WEBVTT

00:00:00.400 --> 00:00:09.519
one of the big stories at computex this year was about removing the bottlenecks

00:00:05.600 --> 00:00:12.240
from high-speed m.2 NVMe ssds pretty

00:00:09.519 --> 00:00:18.560
much every motherboard was bristling with mounts and partners like ASUS even

00:00:15.280 --> 00:00:21.840
showed off this crazy card that could

00:00:18.560 --> 00:00:25.439
hold four drives at a time for a

00:00:21.840 --> 00:00:26.880
theoretical 16 gigabytes per second of

00:00:25.439 --> 00:00:32.640
throughput so when samsung approached us to sponsor

00:00:29.439 --> 00:00:36.000
a video showing off their flagship 960

00:00:32.640 --> 00:00:39.520
pro we had plenty of inspiration for how

00:00:36.000 --> 00:00:42.320
to build the awesomest dream SSD raid

00:00:39.520 --> 00:00:44.960
setup unfortunately getting it to actually work

00:00:43.680 --> 00:00:50.860
has been well an adventure yeah

00:00:48.480 --> 00:00:55.840
let's put it that way

00:00:58.960 --> 00:01:06.479
now as the de facto choice for speed freaks for the entire past year

00:01:03.039 --> 00:01:08.080
samsung's 960 pro needs no introduction

00:01:06.479 --> 00:01:15.040
but they sponsored this video so we're gonna do it anyway each of our four

00:01:10.880 --> 00:01:17.840
drives has 512 gigs of vnan storage is

00:01:15.040 --> 00:01:23.040
rated to a blazing fast 3 500 megabytes per second reads and 2100 megabytes per

00:01:20.479 --> 00:01:28.479
second writes has a custom five core samsung processor on board and a

00:01:25.840 --> 00:01:33.040
five-year warranty and somehow the whole thing is the size of a stick of gum

00:01:32.240 --> 00:01:38.840
so then let's go let's plug them all in and

00:01:36.320 --> 00:01:44.960
rip up some benchmark scores right wrong finding a board with enough slots

00:01:42.320 --> 00:01:49.840
was the first challenge we had to solve i thought i saw one at computex that had

00:01:47.600 --> 00:01:55.600
five but that turned out to be a figment of my imagination so we settled then on

00:01:52.399 --> 00:01:58.079
the ASUS prime x299 deluxe with a

00:01:55.600 --> 00:02:04.560
separate pci express card to handle the last m.2 but one small problem

00:02:01.680 --> 00:02:09.520
as cool as it is to run Windows software raid for eye-watering sequential

00:02:06.640 --> 00:02:14.120
performance numbers that has been done to death

00:02:10.879 --> 00:02:18.160
and besides we wanted to actually

00:02:14.120 --> 00:02:21.040
experience the speed in day-to-day use

00:02:18.160 --> 00:02:26.879
so that means that we need to boot Windows from our array

00:02:23.920 --> 00:02:32.160
let's talk about Intel's virtual radon CPU or vroc

00:02:29.840 --> 00:02:37.280
since the latest skylakex high-end desktop processors share much of their

00:02:34.400 --> 00:02:43.040
pedigree with server level xeon chips they actually have three what are called

00:02:40.239 --> 00:02:50.400
volume management devices built in each of which can give up to four PCIe

00:02:46.480 --> 00:02:53.760
4x devices direct access to the CPU for

00:02:50.400 --> 00:02:54.640
high bandwidth low latency performance

00:02:53.760 --> 00:03:00.560
cool the issue though is that today

00:02:57.840 --> 00:03:05.440
it is not yet officially launched and making matters worse Intel is rumored to

00:03:03.280 --> 00:03:11.040
be planning to enable the feature with a hardware key that will only be available

00:03:08.080 --> 00:03:17.360
through bundles with x299 motherboards or ssds thankfully though AMD decided to

00:03:14.959 --> 00:03:23.200
ride in on a white steed and save the day by announcing a free

00:03:20.000 --> 00:03:25.040
driver update and sadly yes their

00:03:23.200 --> 00:03:28.560
marketing materials specifically emphasize the free part that's the world

00:03:27.200 --> 00:03:35.280
we live in now which would enable bootable NVMe raid on

00:03:31.599 --> 00:03:38.159
the x 399 threadripper platform

00:03:35.280 --> 00:03:42.959
so a quick swap to the ASUS zenith extreme and we are ready to

00:03:40.720 --> 00:03:47.840
cry on the desk because AMD pulled the software almost

00:03:45.840 --> 00:03:53.120
immediately after the announcement due to compatibility issues

00:03:50.640 --> 00:03:57.680
once it was finally re-released we then had to bungle our way through the

00:03:54.879 --> 00:04:04.400
incredibly poorly documented process of a loading the AMD NVMe controller driver

00:04:01.120 --> 00:04:07.200
which allows b the AMD raid controller

00:04:04.400 --> 00:04:12.000
driver to be installed which allows c the AMD raid configurator driver to be

00:04:10.480 --> 00:04:19.680
installed then even when we were in Windows running on

00:04:15.599 --> 00:04:22.079
four ssds we hit another roadblock

00:04:19.680 --> 00:04:25.919
our iometer benchmark results were nowhere near what AMD posted on their

00:04:24.880 --> 00:04:30.960
blog so after still more research

00:04:28.800 --> 00:04:33.840
mad props to gary from ASUS by the way man

00:04:31.840 --> 00:04:40.720
we figured out that because a threadripper CPU is technically two

00:04:37.040 --> 00:04:42.560
separate dies linked by AMD's infinity

00:04:40.720 --> 00:04:47.919
fabric interconnect considerations must be made to ensure

00:04:45.120 --> 00:04:52.400
that the NVMe load is balanced between the dies

00:04:49.199 --> 00:04:55.120
otherwise this 11 Gigabyte per second

00:04:52.400 --> 00:04:57.919
link as fast as it is could actually become a bottleneck

00:04:57.040 --> 00:05:03.600
so armed with a detailed diagram of PCIe

00:05:00.639 --> 00:05:08.160
lane allocations and a pre-release BIOS we pushed onward meanwhile though we

00:05:06.080 --> 00:05:12.240
actually hadn't given up entirely on v-rock

00:05:09.440 --> 00:05:16.080
now while i wouldn't recommend getting your Intel drivers from russian download

00:05:14.960 --> 00:05:22.000
mirrors we did find one that claimed to enable

00:05:18.800 --> 00:05:23.600
v-rock without a raid key and actually

00:05:22.000 --> 00:05:27.759
it technically worked but our performance numbers were way off

00:05:25.520 --> 00:05:33.360
compared to the guys at pc perspective who also got v-rock working via what we

00:05:30.080 --> 00:05:36.000
think was a similar method and anyway

00:05:33.360 --> 00:05:39.680
neither of us could get it to boot with non-Intel ssds

00:05:38.720 --> 00:05:45.600
so back to AMD threadripper then which for better or for worse has bootable raid

00:05:44.320 --> 00:05:49.759
today AMD got back to us with a preset for

00:05:47.600 --> 00:05:53.360
iometer and some guidance to help us replicate their results

00:05:51.600 --> 00:05:57.840
not while actually booted from it mind you but with a raw as an unformatted

00:05:56.000 --> 00:06:01.919
file system and we implemented everything meaning that it is finally

00:05:59.759 --> 00:06:07.919
time to do this so after physically balancing the four

00:06:04.639 --> 00:06:10.960
samsung 960 pros between the dies using

00:06:07.919 --> 00:06:14.319
ASUS's dim.2 for two of them and their

00:06:10.960 --> 00:06:17.280
hyper m.2 x 16 card for the other two we

00:06:14.319 --> 00:06:20.639
booted in pure UEFI mode and configured our array using the built-in raid expert

00:06:19.360 --> 00:06:27.120
2 utility we then f6 all three drivers in order

00:06:24.479 --> 00:06:31.600
and saw our 2 terabyte array as available for os installation

00:06:29.919 --> 00:06:36.960
once booted into Windows we launched diameter with one megabyte reads and

00:06:33.759 --> 00:06:39.840
writes at 32q depth and had to rub our

00:06:36.960 --> 00:06:45.680
eyes at the results 12 gigabytes per second on reads and 7.3 gigabytes per

00:06:43.199 --> 00:06:53.120
second on rights and that is without AMD's bs raw file system stuff we are

00:06:49.440 --> 00:06:56.560
talking a fully operational ntfs

00:06:53.120 --> 00:06:58.160
formatted bootable array with those

00:06:56.560 --> 00:07:02.800
kinds of numbers for comparison a single 960 pro

00:07:00.720 --> 00:07:08.319
delivered three and a half gigabytes per second on reads and two gigabytes per

00:07:05.280 --> 00:07:10.319
second on writes in the same machine now

00:07:08.319 --> 00:07:14.080
crystaldiskmark didn't scale quite as well on reads but

00:07:12.080 --> 00:07:19.280
we do have to keep in mind that this technology is still in its infancy at

00:07:17.039 --> 00:07:24.800
least on the driver's side and we might actually see more of the raw hardware's

00:07:21.840 --> 00:07:29.680
potential unlocked in the future so thanks to samsung for sponsoring this

00:07:26.880 --> 00:07:35.440
exploration of high-end bootable raid on the desktop i don't think this crazy

00:07:32.639 --> 00:07:39.360
ride is quite over yet but we're gonna take a little breather and then maybe

00:07:37.199 --> 00:07:44.800
we'll revisit it once things have settled down particularly on the Intel

00:07:41.520 --> 00:07:46.160
side over the next few months so thanks

00:07:44.800 --> 00:07:51.120
you guys for watching if you disliked this video you can hit that button but if you liked it hit like get subscribed

00:07:49.680 --> 00:07:55.120
maybe consider checking out where to buy the stuff we featured at the link in the

00:07:52.960 --> 00:07:58.720
video description also down there is our merch store which has cool shirts like

00:07:56.479 --> 00:08:02.680
this one as well as our community forum which you should totally join