Blazing Fast NVMEs and Direct Storage API for PCs *spawn*
- Thread starter DavidGraham
- Start date
I would assume that's exactly how it's configured. You have 16x PCIe5 lanes dedicated to the GPU and then another 4x PCI4 lanes dedicated to the first M2 slot. However it may be possible to configure in a different way in the BIOS. Perhaps 8x PCIe5 for the GPU and then 4xPCIe5 for the NVMe. If this kind of messing around is required though then it suggests there won't be any consumer level PCIe5 drives available to take advantage of it.
Then again, even with a fast PCIe4 drive you should be able to fill the VRAM of an RTX 3090 in under 2 seconds so is there really much real world benefit of a drive that's twice as fast?
Rootax
Veteran
For home users , I don't see the urgent need, even if you can't stop this kind of progression. But, if this is really hard to implement for motherboard manufacturers, I guess only the server market in need of more and more bandwitdh will be interested at first. Even if some proprietary techs are already there to answer some needs.
DmitryKo
Veteran
NVMe 2.0 has been released, includes specifications for ZNS (Zoned Namespaces) command set, Simple Copy Command, Endurance Groups, Key-Value command set, and Rotational media.
https://nvmexpress.org/nvme-2-0-specifications-infographic/
https://nvmexpress.org/everything-y...0-specifications-and-new-technical-proposals/
https://www.anandtech.com/show/16702/nvme-20-specification-released
https://www.tomshardware.com/news/nvme-2-0-supports-hard-disk-drives
Typical PCIe 4.0 NVMe controllers only have 3 (three) ARM Cortex cores even for high-end products, with the same amount of flash memory channels (8 or 16), and they already require large heatsinks to operate reliably.
Clearly something needs to be done about LBA sector translation and garbage collection/wear levelling overhead before PCIe 5.0 x4 speeds of 16 GByte/s (and PCIe 6.0 x4 at 32 GByte/s) become viable for consumer drives.
Explicit namespace management (ZNS) does help reduce controller overhead, but it's mostly targeted at enterprise storage at the moment...
https://nvmexpress.org/nvme-2-0-specifications-infographic/
https://nvmexpress.org/everything-y...0-specifications-and-new-technical-proposals/
https://www.anandtech.com/show/16702/nvme-20-specification-released
https://www.tomshardware.com/news/nvme-2-0-supports-hard-disk-drives
It's not just about the availability of PCIe 5.0 lanes in the host CPU. If you look at Marvell Bravera SC5 specs above, they include no less than 10 (ten) ARM Cortex-R8 and Cortex-M7 cores dedicated to LBA sector processing, and an additional Cortex-M3 core for TCG Opal encryption.
Typical PCIe 4.0 NVMe controllers only have 3 (three) ARM Cortex cores even for high-end products, with the same amount of flash memory channels (8 or 16), and they already require large heatsinks to operate reliably.
Clearly something needs to be done about LBA sector translation and garbage collection/wear levelling overhead before PCIe 5.0 x4 speeds of 16 GByte/s (and PCIe 6.0 x4 at 32 GByte/s) become viable for consumer drives.
Explicit namespace management (ZNS) does help reduce controller overhead, but it's mostly targeted at enterprise storage at the moment...
Last edited:
No pcie5 in sight for amd based on this leak.
https://videocardz.com/newz/amd-raphael-zen4-am5-presentation-from-march-2020-leaks-out
https://videocardz.com/newz/amd-raphael-zen4-am5-presentation-from-march-2020-leaks-out
Attachments
No pcie5 in sight for amd based on this leak.
https://videocardz.com/newz/amd-raphael-zen4-am5-presentation-from-march-2020-leaks-out
They're code naming their CCD's Durango? Hmmm.
https://nvmexpress.org/everything-y...0-specifications-and-new-technical-proposals/
https://nvmexpress.org/specifications/
Some intresting NVME2.0 specs.
https://nvmexpress.org/specifications/
Some intresting NVME2.0 specs.
I would guess by the time PCI-5 is adopted on motherboards and consumer drive reaching the 16GByte speeds come they will have moved on to newer arm processors
Aside from that for gaming wouldn't it just be easier to put an ssd directly on the graphics card ? Cut out the cpu completely in that regard
DmitryKo
Veteran
Cortex-M and Cortex-R aren't full-featured processors, but embedded microcontroller cores built for very low power consumption.
That would offer no additional incentives over simply using NVMe dedicated PCIe lanes on the CPU, as provided in Ryzen and Rocket Lake. You cannot cut out the CPU because the OS has to perform disk I/O; GPU cannot process cluster chains and decode LBA sectors or issue NVMe commands.
In theory it is possible to use P2P (peer-to-peer) DMA mode of the PCI Express bus to send data from NVMe disk directly to GPU local memory, bypassing system memory, but storage and display driver models would have to be updated to support P2P DMA operations.
Radeon SSG software libraries were simply using the SSD RAID disk as a very fast file cache - it wasn't directly accessible from the GPU, as product ads led us to believe.
Last edited:
DmitryKo
Veteran
Yep, Windows 11 and 1 TB NVMe only.
https://news.xbox.com/en-us/2021/06/24/windows-11-the-best-windows-ever-for-gaming/
https://wccftech.com/windows-11-will-be-required-to-run-the-new-directstorage-api/
Well, I already have an 1 TB PCIe 3.0 disk; look forward to see the benchmarks to consider an upgrade to a PCIe 4.0 model...
https://news.xbox.com/en-us/2021/06/24/windows-11-the-best-windows-ever-for-gaming/
https://wccftech.com/windows-11-will-be-required-to-run-the-new-directstorage-api/
Well, I already have an 1 TB PCIe 3.0 disk; look forward to see the benchmarks to consider an upgrade to a PCIe 4.0 model...
darn! I only have a 256GB NVMe and making as most space as I can for my favourite games, living a life of witchery.
I wonder why a 1GB drive is required though? The Series S doesn't have a 1TB drive.
Do 1TB drives have a different controller which allows for something necessary that the 500GB and lower drives don't?
DmitryKo
Veteran
I think that was expected, considering they are revamping the entire I/O Manager stack to support batch I/O requests.
1 TB requirement probably has to do with the size of typical game assets rather than performance.
That said, current 1 TB models typically have significantly better IOPs performance and sequential write speeds, due to additional flash memory channels and bigger DRAM cache; 256 GB variants can be 2-3 times slower, and 512 GB variants can be 1.5-2 times slower.
Last edited:
- DirectStorage requires 1 TB or greater NVMe SSD to store and run games that uses the "Standard NVM Express Controller" driver and a DirectX 12 Ultimate GPU.
Interesting to see the NVMe requirement confirmed though. So no SATA SDD support. It's good for devs I suppose as it means they can guarantee at least similar to XSX performance from their DirectStorage path in games.
DmitryKo
Veteran
Maybe they will automatically re-partition the disk to include a separate gaming-only partition, as we discussed earlier? I still think that would be too much hassle for most users...
Anyway, for those with UEFI capable PCs who still use MBR system disks, you will need to convert your disk to UEFI/GPT partition layout even if you're still using HDD or SATA SSD, because CSM has to be disabled to enable TPM 2.0 and UEFI Secure Boot. This will also make the inevitable transition to a bootable NVMe system disk much easier.
Anyway, for those with UEFI capable PCs who still use MBR system disks, you will need to convert your disk to UEFI/GPT partition layout even if you're still using HDD or SATA SSD, because CSM has to be disabled to enable TPM 2.0 and UEFI Secure Boot. This will also make the inevitable transition to a bootable NVMe system disk much easier.
Last edited:
why 1tb mines 960gb
