I am dying to see raid 0 samsung 960 pro benchmarks on TR. It should be 7GBps, less overheads for a pair.
raid also means sequential write speeds should be similar to read speeds, unlike single drives.
Its significant, imo, that write speeds improve disproportionately from raid0. Its more balanced. Fast reads are no good if they are waiting for a write to happen.
I think folks will regret not getting 3 (seems the maximum) or more native m.2 ports on their TR mobo – native will be faster than any pcie3 discrete controller card imo.
Yep,the 8 core 1900x was a surprise, & at that price, its a temptingly small premium (~100usd) to go from the top ryzen to the bottom of the industrial strength league.
As u may know, vega can use either system ram or even nvme raid storage as gpu cache extenders, utilising the HBCC chip.
Just bouncing ideas around, but there is, for example, a solution~ were it gpu datasets.
Vega has hbcc, which allows virtualised STORAGE to be used as a gpu cache extender.
On a lane rich TR or epyc, use a 16 lane, 4x m.2 port nvme highpoint hardware raid controller card ($400usd).
Then build a 4x striped nvme ssd array, with a focus on speed, not capacity – even 4x 128GB el cheapo (samsung oem new ~$75usd aliexpress/ebay ?, google pm961) or same in 256GB for ~1.6x more.
or, get the best, which is 500GB samsung 960 pro (~$350usd ea.) – even 2x striped may be faster than 4x as above.
The sadly dear highpoint controller, is the only option atm afaik. Dunno how it compares with ram prices, but the cheapest option as above would be ~$700usd & would provide 500GB of pretty amazing bandwidth storage resource. An array of 500GB 960 proS as above, would very nearly saturate the cards 16x pcie3 lanes in sequential write (singly, they bench at 3.5GBps), and yield ~12GBps seq. write allowing for overheads.
Vitally imo, the weakest link is strengthened disproportionately by raid 0. i.e., the gap between read and write speeds is narrowed considerably.
So if as u say, u r a ~ dilettante for now, u could test the water with slower but vast gpu workspace via vega and hbcc, and have a valuable temporary very fast workspace for apps that need & profit from it also.
Note that a decent TR mobo will have 3x native nvme ports onboard as well (per the TR spec), but no native raid.
You could then reason that it makes sense to be very niggardly about your ram spend in partial recompense?
As I said, just posing a few ideas.
NVMe RAID support
AMD has also announced a free update to the X399 platform – which Threadripper CPUs run on – which brings NVMe RAID support, allowing for incredibly fast data transfer speeds. This update includes NVMe RAID drivers, and supports bootable RAID 0, 1 and 10 modes for up to 10 devices.
The free update will be made available on September 25.
Semantics help here. TR is covered by a "heat spreader".
It would seem a cooling block covering the whole rectangle would have best effect.
Doh? – but I have never seen it actually said, so maybe it isnt?
"I only buy the best" is a 2 edged sword.
With a CPU etc., the more u pay, the more reluctant u r to upgrade, so u may end up having a worse cpu than the guy who got the cheaper one and put the savings toward his upgrade fund for when it runs out of steam.
A few here have said they plan to double memory later.
What is the optimum plan for a future memory doubling?
The general rule seems to be more sticks, lower ram speeds, so how can we newbs get the best outcome from what we buy now?
If there i a simple answer it would be appreciated.
Ta. So if speed is vital (& it is~ w/ zen, as mem clock sets the fabric clock), u need to replace the whole kit with a larger one.
So if, oth, u do add extra sticks, the original memory sticks will have to be downclocked a bit, correct?
High standard of debate. Kudos guys.
Ta for helpful responses to my memory queries.
re the tipping point betw TR & ryzen – yep,its a bugger.
I feel a voice in the wilderness, but i think a raid0 nvme array will be an important part of a pcS arsenal in the near future, & those who dont at least make provision for that will regret it.
For the first time, its a realistic option which hasnt previosly needed consideration by pc buyers.
Even if you can only manage a 2x array of 256GB ~$150us (there are cheaper oem replicas (sm961?)) samsung evoS, you will have created
the evoS are rated at ~3200GB/s SR & 2250GB/s Seq Write, so say 6GB & 4GB/s for the array for simplicity.
So in fairly short order, we have gone from affordable mainstream system drives of 100MB/s sata HDDs, to ~300-500MB/s sata SSDs to 3500MB/s for top nvmeS, and i doubt it will be long before they saturate the 4GB/s of pcie bandwidth allocated them.
All good, but a plateau, since nobody had lanes for more than ~1 drive.
Ta Daah – OTH, we now have 64 lane TR, meaning we have the lanes to double, triple ... or more, those single drive speeds.
This new mainstream resource, is small by storage standards but huge by memory standards. It is slow by system memory standards (~40GB/s), but incredibly fast by storage standards.
It fills a huge gap/vacuum that used to exist between storage and memory speeds and latencies, and offers the tantalising prospect of ~unlimited memory, sort of, but probably much better than u think..
24 core 1P Epyc is tasty at ~1050US.
u may save a bit on ram using more smaller cheaper sticks.
even 64 lanes is tight for multi gpu and an nvme raid array.
imo, in deciding I would factor in a liklihood that if a vega zen rig could work, there would probably be synergies to the jointly developed processors sharing a common fabric bus and ecosystem, and much more.
I think 8 sticks are slower than 4 sticks. Just check the ram configuration specs for any TR mobo to confirm it.
8 sticks at 3200 may be ambitious.
I have a thought to share on this.
A good strategy if buying 2x pcS (not uncommon in buiness) & want to provide a pragmatic memory upgrade path?
Skip memory on one pc, and get a double size kit for the other, then split the kit between PCs after testing.
If u opt to double ram later, just re-unite the two previously tested halves, and buy another kit for the other pc.
As you can see from any TR mobo manual, if you fill all the memory sockets, it will only support lower speed ram.
Most likely, judging from kits available, 3200 if u use 4 slots/sticks & ~2900 if u use 8 sticks.
aka a single channel rig vs a dual channel ram setup.
If you start with 4 sticks at the desireable 3200, then add 4 for a total of 8 sticks, then u see the problem.
You could start w/ 4 sticks at 2900, then add another 4 at 2900, but they wont be a factory tested 8 stick kit, so there is a risk they may need to be run a bit slower again.
Ram speed is regarded as precious with zen, as the speed setting also dictates the fundamental Infinity Fabric bus speed.
Most would say beg borrow or steal the money to get the end target ram size in a faster 4 stick tested kit at the outset.
I have little experience. I hope the above is clear and correct.
the fastest is one stick per memory channel.
TR is 4 channel, so 4 sticks is fastest for TR
3200 seems a sweet spot to aim for.
(Ryzen BTW, is 2 channel, so 2 sticks is fastest)
A rhetorical point, fyi?, since TR nvme raid is a hot topic ATM, given the imminent drivers for bios raid due from amd.
OPTIMALLY TWEAKING A RAID0 ARRAY
I was reading a raid review of 4x 960 pro nvme drives, which as i recall were new, and in the setup part, it showed the individual speeds of the drives.
The read speeds differed radically – a gap of 400MB/s from best to worst as I recall.
Point being, if we end up using various nvme ports for our raid0 arrays on our TR, the slowest drive will set the pace for all drives in the raid0 array.
We should balance things by putting our best drive on our worst port in the array e.g.
We should rig our array to have minimal variance in the individual performances of each drive.
Yes, they sure give tangled explanations in manuals, & there are many manuals.
It may help to think more of lanes – each full strength nvme ssd needs 4x PCIE3 LANES, not 4x pcie2 lanes (think of them as half lanes – nvme can use these older standard lanes, but at ~half the speed).
SO
in effect,
TR starts with 64 lanes less 4 for the chipset
The norm/specification is for the TR mobo to use 12 more of those lanes for heavy duty ports like nvme or improved sata ports.
That normally leaves 48 lanes to be shared by the pci slots.
The complicated options become simpler/logical, if u just NB that u cant have it both ways. The lanes have to come from somewhere out of the allocated total of 64.
u cant have 3x 4 lane (the total 12 lanes allocated to mobo ports) nvme ssdS, AND the improved sata ports e.g.
More than 12 lanes can be used for onboard ports. I have seen a TR mobo that lets u pinch some lanes out of the other 48 by reducing the lanes on a pcie3 slot from 8x to 4x e.g.
As the above poster said, selecting the sata and NVME on his mobo, meant the nvme port had to use 4x lesser bandwidth pcie2 lanes – in effect, the extra and better sata ports get 2x pcie3 and the nvme gets a lesser 2x pcie3 lanes of bandwidth.
Some moboS are coming with 10Gb lan now, usb3 is quite high bandwidth etc. – & options may similarly need some balancing for your priorities, given available lanes.
imo, the most valuable use for the normal 12 lanes is for native nvme ports, & i prefer a mobo w/ 3x ports – like msi or asrock.
If u establish u cant pinch lanes from the pci bus as above, then u have 12 lanes total to play with.
If you tally up the lanes u need for the options u desire, u should know if its doable or not & the options will begin to make sense.
TR NATIVE RAID DRIVERS RELEASED
Download NVMe RAID Drivers For AMD Ryzen Threadripper X399
http://www.guru3d.com/files-
REVIEW
Re the deabate on allowing for enough ruinously expensive ram on our chosen rig.
Its most likely we will overspend for 90 % of our needs, and it may be we could live with lesser performance for the peripheral ram hungry jobs.
Remember the bad old days when swapping out to disk at times was unavoidable due to ram scarcity?
Horrible right?
Its still a solution though, and storage has rather changed since then but our perceptions on swapping have not. Since ssd became a mainstream device, we have gone from 100MB/s sata hdd to 3500MB/s nvme read speeds, & writes ~2/3 of that.
Amd have just given us the bios raid means to cost effectively ~triple single drive read speeds & get ~2.5x those nvme write speeds. That's 105x faster reads than our recall of swapping to storage on a 100MB/s hdd.
Maybe a bit of swapping occasionally isnt the end of the world? What is old is often new again.
Cost effectively using the stock 3x nvme ports on many mobos, 3x samsung sm961 250GB nvme @$150~ US ea. should yield ~9GB/s+ reads from a 750GB array.
BTW, FYI, Ryzen seems to crack 40GB/s, & TR 60GB/s in memory bandwidth in read and write, but gpu access to system memory would be limited to 16GB/s at best on its pcie slot.
Its comforting to know that while slow running perhaps, no job is too big for your rigs "memory".
"Need" for most, must be balanced with cost.
lets say $1500US for 128GB, & $750 for 64GB ddr4. That's 60-80GB/s to 4 channel TR for the cpu, but importantly, only 16GB/s to the gpu.
Quad 256GB (dont get 128GB, false economy, slow) 3200MB/s sm961 (~ oem Evo nvme) x4 @ $150 US is $600 for 1TB of respectable ~12.5 GB/s on your 3x nvme port TR (+ ~$15~ for a pcie/nvme adaptor for the fourth drive).
Write speeds sadly lag, at about 2/3 that sequential read speed.
Given the nature of swapping, those ideal sequential speeds, reflect the reality of shuffling largish pages of "memory". The task is ideal for achieving high throughput.
Poorer punters could redo the above numbers for 16GB ddr4 & 3x nvme, & find them attractive also.
Its wrong to focus on the above 12GB/s speed. The app continues to use the 64GB e.g. dram, & only when exceeded, does the 12GB/s partially apply.
(of course you could get considerably faster but dearer 500GB 960 pro drives as used by amd, and more drives, but the above is; cheap, no fancy cards, and only needs one slot for a cheap adaptor.)
I can imagine many situations where it's useful to simulate a much more powerful platform – learning, trialling software, students, dilettantes, simulating a similar memory spec alternative to cloud VM workstation fees for lesser jobs like training.
Just musing.... 1TB is more than you would want or is wise? as virtual ram. I suppose we could be guided by the relative size patterns of l1, l2, l3 cache. AMD as I recall, use 256GB of nvme array as cache on their 16GB ssg vega pro card, a factor of 16x, if that's any guide, but it seems a lot for just memory cache. A multiple of 2-3x installed system ram size should work pretty well in most cases imo. It depends on many factors, which are exactly what hbcc is designed to deal with.
A big surprise to vega ssg reviewers, was that due to the nvme virtual cache being non volatile (doh), games etc. which take ages to load into gpu ram normally, are ~instantly ready to resume where you left off and shut down.
Many high end apps (inc. windows) have configurable temp or scratch file locations, so surplus fast array space neednt be wasted.
afaik, TR doesnt allow 3x 16 lane cards, so dual GPUs precludes using 16 lane nvme adaptors.
If you opt for vega, a fast scratch drive is doubly relevant, as vega's hbcc allows gpu cache to be extended in this way also.
folks may wish to note what raid0 isn't any better at than a single ssd, & thats iops. Just saying.
Where is the risk? If it doesnt work out for you on your TR – meh. You have 4x individually very usable 256GB nvme boot drives as upgrades for office desktops to sell or use in house.
I can see the appeal of a well backed up bootable array, but note that nowhere above have I suggested trusting the array with valuable data, so arguments about the increased risks of raid 0 are moot (and a tad anachronistic btw).
Raid 0 nvme is CERTAINLY more reliable than the volatile ram it's standing in for. Good apps will have a fast recovery plan for fully expected power outages, so any swap file raid failure is similarly insured.
The IT industry has a huge problem holding it back. A perpetual~ inability to satisfy demand for dram.
Only AMD offer a glimpse of a partial solution, and all they get is yawns and naysayers.
Nvme is the best thing in years. Amd have given us the lanes on TR, and now the tools to multiply already astonishing nvme storage speeds by a factor of up to 6x, at which point, it approaches 33% of cpu memory bandwidth, and saturates any 16 lane link for reads (~75% for writes).
i dont get this, i have uninstalled AMD chipset drivers and now my PC is flying along, more stable in BF1 as well rapid mode now working in samsung magician
Yes, I find it hard to get any simple synthesis on this software, which it seems doesn't come with oem versions of samsung drives (like the sm961).
afaik, the sm961 has as good or better specs (I fancy the 256GB) , but I hear great things about Magic.
it's further complicated by raid, which appeals to me, and trim when using raid.
It all seems too hard to use oem drives, but they could be a major saving?
Whimsically, nasty gasses aside, I had a good look at a junked fridge today.
Not sure if it was unusual, but it had 2 radiators – a smaller one in the base the size of the fridge footprint.
Plenty of the small dia copper tubing, which could be bent to shape to ensure no joins inside the pc if simply supplementing air cooling.
Its such a big radiator, near silent should be fine.
An aquarium pump may work but could have heat problems.
If water blocks can be bought separately, its a cheap solution.
Another idea is the radiator from a car heater.