Is it true if you have 2x 16GB of Ram its half as fast as 4x 8GB on a Xeon based 2011 V3 socket?

Just saw a post on Toms Hardware, where they claim if you have 4 Channel memory then having only 2 sticks of RAM means it only performs at 50% of the speed of the same amount of RAM if there are 4 sticks.

I went for 2x16GB to provide for an easier upgrade. I.e. If needed add 2 more to deliver 64GB.

I would say its extremely quick, so this is definitely not one of those "my workstation is not performing properly" threads. It actually outperforms my expectations.

I am just very interested.

Any views on this. The MB is a Supermicro X10SRA-O with the Intel Xeon E5 1650 V3 cpu. Running Windows 10 Home

The manual doesnt specifically warn or comment on this and shows both 2 and 4 sticks. Converseley the HP Z640 manual rates them as Good and Excellent respectively.

As I doubt I will need 32GB atm thats why I went this 2 stick route.

On a synthetic benchmark, you probably would see a 50% hit on memory bandwidth for Quad-channel DDR4 that only have two slots populated.

Normally, if I spend $900 on a server-grade motherboard and CPU, I wouldn't want gimp the performance of the ECC DDR4 memory by installing only 2 sticks. However, unless you do run multiple hardcore x265 encoding projects or have a ton of VMs, I don't think you would notice any performance hit in real-world usage since memory bandwidth isn't going to be the bottleneck for most applications.

So why would you not post this question on Tom's Hardware guide forum? You will believe the answers posted here but not there? Here we go again.

cuongduong wrote in post #17985727
On a synthetic benchmark, you probably would see a 50% hit on memory bandwidth for Quad-channel DDR4 that only have two slots populated.

Normally, if I spend $900 on a server-grade motherboard and CPU, I wouldn't want gimp the performance of the ECC DDR4 memory by installing only 2 sticks. However, unless you do run multiple hardcore x265 encoding projects or have a ton of VMs, I don't think you would notice any performance hit in real-world usage since memory bandwidth isn't going to be the bottleneck for most applications.

Thanks

Shaymus wrote in post #17985960
So why would you not post this question on Tom's Hardware guide forum? You will believe the answers posted here but not there? Here we go again.

Sadly I was born with an enquiring mind, so am interested in different opinions.
You on the other hand must be of the opinion if you read it on the Net it must be true
After all every person who posts is a genius! And must be believed. no ?

As they say if you are not interested why post?

Just trying to learn everything I can in 2 months!
Fortunately some kind people are willing to share their knowledge.

Submariner wrote in post #17986262
Sadly I was born with an enquiring mind, so am interested in different opinions.
You on the other hand must be of the opinion if you read it on the Net it must be true
After all every person who posts is a genius! And must be believed. no ?

As they say if you are not interested why post?

Just trying to learn everything I can in 2 months!
Fortunately some kind people are willing to share their knowledge.

The issue is that this isn't just a post by one random guy on the internet... this is a well documented and talked about fact (which the manufacturer will usually disclose on the product information page).

EverydayGetaway wrote in post #17986534
The issue is that this isn't just a post by one random guy on the internet... this is a well documented and talked about fact (which the manufacturer will usually disclose on the product information page).

Thanks Tim
OK Supermicro manuals are really sparse, not like MSI MB manuals (really more geared to knowledgeable corporate IT Pros I guess ) but there is no mention in this manual or on their support pages.
Other than showing you can have 2 sticks or 4 sticks, or 6 or 8 ( up to 512GB RAM ) and which order to populate the memory slots first and the order, and of course to say they must be the same Memory.
The only slight warning/advice given is if you have only 4GB that system devices are allocated some memory so it only leaves 2.8GB to run the OS and all applications.

Conversely
HP does suggest for maximum performance on the Z series workstations that 1 stick of 16GB = good, and 2x 16GB sticks = Better, and 4x 16GB sticks = excellent. They also show that 4x8GB perform better than 2x16GB but say its only slightly better.

No mention of 50% performance
I wonder if this is somewhat motherboard dependant?

Maybe my good performance is because I am actually not really needing more than 16GB atm?

Oh well if its true, I guess if I ever see it slow down, then get 2 more sticks. ( sure I wont need it ).

Benefits are theoretical, only really help if you're using massive amounts of memory bandwidth, such as perhaps video encoding or benchmarks. In practice benefits will be very limited and practical gains will be close to zero for most common interactive scenarios. Cache in the CPU (L1, L2, L3, sometimes L4) mitigates memory latency, and in-cpu cache is (approx, guessing) 100x faster than RAM.

Dual, triple, quad-channel memory is only a percent or two faster than the previous. It's hardly noticeable, except for FLOPS-peepers (the computer equivalent of measurebating pixel-peeping for noise).

More importantly is the bus-type, and optimization of the software. What really kills performance in multi-channel boxes is poorly written software that causes cache-thrash and bus overutilization.

Cache-thrash is when two different cores are processing data simultaneously on shared-cache architectures. As one core pulls data and fills the cache, the other core pulls data, realizes that the cache is full, and dumps it to cache it's own data.

Having 2x 8GB in a quad-channel box vs 4x 4GB for 16GB total is hardly noticeable except to those who are eeking a few FPS of video or for purely sequential data sequences.

tim wrote in post #17987172
Benefits are theoretical, only really help if you're using massive amounts of memory bandwidth, such as perhaps video encoding or benchmarks. In practice benefits will be very limited and practical gains will be close to zero for most common interactive scenarios. Cache in the CPU (L1, L2, L3, sometimes L4) mitigates memory latency, and in-cpu cache is (approx, guessing) 100x faster than RAM.

Thanks Tim
My CPU has 15MB of cache - no idea if that is a lot or not

docholliday_sc001 wrote in post #17987195
Dual, triple, quad-channel memory is only a percent or two faster than the previous. It's hardly noticeable, except for FLOPS-peepers (the computer equivalent of measurebating pixel-peeping for noise).

More importantly is the bus-type, and optimization of the software. What really kills performance in multi-channel boxes is poorly written software that causes cache-thrash and bus overutilization.

Cache-thrash is when two different cores are processing data simultaneously on shared-cache architectures. As one core pulls data and fills the cache, the other core pulls data, realizes that the cache is full, and dumps it to cache it's own data.

Having 2x 8GB in a quad-channel box vs 4x 4GB for 16GB total is hardly noticeable except to those who are eeking a few FPS of video or for purely sequential data sequences.

Thanks for the detailed info, if its only a few % less rather than 50% I would still repeat buying the 2x 16GB for future proofing reasons.

Submariner wrote in post #17987839
Thanks for the detailed info, if its only a few % less rather than 50% I would still repeat buying the 2x 16GB for future proofing reasons.

That is the best way to go. That way, when an Amazon sale appears on RAM, you can throw another 2x16 in!

In the early days of dual and quad channel memory, a few colleagues and I tested the difference. In compiling an MPEG video with purely CPU (no GPU assist in those days), the difference amounted to 30s-1m less compile time in a 1 hour video. With the new design multi-core processors (and lack of FSB design in the new chipsets), that is going to be a few seconds per hour of footage difference at most.

It's not enough to make a difference for general apps. The 15MB of L3 cache in your process is good, fairly common for most Xeon's nowadays. My personal box currently runs dual-physical procs with 16-cores and 40MB cache per proc. I'm running 256GB of RAM - 16x 16GB ECC. That makes compiling code happen in seconds instead of minutes and videos at a rate of minutes per hour of footage. And, since Lightroom is so poorly written for high-end machines, it still chunks a bit. The codebase optimization sucks balls.

Spend your money on a good, high end SSD instead. Either an Intel or Samsung 850 PRO...

docholliday_sc001 wrote in post #17987916
That is the best way to go. That way, when an Amazon sale appears on RAM, you can throw another 2x16 in!

In the early days of dual and quad channel memory, a few colleagues and I tested the difference. In compiling an MPEG video with purely CPU (no GPU assist in those days), the difference amounted to 30s-1m less compile time in a 1 hour video. With the new design multi-core processors (and lack of FSB design in the new chipsets), that is going to be a few seconds per hour of footage difference at most.

It's not enough to make a difference for general apps. The 15MB of L3 cache in your process is good, fairly common for most Xeon's nowadays. My personal box currently runs dual-physical procs with 16-cores and 40MB cache per proc. I'm running 256GB of RAM - 16x 16GB ECC. That makes compiling code happen in seconds instead of minutes and videos at a rate of minutes per hour of footage. And, since Lightroom is so poorly written for high-end machines, it still chunks a bit. The codebase optimization sucks balls.

Spend your money on a good, high end SSD instead. Either an Intel or Samsung 850 PRO...

Hi
Thanks Doc, thats most encouraging news. And all the details so much appreciated
Obviuosly thats why they didnt highlight it too much in the manual.
My plan was to stick to 32GB till I start rendering 4K video say in a year ( my 5DS R only does 1080p )
I have an Intel 750 PCIe for Win 10 and Programs and the current shoot. I have 2x samsungs 850 evos 500GB to copy the current shoot to when I complete editing.
I was going to buy the 950 Pro, bit seeing the high heat at the M2 connector think I prefer to buy another intel 750 when my supplier has another sale.

Lets hope this MB works :-/

Submariner wrote in post #17988499
Hi
Thanks Doc, thats most encouraging news. And all the details so much appreciated
Obviuosly thats why they didnt highlight it too much in the manual.
My plan was to stick to 32GB till I start rendering 4K video say in a year ( my 5DS R only does 1080p )
I have an Intel 750 PCIe for Win 10 and Programs and the current shoot. I have 2x samsungs 850 evos 500GB to copy the current shoot to when I complete editing.
I was going to buy the 950 Pro, bit seeing the high heat at the M2 connector think I prefer to buy another intel 750 when my supplier has another sale.

Lets hope this MB works :-/

Very good choices in hardware. I've been designing high performance/reliabilit​y boxes for over 20 years. I'm also a low-level coder, mostly asm and C++ work while doing some .NET stuff for databasing and parallel image processing code (medical imaging). The same box runs my Lightroom/Capture One/Photoshop/InDesign stuff. Stick to the good stuff and it'll last a long time...my boxes get pounded hard day after day and I've routinely had RAM loads > 24GB constantly.

I use the i750 drives, as well as the 850 Pros (the evos are ok for desktops, but do not like RAID - they use an SLC buffer to give the feeling of speed. In the long run, the Pros are "pure" speed. Since RAID isn't about direct file transfers, the evos go nuts after a while and will slow down). As far as other hardware, I stick to Supermicro boards, Intel Xeon procs, Micron or Samsung RAM, NVIDIA Quadros, and Seasonic 1.2kw+ power supplies. There's no advantage to liquid cooling either - a good set of Panaflo or Delta fans do the job perfectly; just remember that the CFM of the exhaust fans should be close to the amount of cold/room air that is entering the case, otherwise you'll have a vacuum and not cool anything (look at the gamer kiddies with their 10 fans sucking air out, but little air coming in!)

For archival data, I use physical SAS 15K drives in a fiber channel NAS RAID'd for storage. A bad drive can still be repaired, but most SSDs cannot be recovered as easily. It's still important to have a good backup plan to keep the data. My LR catalog is around 5GB and the images folder on this box is sitting at about 2.8TB currently, and it loads, runs smoothly

The trick is to find the bottlenecks in the system and address them. It used to be the main bottleneck was the disk path. With SSDs, that's gone away. RAM bandwidth is rarely an issue nowadays. The bus speed was a bottleneck for some time, but with QPI, that's going away. It's now down to the poorly written/optimized software that is moving data (very rapidly) into the processors and causing cache thrash.

With video, the graphics card can be a bottleneck, if the system doesn't have a hardware encoder for the output. Matrox makes quite a few hardware MPEG encoders, as well as other formats, that can help that offload. Adobe Premiere takes advantage of many of these. Also, with Sony Vegas, you can also split your rendering across multiple boxes over a network connection, so that "cluster render" can also help. Of course, with CUDA and OpenCL, GPU assisted rendering *can* be faster than pure CPU rendering, but not always - a small, slow card can also be slower. The hardware encoders are still the best choice.

Hope that gives you some insight!

docholliday_sc001 wrote in post #17988970
There's no advantage to liquid cooling either - a good set of Panaflo or Delta fans do the job perfectly; just remember that the CFM of the exhaust fans should be close to the amount of cold/room air that is entering the case, otherwise you'll have a vacuum and not cool anything (look at the gamer kiddies with their 10 fans sucking air out, but little air coming in!)

I have to respectfully disagree. There is definitely advantage to liquid cooling, especially when you utilize Supermicro Hyper-Speed motherboard. I would never be able to get those E5 processors to push more than 8% BCLK with normal cooling, and those Delta case fans are about as loud as it could get.

docholliday_sc001 wrote in post #17988970
For archival data, I use physical SAS 15K drives in a fiber channel NAS RAID'd for storage.

Hmm, a fiber channel NAS? Is your NAS off-site that you would need to utilize FCIP? Anytime I set up a NAS, iSCSI is the protocol I would use. Anytime I set up a SAN, it's Fiber Channel. The upfront cost of deploying a FC device is just too high for most of the SMBs.

Also just curious, but what RAID level are you using?