've already bought the higher density DIMMs - because you can't sell them before you bought the bigger DIMMs.
You actually have it reversed. The technical term is rank, and there's at least 1 rank per DIMM, possibly 2 or 4. The memory controller can precharge a certain amount of data per rank (prefetch) and interleave them. So when one rank is transferring data, another rank can be preparing different data for access.
DDR4 has multiple layers of this (channel, rank, bank group, bank).
When you have more ranks per channel, the memory controller can do more "prefetching" and reduce the time the channel spends idle waiting for the RAM to retrieve data.
Thus more DIMMs means slightly better performance.
I am thinking more along the lines R-DIMMS vs LR-DIMMS which in this care are represented by 32 gig offering and 64 gig offering by Apple. This is where the cut exists. As well, there is the matter of the memory controller. Given that the total here is below 512 gigs, the only difference should be found in power requirement and the latency associated within LR really isn't going to be apparent due to the trade off in number of banks and the memory controller as mentioned. Again this is on paper and in hands on use, no one is going to know which combination is in the box. There is no diminished return here as related to ranks.