DirectStorage feature test discussion and comparisons

[Migz - DH]

3DMark released its DirectStorage feature test today, and I thought that a thread dedicated to discussion of the topic and comparing scores would be useful since there doesn't seem to be an easy way to compare scores through 3DMark's interface for this feature test.

Please feel free to populate the thread with your system specs and scores!

[Migz - DH]

My Build:
OS - Windows 11 Home, (Version 24H2, Build 26100.2454) Windows Feature Experience Pack 1000.26100.36.0
Processor - AMD Ryzen 9 5900X
RAM - 32GB G-Skill DDR4@3600MHz (F4-3600C16D-16GTZN)
Motherboard - ASUS PRIME X570-PRO, BIOS 4802, July 2023
---- AMD Chipset Drivers 6.10.17.152, October 2024
GPU - NVIDIA GeForce RTX 3080 ASUS ROG Strix O10G Gaming
Video Driver - 566.14 - WHQL, November 2024
Dual Monitor
---- LG UltraGear 34GP83A-B @ 3440x1440
---- AOC Agon AG352UCG6 @ 3440x1440
System Drive - WD_Black SN850 2TB, NVMe
Power Supply - KINGWIN Model: ABT-1220MA1S, 1220W, 80 Plus Bronze
Sound Solution - Creative Labs Sound BlasterX G6, firmware 2.1.201208.1030
Sound Driver - 1.16.4.14, October 2020 - Creative Labs

My Score:
http://www.3dmark.com/ds/516

187.7% faster with DirectStorage enabled

Storage to VRAM
DirectStorage on, GDeflate Compression 14.53 GB/s
DirectStorage on, Uncompressed 6.45 GB/s
DirectStorage off, Uncompressed 5.05 GB/s

Storage to RAM
DirectStorage on, Uncompressed 6.43 GB/s
DirectStorage off, Uncompressed 5.21 GB/s

RAM to VRAM
DirectStorage on, Uncompressed 15.81 GB/s
DirectStorage off, Uncompressed 11.96 GB/s
DirectStorage on, GDeflate Compression 52.66 GB/s

[Migz - DH]

Although I am dubious of the advantages to be gained in games with my system, I am genuinely surprised by how little GPU usage was involved in this test,

I am also surprised by the drive temperature graph. In each sub-test, the temperature rises dramatically during the cooling down phase. Still, it's running within the safe temperature envelope. d-(^.^")

[umeng2002]

5800X, RTX 4080 Super, WD SN850X,

Direct Storage Percentage
174
DirectStorage on, GDeflate Compression
15.4 GB/s
DirectStorage off, Uncompressed
5.59 GB/s
DirectStorage on, Uncompressed
6.75 GB/s
DirectStorage off, Uncompressed
6.26 GB/s
Direct Storage on RAM to VRAM
31.6 GB/s
DirectStorage on, Uncompressed
13.7 GB/s
DirectStorage off, Uncompressed
12.0 GB/s
DirectStorage on, GDeflate Compression
61.0 GB/s
DirectStorage on, Uncompressed
6.53 GB/s

http://www.3dmark.com/ds/1340

Now all we need are games to use it.

GPU deflate is the only feature that uses the new stuff. Without GP deflate, it's just an API change... changed for better performance, but still just an API change. GPU deflate is the real feature in Direct Storage.

[SHREVE]

Score
164.8% faster with DirectStorage enabled
Storage to VRAM
DirectStorage on, GDeflate Compression14.79 GB/s
DirectStorage on, Uncompressed6.54 GB/s
DirectStorage off, Uncompressed5.58 GB/s
Storage to RAM
DirectStorage on, Uncompressed5.72 GB/s
DirectStorage off, Uncompressed5.17 GB/s
RAM to VRAM
DirectStorage on, Uncompressed17.17 GB/s
DirectStorage off, Uncompressed14.15 GB/s
DirectStorage on, GDeflate Compression74.16 GB/s

Graphics CardNVIDIA GeForce RTX 4090VendorAsustek Computer, Inc.# of cards1SLI / CrossFireOffMemory24,576 MBClock frequency2,775 MHz (2,235 MHz)Average clock frequencyN/AMemory clock frequency1,313 MHz (1,313 MHz)Average memory clock frequencyN/AAverage temperatureN/ADriver version32.0.15.6614Driver statusApprovedECC video memoryDisabled
Processor
ProcessorIntel Core i9-13900K ProcessorClock frequency5,800 MHz (3,000 MHz)Average clock frequencyN/AAverage temperatureN/APhysical / logical processors1 / 32# of cores24PackageFCLGA1700Manufacturing process7 nmTDP125 W
General
Operating system64-bit Windows 11 (10.0.22631)MotherboardASUSTeK COMPUTER INC. ROG MAXIMUS Z790 HEROMemory32,768 MBModule 116,384 MB G.Skill @ 6,400 MHzModule 216,384 MB G.Skill @ 6,400 MHzHard drive model2,000 GB SHPP41-2000GMVBS statusDisabledHVCI statusDisabled

[Lynxx83]

what I do like,
is that UL is finally releasing something like this after all,
personally I've been asking for a DirectStorage Test long time ago.

what I don't really care about,
is that there's no fancy 3D imagery to look at during tests,
so no Avocados flying around or smth similar,
as we know it from the DirectStorage BulkLoad Demo Bench ;)
tho, it would have been nice.

what I do not like,
is that you can't select/de-select individal sub-tests,
that you can't loop them,
and most of all, that you can't deactivate the 60sec cooldown phase inbetween each sub-test. I guess I know why it's there,
but it should definitely be possible to have the option to disable it.
that would be the most important feature I'm missing right now.

[UL_Jarnis]

The cooldown between the tests is there to ensure you can actually compare devices - without it many devices would thermal throttle depending on the device model and the quality of the heatsink (if any). But we do agree that a custom option to disable the cooldown periods would be useful for some specific testing. Will consider adding it in a future udpate.

Looping or running individual components separately is also something that is possible. Can't promise anything outright, but we are looking into adding more custom options.

No fancy 3D imagery because the test is intended to isolate the storage performance part. It is true that especially the GDeflate part is very very best case scenario (by design) as usually rendering the game would also be taking GPU resources.

[Ext3h]

That "DirectStorage Off" "RAM to VRAM" test is doing something fishy though. Is that thing seriously comparing a low-threadcount copy by shader to just delegating the copy to the copy engine with "DirectStorage On"?

I mean ... that's literally ignoring any "Don't do that" advice about memory uploads given in the past 5 years. If you do copy by shader, you go high occupancy / thread count, or the results look exactly as bad as they do in this benchmark. But generally you shouldn't even try bulk upload on the 3D engine...

Also the choice you made when displaying the "DirectStorage on, GDeflate Compression" number above the "DirectStorage on, Disk -> RAM / RAM -> VRAM" visual display is more than just confusing / misleading. Those numbers don't fit together, at all. If anything you should have added a 3rd display, where you had compared CPU-Deflate vs GPU deflate, and placed that inflated number over there. (But then you should of course compare LZ4 vs GDeflate, as GDeflate is actually not exactly a CPU or GPU friendly algorithm in the first place, and LZ4 at around 4GB/s per CPU core can easily match current generation SSDs speed as well.)

Something is also really weird about the GDeflate benchmark in isolation, somehow it looks as if you had limited that to unpacking only a single resource / single low thread count dispatch, seeing as GPUs of all performance classes hit quite close together regardless of SM count or cache sizes, and also a quite low peak compute engine utilization.


Well, and don't get me started on how the Disk<->RAM portion of the benchmark is literally just an obfuscated IORing vs Overlapped API comparison. Weirdly neither API has been properly pushed to it's limit as both can easily outperform the performance numbers this benchmark was able to get from them. In fact it's a first that I see someone pushing IORing into the CPU limit before the NVMe's device controller is hitting 100% active time. That's a strong indicator of insufficient queue depth on both of the underlying APIs and/or lack of threading.

[Mhill]

14900K + 4080 Super + 4TB Samsung 990 Pro

137.8% faster with DirectStorage enabled

Storage to VRAM
DirectStorage on, GDeflate Compression15.09 GB/s
DirectStorage on, Uncompressed6.59 GB/s
DirectStorage off, Uncompressed6.35 GB/s
Storage to RAM
DirectStorage on, Uncompressed6.56 GB/s
DirectStorage off, Uncompressed6.54 GB/s
RAM to VRAM
DirectStorage on, Uncompressed16.58 GB/s
DirectStorage off, Uncompressed13.69 GB/s
DirectStorage on, GDeflate Compression59.35 GB/s

[Big Italy]

3DMark Link - http://www.3dmark.com/ds/3698

Operating system: 64-bit Windows 10 (10.0.19045)
Motherboard: Gigabyte Technology Co., Ltd. Z590 VISION D
Memory: Corsair Vengeance RGB Pro 32,768 MB
Hard drive: 2GB Sabrent Rocket 4 Plus Gaming
CPU: Intel Core i7-11700K Processor 5,200 MHz (3,600 MHz)
GPU: MSI GeForce RTX 4090 Suprim Liquid X 24G

246.7% faster with DirectStorage enabled

Storage to VRAM
DirectStorage on, GDeflate Compression 14.60 GB/s
DirectStorage on, Uncompressed 6.40 GB/s
DirectStorage off, Uncompressed 4.21 GB/s
Storage to RAM
DirectStorage on, Uncompressed 6.40 GB/s
DirectStorage off, Uncompressed 5.05 GB/s
RAM to VRAM
DirectStorage on, Uncompressed 12.48 GB/s
DirectStorage off, Uncompressed 9.87 GB/s
DirectStorage on, GDeflate Compression 67.03 GB/s

Detailed scores:
Direct Storage Percentage 246
DirectStorage on, GDeflate Compression 14.6 GB/s
DirectStorage off, Uncompressed 4.21 GB/s
DirectStorage on, Uncompressed 6.4 GB/s
DirectStorage off, Uncompressed 5.05 GB/s
Direct Storage on RAM to VRAM 26.8 GB/s
DirectStorage on, Uncompressed 12.5 GB/s
DirectStorage off, Uncompressed 9.87 GB/s
DirectStorage on, GDeflate Compression 67.0 GB/s
DirectStorage on, Uncompressed 6.4 GB/s

[Hampster-Style]

I immediately jumped on the storage bench DLC after this dropped and it has already earned its keep. I had a laptop that would randomly switch between APU and discrete GPU at short quick intervals during operation (power profile related) and this test along with the PCI-E bandwidth one made it simple to quickly identify and address. Big thanks to 3DMark for DirectStorage update.

[Mirec12345]

GPU
NVIDIA GeForce RTX 4070 Laptop GPU (8192 MB)
1333 MHz (1305 MHz)
CPU
13th Gen Intel Core i9-13900HX
2200 MHz (5586 MHz)
32768 MB(2 x DDR5-5586)
Storage
SKHynix_HFS001TEJ9X115N
169 GB (C:)

Score
195.6% faster with DirectStorage enabled
Storage to VRAM
DirectStorage on, GDeflate Compression13.66 GB/s
DirectStorage on, Uncompressed6.19 GB/s
DirectStorage off, Uncompressed4.62 GB/s
Storage to RAM
DirectStorage on, Uncompressed6.08 GB/s
DirectStorage off, Uncompressed4.51 GB/s
RAM to VRAM
DirectStorage on, Uncompressed9.55 GB/s
DirectStorage off, Uncompressed7.99 GB/s
DirectStorage on, GDeflate Compression29.78 GB/s

[Mirec12345]

GPU
Intel(R) UHD Graphics
300 MHz (300 MHz)
CPU
13th Gen Intel Core i9-13900HX
2200 MHz (5586 MHz)
32768 MB(2 x DDR5-5586)
Storage
SKHynix_HFS001TEJ9X115N
169 GB (C:)

Score
85.3% slower with DirectStorage enabled
Storage to VRAM
DirectStorage on, GDeflate Compression0.66 GB/s
DirectStorage on, Uncompressed6.21 GB/s
DirectStorage off, Uncompressed4.50 GB/s
Storage to RAM
DirectStorage on, Uncompressed6.05 GB/s
DirectStorage off, Uncompressed4.62 GB/s
RAM to VRAM
DirectStorage on, Uncompressed9.16 GB/s
DirectStorage off, Uncompressed8.62 GB/s
DirectStorage on, GDeflate Compression1.56 GB/s

[chrcoluk]

Cant run it, as installing just stays stuck at 0%, how did you guys install it?

[chrcoluk]

The reason might be it depends on the normal storage benchmark which I havent paid for, as when I select the directstorage benchmark, the normal storage benchmark is stuck on installing as well.

[Your Long Lost Cuzin]

Is this top 100 for "Storage to VRAM" ??

SPECS
AMD Ryzen 9 7950X3D 16-Core Processor
CT4000T705SSD3
PCIe 5.0x4 (Max. 16 GB/s)
Firmware: PACR5111
AMD Radeon RX 7900 XTX
PCIe 4.0 x16
AMD Radeon RX 7900 XTX
PCIe 4.0 x16
AMD Radeon(TM) Graphics
PCIe 4.0 x16


DirectStorage feature test
Valid result
Score 190.0% faster with DirectStorage enabled

Storage to VRAM

DirectStorage on, GDeflate Compression
23.48 GB/s

DirectStorage on, Uncompressed
11.22 GB/s

DirectStorage off, Uncompressed
8.10 GB/s

Storage to RAM

DirectStorage on, Uncompressed
11.22 GB/s

DirectStorage off, Uncompressed
7.69 GB/s

RAM to VRAM

DirectStorage on, Uncompressed
13.02 GB/s

DirectStorage off, Uncompressed
11.27 GB/s

DirectStorage on, GDeflate Compression
65.10 GB/s

Drive model CT4000T705SSD3
Firmware PACR5111