Faster drive for boot or storage - directstorage

[Shott92]

So speccing up a build to help me do some AI driving simulation for my uni course.
Wanting to save a couple of quid while getting best performance, also it appears MoBos may be limited in number of pcie5 slots anyway.
These simulations are super compute intensive and can an hour to run just a 1 minute, simulation. So every little helps, but I'm still on a budget.
So would I use a pcie 5.0 SSD as my boot and program drive (which will have Clara simulator and airsim installed) and then have my footage, maps, programs installed on the PCIE 4.0 drive, or the other way round.
Or should I have it all on the same drive so that there isn't any shared I/O bandwidth?
I'll be looking at a 12700/13600 or 6600X/6700X (if available at launch), with a RTX4070/RX7800.
Two SSDs probably one 2TB and one 500 GB
And then as much DDR5 ram as I can afford.
Any help or guidance is greatly appreciated.
Thanks

 

[EsaT]

All Alder Lake's 16 PCIe v5 lanes go to graphics card slot.
(or are shared between it and another "x16" slot)
Lanes connected to M.2 slot are PCIe v4.

And anyway if you're on more limited budget, PCIe v5 drives are going to cost arm and both legs.
(and run hot as hell under sustained load making whole sustained performance questionable)

 

[Tetras]

You sure they actually care about the storage? From what you describe it sounds like it'll be more CPU and memory intense. My guess would be that if it is I/O intensive then you'd want the footage, maps and programs that are involved in the compute task on the fastest drive, not Windows.

 

[alec]

directstorage is not a thing for a while
pcie 5.0 storage is not released yet, so not tested. Can't give any advice, except that for the first while only AMD AM5 socket motherboards will have a Pcie 5 M.2 slot.

But like others pointed out, your workload sounds to be cpu/memory intensive. So any nvme drive, even pcie 3.0, should do fine for loading resources quickly. It is rare to be limited by max storage bandwidth in compute tasks.

Minimise CPU/memory latency.
Pay attention to single/low thread performance of CPU you are picking. Both Zen 4 and Raptor Lake should be fine on that front, but there may be advantages for higher tier CPU that come with higher boost clocks.
DDR5 makes it easy to go 64GB, but try to get it in 2 sticks not 4. Current memory controllers struggle with 4 sticks in general, will not clock as high. This makes latency go up significantly, which is usually bad for any simulation workloads

TL;DR
don't chase storage performance, do chase latency

 

[Shott92]

Hi guys thanks very much for the feedback, so my thoughts was that the way the simulation will run, is it will build a map from the SSD, I assume mostly held in VRAM, but a full map will be too big, especially if all assets are turned on.
It will then try to run my code, which will be a mix of standard image manipulation and performing mathematics and matrix functions on the image, which will be converted into essentially a data array/ table, with raw data, and some AI workloads on top, these will both mainly be done with GPU enhancement.
Again due to size it won't fit maps, simulation assests, data outputs and all these data frames into VRAM or RAM so I suspect it'll have to dump to the SSD at pretty regular intervals.
But I'm not certain, i think what I'm going to do is build the PC with a single Samsung 980 pro.
And see what the resource utilisation is and if the I/O is high and I'll try to remember to update this post and let you guys know the outcome, but the build will be a couple months away yet as the rumours suggest the rtx 4000 and rx 7000 will offer significantly better value than current cards.
Thanks again everyone

 

[alec]

Oh, if it is a case of regular overflow of VRAM and RAM onto storage, then forget everything I said about latency.
Even the fastest storage is order(s) of magnitude slower than RAM. So definitely try and got enough memory or reduce the workload size.
Else all your fancy new CPU and GPU will be doing is wait for storage.

If there is no other way and you're dealing with TB size arrays regularly, good plan with Samsung 980 pro. Something that could be hammered for a logn time without a drop in performance. Also in same tier are WD black SN850 and SK Hynix Platinum P41.