GPU Megat[H]read - the cores of wrath grew heavy on the die that day

The Something Awful Forums > Discussion > Serious Hardware/Software Crap > GPU Megat[H]read - the cores of wrath grew heavy on the die that day

«‹›3877 »

AutismVaccine: Feb 26, 2017; SPECIAL NEEDS
SQUAD

Rinkles posted:

There�s a utility for testing direct storage speeds, and that�s in line with what I saw comparing a fairly slow nvm with SATA. I�ll post it when I�m at a PC.

1080Ti, 5800X, 980Pro 1TB M.2

# ? Jan 25, 2023 14:45

Adbot: ADBOT LOVES YOU

# ? Jun 5, 2024 21:00

power crystals: Jun 6, 2007; Who wants a belly rub??

Inland "Performance Plus" 4TB, no idea on any further model number, with a 3080 10GB:

Apparently this thing is better than I thought!

e: VVV this is W10, for what that's worth.

power crystals fucked around with this message at 15:12 on Jan 25, 2023

# ? Jan 25, 2023 15:03

repiv: Aug 13, 2009

shrike82 posted:

wow, the directstorage stuff shows a big gap between m2 and sata ssds
https://www.youtube.com/watch?v=8taBd7EO2Dc

someone also needs to test win10 vs win11 to see if the fast path makes any real difference in practice

if you're on 11 you can test if the fast path is active like so (it only works with nvme and some 3rd party drivers might break it)

code:

(as admin)
> fsutil bypassIo state c:\
BypassIo on "c:" is currently supported
    Storage Type:   NVMe
    Storage Driver: BypassIo compatible

# ? Jan 25, 2023 15:07

gradenko_2000: Oct 5, 2010; HELL SERPENT; Lipstick Apathy

if I'm understanding this technology correctly, you can already achieve fairly high data transfer rates just by the native NVme / PCIe interface, but the problem is that all that data has to be uncompressed, which means games will be large

what directStorage does is it allows for GPU-powered decompression, such that you can take a big game, compress it down to a relatively much smaller size, and then use the GPU to decompress it in "real-time", but doing so at a speed that makes it relatively un-noticeable, that you couldn't do before because CPU-powered decompression would take so long that it wouldn't work

in the end, games can take up less space on a drive, while still being quick to load-up and play (or even quicker than without this tech)?

# ? Jan 25, 2023 15:17

repiv: Aug 13, 2009

directstorage is a few different things under one umbrella

1.0 was focused on minimising overhead for disk to CPU transfers, similar to io_uring on linux or analogous to DX12 but for storage. compression wasn't integrated but the game could still decompress data manually after it arrives on the CPU.

1.1 added compression, using a codec developed by nvidia but now open sourced. transfers still have to go via the CPU, but data can remain compressed all the way through the PCIe bus and be unpacked on the GPU itself. currently this is based on compute shaders as no current GPUs have hardware decompression (ala the consoles), but the API seems to be deliberately designed to leave that option open for future GPU architectures.

the next step would be to cut the CPU out of the loop entirely, and have the SSD DMA data directly to the GPU

# ? Jan 25, 2023 15:24

repiv: Aug 13, 2009

not sure if forspoken is taking advantage of GPU decompression, that feature only landed in november and that's pretty late to be retooling the asset pipeline like that

# ? Jan 25, 2023 15:28

Twerk from Home: Jan 17, 2009; This avatar brought to you by the 'save our dead gay forums' foundation.

gradenko_2000 posted:

if I'm understanding this technology correctly, you can already achieve fairly high data transfer rates just by the native NVme / PCIe interface, but the problem is that all that data has to be uncompressed, which means games will be large

what directStorage does is it allows for GPU-powered decompression, such that you can take a big game, compress it down to a relatively much smaller size, and then use the GPU to decompress it in "real-time", but doing so at a speed that makes it relatively un-noticeable, that you couldn't do before because CPU-powered decompression would take so long that it wouldn't work

in the end, games can take up less space on a drive, while still being quick to load-up and play (or even quicker than without this tech)?

I haven't seen detail about compression ratios yet using directstorage, but Autism Vaccine's example doesn't look too much faster than LZ4 on a CPU, which can already decompress at a couple GB/s on just a single core: https://github.com/lz4/lz4

I'm really optimistic about this tech, but also think that PCs with their much faster CPUs than consoles and future sky-high core counts probably don't need GPU based decompression as much as the consoles do.

PowerCrystal's example though, that is super impressive. That's faster than memcpy on a Core i7-9700K CPU reference system. I just realized that the whole point of this is that you direct load textures from disk to GPU without touching main memory. Awesome tech.

# ? Jan 25, 2023 15:30

repiv: Aug 13, 2009

even if the CPU is up to the task of decompressing the data, it comes with the disadvantage of having to shove the decompressed data over PCIe to GPU memory

part of the benefit of moving decompression to the GPU is that data can remain in compressed form as it travels over PCIe saving a bunch of bandwidth

Twerk from Home posted:

I just realized that the whole point of this is that you direct load textures from disk to GPU without touching main memory. Awesome tech.

not quite yet, DS 1.1 still only supports disk to CPU transfers, so data headed for the GPU still has to hop via the CPU first

it's technically feasible for them to do direct NVMe to GPU transfers though, and that's surely what they're working towards

repiv fucked around with this message at 15:46 on Jan 25, 2023

# ? Jan 25, 2023 15:35

Gyrotica: Nov 26, 2012; Grafted to machines your builders did not understand.

Zedsdeadbaby posted:

Uuuugh oh my god

Why do they hate textures?

# ? Jan 25, 2023 16:34

Fats: Oct 14, 2006; What I cannot create, I do not understand; Fun Shoe

repiv posted:

someone also needs to test win10 vs win11 to see if the fast path makes any real difference in practice

This is on Win11 with the bypassIo stuff enabled, 4090 and a 980 pro nvme drive

Edit: whoops, forgot to add, 5950x for the processor

Fats fucked around with this message at 17:17 on Jan 25, 2023

# ? Jan 25, 2023 16:47

repiv: Aug 13, 2009

well that's certainly much lower CPU overhead than the win10 result above (assuming you have similar-ish CPUs)

# ? Jan 25, 2023 16:52

UHD: Nov 11, 2006

win11, 980 pro nvme, 4070 ti, i5-12600K cpu

e: "Storage Driver: Not BypassIo Compatible"

UHD fucked around with this message at 18:00 on Jan 25, 2023

# ? Jan 25, 2023 17:00

Wiggly Wayne DDS: Sep 11, 2010

3080 10GB, 8700k:
960 EVO 500GB: 5.85GB/s
850 EVO 500GB: 2.00GB/s

# ? Jan 25, 2023 17:01

MarcusSA: Sep 23, 2007

Ok what�s the deal with the different file sizes here?

# ? Jan 25, 2023 17:01

repiv: Aug 13, 2009

it looks like it's roughly correlated with the amount of VRAM available, it fills about half of it

# ? Jan 25, 2023 17:05

power crystals: Jun 6, 2007; Who wants a belly rub??

repiv posted:

well that's certainly much lower CPU overhead than the win10 result above (assuming you have similar-ish CPUs)

For mine at least I have a 5800X3D but I have no idea how the extra cache affects this.

# ? Jan 25, 2023 17:20

Zero VGS: Aug 16, 2002; ASK ME ABOUT HOW HUMAN LIVES THAT MADE VIDEO GAME CONTROLLERS ARE WORTH MORE; Lipstick Apathy

Is there any way to �force� VRR for a 12th gen Intel iGPU laptop? The option doesn�t appear in the Intel or Arc control panels and I see no reason why it shouldn�t be supported. This is a late 2022 Asus laptop with a 120hz OLED.

Also, the Intel control panel complains that resizable bar isn�t on. No sign of it whatsoever in the laptop bios. Again, shouldn�t this poo poo be enabled in really recent hardware?

Zedsdeadbaby posted:

Uuuugh oh my god

I got booed out of the Final Fantasy thread months ago for complaining about all the awkward anime-style grunts in the voice acting of the FF7 remake. I didn�t realize �American-style� could be even worse.

# ? Jan 25, 2023 17:26

Evil Fluffy: Jul 13, 2009; Scholars are some of the most pompous and pedantic people I've ever had the joy of meeting.

What's the speed/result like with a WD SN770 TB2, or equivalent?

# ? Jan 25, 2023 17:30

Theophany: Jul 22, 2014; SUCCHIAMI IL MIO CAZZO DA DIETRO, RANA RAGAZZO

2022 FIA Formula 1 WDC

7900XTX, 3900X, 1TB Sabrent Rocket 4.0

# ? Jan 25, 2023 17:45

repiv: Aug 13, 2009

if more people note whether they have bypassio active (you definitely don't if you're still on win10) then we can better determine how much load it takes off the CPU

code:

(as admin)
> fsutil bypassIo state c:\
BypassIo on "c:" is currently supported
    Storage Type:   NVMe
    Storage Driver: BypassIo compatible

# ? Jan 25, 2023 17:52

kliras: Mar 27, 2021

discord finally rolling out av1 support

https://twitter.com/gerdelgado/status/1618285964308402180

# ? Jan 25, 2023 17:55

Rinkles: Oct 24, 2010; What I'm getting at is...
Do you feel the same way?

repiv posted:

if more people note whether they have bypassio active (you definitely don't if you're still on win10) then we can better determine how much load it takes off the CPU

It was on for me

Rinkles posted:

WD SN550

WD BLUE SATA

w/ a 3060ti

# ? Jan 25, 2023 17:56

Lockback: Sep 3, 2006; All days are nights to see till I see thee; and nights bright days when dreams do show me thee.

Gunna connect to people's streams with my InternetTV box just to ruin everyone's good time.

# ? Jan 25, 2023 18:00

repiv: Aug 13, 2009

getting 5GB/sec on my 970 EVO 1TB but 9GB/sec on my SN550 2TB :thunk:

both have bypassio, both are pcie3, but the latter is a slower drive by conventional metrics

what's going on there then

e: maybe something to do with the 970 being NVMe 1.3 and the SN550 being NVME 1.4? bypassio fast path taking advantage of some newer NVMe feature(s)?

repiv fucked around with this message at 18:26 on Jan 25, 2023

# ? Jan 25, 2023 18:01

Rinkles: Oct 24, 2010; What I'm getting at is...
Do you feel the same way?

Re-running it on the nvme, there seems to be a fair bit of variability in CPU use

# ? Jan 25, 2023 18:03

njsykora: Jan 23, 2012; Robots confuse squirrels.

kliras posted:

discord finally rolling out av1 support

https://twitter.com/gerdelgado/status/1618285964308402180

Love the concept of a no poors allowed Discord stream.

# ? Jan 25, 2023 18:12

Cross-Section: Mar 18, 2009

5900x/4090/GIGABYTE AORUS NVMe Gen4 1TB, bypassio enabled

I still don't really know what any of this means :shrug:

# ? Jan 25, 2023 18:12

kliras: Mar 27, 2021

njsykora posted:

Love the concept of a no poors allowed Discord stream.

it's horribly worded in general. enabling software decode sounds like the most sensible choice, so hopefully it will be available on launch

# ? Jan 25, 2023 18:18

UHD: Nov 11, 2006

repiv posted:

if more people note whether they have bypassio active (you definitely don't if you're still on win10) then we can better determine how much load it takes off the CPU
code:
(as admin)
> fsutil bypassIo state c:\
BypassIo on "c:" is currently supported
    Storage Type:   NVMe
    Storage Driver: BypassIo compatible

my results didn't have bypassio enabled and damned if i know how to turn it on :shrug:

board is msi pro z690-a ddr4

code:

BypassIo on "c:" is partially supported
    Volume stack bypass is disabled (EPMVolFl.sys)
    Storage Type:   NVMe
    Storage Driver: Not BypassIo Compatible

UHD posted:

win11, 980 pro nvme, 4070 ti, i5-12600K cpu

e: "Storage Driver: Not BypassIo Compatible"

# ? Jan 25, 2023 18:33

gradenko_2000: Oct 5, 2010; HELL SERPENT; Lipstick Apathy

Ampere doesn't support AV1?

# ? Jan 25, 2023 18:36

repiv: Aug 13, 2009

UHD posted:

my results didn't have bypassio enabled and damned if i know how to turn it on board is msi pro z690-a ddr4

apparently EPMVolFl.sys is used by EaseUS partition tools, if you have that installed it may be the culprit

gradenko_2000 posted:

Ampere doesn't support AV1?

it has decode, not encode

# ? Jan 25, 2023 18:51

UHD: Nov 11, 2006

repiv posted:

apparently EPMVolFl.sys is used by EaseUS partition tools, if you have that installed it may be the culprit

that was it, uninstalling EaseUS and rebooting worked, thank you

this appears to be a best case screenshot. on average - because the numbers vary quite a bit from run to run - bandwidth went up but cpu usage didn't really change

e: from when bypassio was off vs on

UHD fucked around with this message at 19:16 on Jan 25, 2023

# ? Jan 25, 2023 19:13

Rinkles: Oct 24, 2010; What I'm getting at is...
Do you feel the same way?

UHD posted:

that was it, uninstalling EaseUS and rebooting worked, thank you

this appears to be a best case screenshot. on average - because the numbers vary quite a bit from run to run - bandwidth went up but cpu usage didn't really change

e: from when bypassio was off vs on

that's a pretty big jump

# ? Jan 25, 2023 19:16

Nfcknblvbl: Jul 15, 2002

So many avocados in this thread, it�s making me hungry!

# ? Jan 25, 2023 19:22

mobby_6kl: Aug 9, 2009; by Fluffdaddy

i5-3470, GTX 1070, some sort of WD SATA SSD

Run trash computer, be free.

# ? Jan 25, 2023 19:45

Kibner: Oct 21, 2008; Acguy Supremacy

5950x in Eco Mode
32 GB of 3200MHz ECC DDR4 RAM (so, very slow timings)
Asus Pro WS X570-ACE
MSI 1070 (at factory mildly OC'd settings)

480GB Intel Optane 905p NVMe

2TB Intel 660p NVMe

I assume both drives are running at full bandwidth and not limited because I have something plugged into the wrong SATA slot.

e: neither drive is BypassIo Compatible (not sure how to enable; both say "partially supported")

Kibner fucked around with this message at 20:41 on Jan 25, 2023

# ? Jan 25, 2023 20:37

repiv: Aug 13, 2009

Kibner posted:

e: neither drive is BypassIo Compatible (not sure how to enable; both say "partially supported")

if there's a driver conflicting then fsutil should name it, as it did for UHD

# ? Jan 25, 2023 20:43

Shipon: Nov 7, 2005

UHD posted:

my results didn't have bypassio enabled and damned if i know how to turn it on board is msi pro z690-a ddr4
code:
BypassIo on "c:" is partially supported
    Volume stack bypass is disabled (EPMVolFl.sys)
    Storage Type:   NVMe
    Storage Driver: Not BypassIo Compatible

have the same board as you but got this problem instead (12900k, 4090, 970 Evo Plus 2TB), weird

code:

> fsutil bypassIo state c:\
BypassIo on "c:" is not currently supported
    HResult: 0x80820003 (BypassIO cannot be enabled while a volume snapshot exists)
    Driver:  volsnap.sys
    Reason:  One or more snapshots are active on the volume
    Storage Type:   NVMe
    Storage Driver: BypassIo compatible

# ? Jan 25, 2023 20:46

Kibner: Oct 21, 2008; Acguy Supremacy

repiv posted:

if there's a driver conflicting then fsutil should name it, as it did for UHD

I don't see anything giving a reason. Am I missing something?

code:

PS C:\Users\kibne> fsutil bypassIo state c:\
BypassIo on "c:" is partially supported
    Volume stack bypass is disabled (IaNVMe.sys)
    Storage Type:   NVMe
    Storage Driver: Not BypassIo Compatible
PS C:\Users\kibne> fsutil bypassIo state d:\
BypassIo on "d:" is partially supported
    Volume stack bypass is disabled (IaNVMe.sys)
    Storage Type:   NVMe
    Storage Driver: Not BypassIo Compatible

edit: is it that IaNVMe.sys file?

# ? Jan 25, 2023 20:48

Adbot: ADBOT LOVES YOU

# ? Jun 5, 2024 21:00

repiv: Aug 13, 2009

yeah IaNVMe.sys is blocking it, apparently that's the intel NVMe driver

AFAIK there's no reason to use manufacturer NVMe drivers nowadays? bypassio only works with the generic windows NVMe driver in any case

# ? Jan 25, 2023 20:51

The Something Awful Forums > Discussion > Serious Hardware/Software Crap > GPU Megat[H]read - the cores of wrath grew heavy on the die that day

«‹›3877 »