As mentioned elsewhere, my colleagues and I are gearing up to receive and work with a large area (~24,000 km2) of LiDAR data (las file point clouds, DEM, DSM, intensity grids and contours).

I will be working with this as far as is possible in Manifold 9 and a portion of the overall budget has been put aside for purchase of suitable hardware to work effectively with such a dataset now and into the future.

I am still waiting for feedback from our IT people but in terms of a high level spec, we have been looking at the following:

• 64GB of fast RAM.
• Dedicated PCIe SSD to host Windows, other software and cache (1 or 2TB).
• Dedicated PCIe SSD to host working area (4TB or greater and additional slots for future expansion).
• Multi-core processor e.g. AMD Ryzen Threadripper.
• Nvidia CUDA graphics card. Ideally more than 1 graphics card slot.
• Motherboard and PSU to support all this.
• Fast local or network store for storing the main dataset and derivatives.

Does this make sense in the context of Manifold 9 current or future additions (and in the context of multi-terabyte LiDAR datasets)?

Is there anything that is redundant or missing in the list above?

Does anyone have any thoughts as to particular components which offer a price/performance sweet spot?

Any thoughts are welcome as I am a bit out of my depth which computer hardware. Manifold 9’s infrastructure appears to be fast enough to feed the cores, so I am guessing that the ability to get the data on and off storage fast is key.

Thanks in advance

Landsystems Ltd ... Know your land | www.landsystems.co.nz

The potential bottleneck:

• Fast local or network store for storing the main dataset and derivatives.

Might be unavoidable.

Thanks Tim, we have been looking at Thunderbolt 3 enabled DAS storage which looks promising.

So, not shared, only accessible to the local user?

What advantage does that have over internal storage?

Shared between a couple of local users

Regarding sweet spots... AMD Ryzen processors that support PCIe 4.0 allow significantly faster M.2 ssd storage as compared to PCIe 3.0. Having at least two fast M.2 slots is wise.

By way of pure speculation, I'd also consider having a growth path for more RAM, like up to 128 GB. That's not going to be a big factor processing terabytes of data, which will exceed the size of RAM in an event, but it might be a help. Future editions of Manifold will probably provide options for "in memory" use, more aggressively grabbing large swaths of RAM to dedicate to doing all work 100% in RAM. For tasks that could fit entirely within 100 GB or so that could make a difference.

Many thanks for your informative and comprehensive responses. I will digest this today but this is certainly useful information to go into my list.

Landsystems Ltd ... Know your land | www.landsystems.co.nz

1) chia money that is like bitcoin money ( use farm of server ) need for "plot" M.2 NVMe SSD ( many SSD can have failure with time because intense R/W process ) .

2) some info graphic about PCI version history and SSD card to choose

3) only some chipset on motherboard ( for AMD ryzen) can support PCI 4.0 specification ( first offcial X470 then "upgrade" version of B550 ) .

in the feauture Alls connectors ( if number of wire is sufficient wit Ampere watt Voltage) willl support many prootocol. It is the case for USB -c . USB-C Alt Modes support DisplayPort, MHL, HDMI 1,4, Thunderbolt. USB-C.

If the electric current is greater than 3A, it must use a 5A wire with an E-Mark chip.

The wire’s capability level of the electric current depends on the thickness of the wire.

L'USB Type-C USB 3.1 max 20 volts et 5 ampères = 100 WATT (ancienne norme 5 V et 1,8A )

Boyle surface fr ,en

here some informations ( not up to date but try to update it with time )

so it is not because you have PCI 4.0 that you have the best spec you need also hardware connector length that support 16x !! so take care on number of lane and key ( slot / card ) !

Attachments:
PCIE_transfert_rate.jpg
PCI_card_speed.png
SSD_bus_size_connector.png
USB_protocol.png

Boyle surface fr ,en

in my doc i have this sentence retlative to CPU : "" Because Ryzen 3000 “only” can support a single-slot x16 PCIe 4.0, you have to choose whether to put your x16 PCIe 4.0 graphics card or your x16 PCIe 4.0 SSD in that slot""" . The ryzen 3000 seem to be an old generation AMD CPU !!! i hope (to check) that new ryzen CPU support many PCI 4.0 x16 lane with graphic card and memory ( RAM ) . The chipset should also manage many PCI 4.0 x16 protocol ( max lanes)

like Microsoft OS officialversion/ dev ...i have some difficulties to understand the term use by AMD for name their multicore CPU .

Zen4 in design

Zen 3 Design complete

Zen2 : 7nm

Zenand Zen+: 14/12 nm

AMD_CPUchoice_noTR.pn : price in euros ( more higher today ) for Zen 2 Z2 or Zen 3 Z3 generation of CPU . Green is the best choice to buy , Yellow for video editing encoding.

Depend of task that use CPU some old CPU architecture can perform better that new CPU because task use sub group of CPU !!

Hope all i write make sense

NB TR =ThreadRipper

Attachments:
AMD_CPUchoice_noTR.png
AMD_generation.png

Boyle surface fr ,en

On the off-chance this is useful to anyone else, I sent the high level spec I posted at the start of this thread to a local shop that builds custom PC's. The attached is what they came up with. Its not cheap but put in the context of a single use government license of Lastools setting you back 4000 euros (M9 LiDAR wish list (so far) (georeference.org)) it doesn't seem so bad.

Much of it doesn't mean a huge amount to me, but the case is cool

Attachments:
Clipboard-1.png

Landsystems Ltd ... Know your land | www.landsystems.co.nz

The PSU is too weak, especially when you consider durability. Try substituting a Corsair 1000-1200W.

16GB RAM is much too small. (After all, you are thinking of 32GB just on your GPU.) Get at least 64GB system RAM, and make sure there is room for twice that later. Get 128GB up front if you can. [Sorry, I misread. You are thinking of 4 x 16GB for 64GB total.]

Instead of the Threadripper 3975WX CPU, consider the Ryzen 9 5950X. Half as many cores (16/32 vs 32/64) but 75% the multi-core performance and 130% the single-core performance [not a typo] at 25% the price [again not a typo], less than half the power consumption, less noise... and choosing the 5950X would save similar amounts on the case and motherboard, and also case size. Upgrade in two or three years if you begin to find tasks CPU bound (very unlikely).

(Not quite as good for the salesperson, but better for you.)

Agree with those comments. Another thing to consider is that for CPU parallelism Manifold is currently optimized for the low tens of threads, not the mid or higher tens of threads. Running 64 threads won't slow you down, but it won't be as big a proportional gain as going from 16 to 32 threads and might not be significantly faster than 32 threads, depending on what's being done and everything else that is going on in the system.

Over time that is likely to change, with algorithms being adjusted to make better use of 64 and 128 threads, but there will always be optimums driven by the nature of the jobs being done. LiDAR processing certainly benefits from parallelism, but given the rest of the hardware (data access in particular), very many parallel threads could end up waiting on data.

That too, will improve, as Manifold has plans for newer generation architecture that would allow a more memory-resident approach, forcing everything to happen in RAM if you want and you have enough RAM. So notching up DRAM to 128 GB is a good idea for the future.

Also, I think the RTX 3090 GPU is way overkill. At over 10,000 very fast CUDA cores most of it will never be used. There's no heavy math involved in LiDAR processing, just very simple math repeated for a lot of data points, so much of it isn't CPU or GPU bound, it's limited by disk access. Given 32 CPU threads and very simple math, the optimizer might even choose CPU parallelism over GPU parallelism because that will be faster. It pains me to write that, because the 3090 is such a wonderful beast. :-) But latest gen Ryzen and Threadripper cores are also phenomenal compute engines, so running many of them in parallel can get a job done faster than it is worth dispatching to a GPU.

I'd therefore take the extreme price of the 3090 and spend part of the money saved on a less expensive GPU with around 1500 cores, with part of the money saved increasing the size of the M.2 SSD from 1 TB to 2 TB, and contributing the rest to dialing up DRAM to 128 GB.

[Overlapped with Dimitri's better post]

Just to expand on my post above, a Ryzen 9 5950X will be about 25% slower than a Threadripper 3975WX on a task that can fully saturate 32 cores, but about 30% faster on a task that can only utilize one core (for example: updating pyramids in Manifold 9, most UPDATE and INSERT operations [unless they are storage-bound... it gets complicated], many others).

Much of anyone's day is taken up with tasks in the latter category. Once tasks have been fully optimized, they can migrate to the former category--but only their best parts. Most tasks fall squarely in the middle.

So choosing the Threadripper over the Ryzen 9 would amount (in NZ) to paying more than four times the price, for, roughly and overall, almost a 20% performance penalty (but with more heat).

Threadripper CPUs are a bit of vanity exercise in my opinion, not a net performance gain.

Those comparisons are very very rough, based on public Passmark scores.

Which CPU would really be faster on a task that can fully saturate "only" 16 cores...? You would have to measure that with real data. It could be either.

It certainly would not be in the order of 4x either way.

(I would reiterate Dimitri's suggestions, no need to repeat them except to say, spend as much as you can get away with on super fast storage. Rotow's law.)

Wonderful advise. This is such useful (and insightful) information and very much appreciated.

I will definitely take onboard the comments about CPU, PSU, GPU, motherboard and RAM/SSD and get a revised quote. Any particular thoughts around a suitable cut down GPU that represents a sweet spot?

This. Not cut-down!

Quadro RTX4000

Quadro quality in hardware and driver. "Only" 8GB GRAM, but even Manifold 9 will leave most of that unused. Beautiful for graphics including the very best CAD.

And less than half the price of the RTX 3090 you were quoted.

If you do want cut-down, then this is what I would buy for myself, if one of my current cards broke.

GTX1660

Less than one-third the price of the above.

Perfect for Manifold. A Quadro card can be better for CAD and possibly Photoshop, but not much difference otherwise.

Good advice. The Threadripper machine used in the videos has a GTX 1660 GPU card in it.

Great thanks Tim. I think I have my system! Many thanks for yours and others invaluable assistance

Landsystems Ltd ... Know your land | www.landsystems.co.nz

It would be lovely to see the final spec!

Will do. It will essentially be the original swapped out with the recommendations you and Dimitri have made. I will post it once I get the quote revised. Thanks again.

Landsystems Ltd ... Know your land | www.landsystems.co.nz

It took a while to work its way through the machinery but in case it is of use to others, this is where we landed:

Case: Phanteks Enthoo Pro Black Full Tower Gaming Case

PSU: EVGA SuperNOVA 1000W G5 80+

CPU: AMD Ryzen 9 5950X

MOBO: ASUS TUF Gaming X570-PLUS (WI-FI) ATX

Cooler: NZXT Kraken X73 All in one Liquid Cooler

RAM: G.SKILL Ripjaws V Series Black 128GB DDR4 Desktop Memory 3600Mhz (4 x 32GB)

HDD1: Samsung 980 Pro 2TB NVMe PCIe 4.0 M.2 SSD

HDD2: Samsung 870 EVO 4TB SSD

GPU: Leadtek Quadro RTX4000

We ended up ordering 2 machines of this spec thanks to the excellent advise provided by Tim and Dimitri. This was incredibly useful and meant that to purchase 2 machines was only a little over NZ$1000 more than a single machine of the original spec. Thanks again for all your help. A great result I think

This is very timely. I think I've said that before in this forum, and I still reserve the right to not quite pull the trigger. But we're moving to a new house, and I want to do it this time...after we sell the current house.

There are some 'power users' on another forum I'm on. One guy has a raised floor and stuff I never heard of, so it must be good, right? He mentioned he was getting a 43" monitor. I asked him about it and it's a TCL 43S435, 4K, Roku television set. Apparently it has 4:4:4 chroma(insert appropriate tech jargon) which allows it work as a monitor, because the fonts are legible in all colors and backgrounds. The 43-inch model is the small one in their series 4 lineup, so if 43 inches isn't enough for you, they get bigger. He's very happy with it and got another for his son. He was buying last winter when the holiday pricing was in place, and got them for $225 USD. It's still under $300 at Best Buy and even Walmart. I know Dimitri talks about using 3 monitors. What would you change if one monitor was 43 inches?

What would you change if one monitor was 43 inches?

I'd buy two more. :-)

(Just kidding...) The key is getting more screen real estate so you have plenty of room to spread out all the different windows that are used by the applications you have going. If you have several different Manifold instances running it is very convenient to have room to keep them all in view.

Within a Manifold instance, it is very productive to have plenty of room to undock windows and panes, and to be able to resize them as you like. Sometimes, for example, I'll have a really wide table window going so I can see many columns at once, and sometimes I like having a really big map window open, so I can see a very large region at once.

How you get to "more screen real estate" depends on the details of monitors available, their cost, your budget, and so on. There are lots of different ways to get to what works for you.

You have to be careful matching screen resolution to monitor size so your text at 100% sizing doesn't end up being too big or too small. You also have to be careful with viewing angles, as you sit closer to a monitor than to a TV, so if it is a large, flat monitor the pixels away from you on the edge may require different focus for your eyes and they could be different brightness, as well as distorted by perspective. Gamers use curved monitors to get around those issues.

A high end solution for many is to use a 49" super ultra wide, curved monitor with a 32:9 aspect ratio, at a cost of from around $900 to around $2500. Curved monitors make for better viewing and similar pixel brightness on the edges of the screen (since the viewing angle for edge pixels is more perpendicular) compared to flat monitors.

But spending on a super ultra wide doesn't necessarily get you as much screen real estate as you want. A 49" super ultra wide at 32:9 aspect ratio only has twice the screen real estate of two, comparable pixels per inch, 27," 16:9 monitors. So you could end up with both a cheaper solution and more screen real estate using multiple monitors.

When using multiple monitors I like three monitors because even with ultrathin bezels there is still a slight gap between multiple monitors. If you use two monitors that seam is right in the middle of your view. If you use three monitors, the central view is free of a seam with accessory windows appearing in the left and right "wing" monitors.

The wing monitors can also be arranged in an arc with the middle monitor so they are perpendicular to your view, like the edge pixels in a curved monitor. I arrange my monitors so there is a slight overlap where the center monitor is placed ahead of the left and right monitors, so that the seams are only the thickness of the center monitor's bezel and dead screen space next to the bezel.

Or, for about $350 on Amazon, you could get a 34" curved monitor in 21:9 resolution that is "frameless". Use three of those for lots of screen real estate, while having a wide, seamless, field of view in the center monitor.

By the way, that rig LandSystems cobbled up sounds really super. 128 GB of RAM and 32 threads from an AMD Ryzen 9 5950X is just awesome!

Good discussion. Thanks.

Perhaps consider adding a small touchscreen just to vary things up for your health.