Blazing Fast NVMEs and Direct Storage API for PCs *spawn*

I didn't see exactly where it came from, but it smelled like PR of some sort. It might have just been from an anti-Windows pro-Linux writer.
It's referred to as proprietary because Microsoft have proprietary rights to the API. DirectX is proprietary, WMV and WMA were proprietary media formats.

These may be used by the biggest desktop operating system, but they remain proprietary. :yes:
 
Maybe this is an idiotic question, but if you have and APU+GPU system, could you use the graphic side of the APU in the "Custom Decompression Queue"?
 
Maybe this is an idiotic question, but if you have and APU+GPU system, could you use the graphic side of the APU in the "Custom Decompression Queue"?

In theory you should be able to as it's likely that MS will use compute for that purpose, at least initially, which should work on any GPU of X level of compute capability. Integrated graphics aren't disabled as, for example, you can drive a monitor with the integrated GPU on a desktop system while simultaneously driving a monitor on a dedicated GPU.

If it's performant enough at GPU based decompression, I see little to no reason why they wouldn't want to leverage the capabilities of the integrated GPU for such tasks, especially if it's for assets whose destination is main system memory versus the GPUs VRAM. In the latter case it might be preferable to have the dedicated GPU do it if it results in lower latencies and lower latencies is desirable. Although it could still be potentially more beneficial to leverage the integrated GPU reduce work done by the dedicated GPU WRT non rendering related work.

Regards,
SB
 
Adding onto what Silent_Buddha just mentioned, native DX12 SM6 hardware fully supports heterogenous compute models for what seems like basically the whole pipeline. Ostensibly even today a developer who really wanted to could create an application to "meld" IGP/APU and add-on GPU devices, or even add-on GPUs from completely different vendors or different classes/subclasses.

It stands to reason a developer could spend the time to create a platform where an otherwise unused IGP/APU could offload work on a system with a big add-on GPU. I have no idea how difficult that may be; naively I might presume it's easier to "offload storage" to one GPU and "everything else" to another GPU? I have no clue, honestly.
 
Using iGPU for storage stuff like decompression would eat a big portion of potential performance gains, as you would be transferring unpacked data from RAM to VRAM instead of packed.
 
Using iGPU for storage stuff like decompression would eat a big portion of potential performance gains, as you would be transferring unpacked data from RAM to VRAM instead of packed.
While technically true, the DMA transfer from iGPU (which was leveraging main memory the whole time anyway) to dGPU will be limited by the PCI-E bus and however fast the iGPU can chew through the unpack algorithm against whatever data it was supplied. Said another way, leveraging an iGPU as a "decompression accelerator" is still going to end up faster than using the CPU to do the same.

If the dGPU is otherwise occupied, this would be a fantastic way to continue streaming compressed data directly into the dGPU's memory pool without ever really touching the pCPU.
 
In theory you should be able to as it's likely that MS will use compute for that purpose, at least initially, which should work on any GPU of X level of compute capability. Integrated graphics aren't disabled as, for example, you can drive a monitor with the integrated GPU on a desktop system while simultaneously driving a monitor on a dedicated GPU.

If it's performant enough at GPU based decompression, I see little to no reason why they wouldn't want to leverage the capabilities of the integrated GPU for such tasks, especially if it's for assets whose destination is main system memory versus the GPUs VRAM. In the latter case it might be preferable to have the dedicated GPU do it if it results in lower latencies and lower latencies is desirable. Although it could still be potentially more beneficial to leverage the integrated GPU reduce work done by the dedicated GPU WRT non rendering related work.

Regards,
SB

I think I did not explain myself well. I was thinking that, until DirectStorage supports GPU decompression, it could be possible to "cheat" and use the graphic part of an APU at the CPU only stage. If I am not wrong you just can use the CPU at "Custom Decompression Queue" stage.
 
Something for the SSD/NVME/DirectStorage topic as well @ https://www.eurogamer.net/digitalfo...-ssd-interview-pcie-50-directstorage-and-more

The Kingston SSD interview: PCIe 5.0, DirectStorage and more
Talking about next generation.

It's an exciting time in the PC hardware space. While we've been spending our time testing DDR5 memory and 12th-gen Intel processors, we're also on the cusp of some interesting storage innovations. Microsoft's DirectStorage API was released just a few weeks prior, and as PCIe 4.0 SSDs have become increasingly affordable, PCIe 5.0 has debuted on Intel motherboards, soon to be replicated with AMD's next generation platform. With that in mind, we had plenty of questions ready when we were approached by storage giant Kingston to speak to one of their longest-serving executives on the SSD side of the firm.

What follows is an interview with Tony Hollingsbee, EMEA Business Manager - SSD at Kingston. He's been at the company for two decades, and therefore had a hand in the transition from hard drives to SATA and then NVMe SSDs. We asked him about what the next big storage technologies would mean for gamers, whether Kingston would be up for creating SSDs for the Xbox, not just the PlayStation 5 and PC, and what kind of testing drive manufacturers do when preparing their latest products for the public - and reviewers - to test out. Enjoy!

...
 
This seems interesting.

Samsung Electronics Develops Second-Generation SmartSSD Computational Storage Drive with Upgraded Processing Functionality - Samsung

Samsung Electronics Co., Ltd., the world leader in advanced memory technology, today announced that it has successfully developed a second generation of its pioneering SmartSSD.

The new proprietary computational storage incorporates data processing functionality within a high-performance SSD. Unlike existing SSDs, Samsung’s SmartSSD can process data directly, thereby minimizing data transfers between the CPU, GPU and RAM. This technology can avoid the bottlenecks that often occur when moving data between storage devices and CPUs, resulting in markedly improved system performance and much higher energy efficiency.
 

Awesome find! Some really interesting data there!

  • Surprisingly soon launch. I wasn't expecting it this year.
  • Actual benchmark (kind of) result showing a massive speed up. Obviously result will vary by CPU (which they don't mention) and content (this was highly optimised).
  • We now know the compression format, and it's not BCPack! Welcome to GDeflate! Interesting that it is one specific format which seems to be mandated.
  • GDeflate was developed by.... Nvidia! This will have implications for the walled garden arguments that often follow Nvidia. This is obviously an open standard format.
  • It's unclear whether GDeflate will be free to devs. On the one hand they say the compressors will be included with the DirectStorage 1.1 SDK, but on the other hand they say they "will publish Apache 2.0 licensed reference implementations of GDeflate compressors and decompressors, allowing tooling to be integrated with existing asset pipelines." I assume then that it will not be free, which may limit uptake :(
 
7gb/s thats quite fast, not many games that are going to use that though, if any. Has its uses for many other things i can imagine.
 
  • It's unclear whether GDeflate will be free to devs. On the one hand they say the compressors will be included with the DirectStorage 1.1 SDK, but on the other hand they say they "will publish Apache 2.0 licensed reference implementations of GDeflate compressors and decompressors, allowing tooling to be integrated with existing asset pipelines." I assume then that it will not be free, which may limit uptake :(
Apache 2.0 is permissive open source license, you can't really get much more open or free.
 
Apache 2.0 is permissive open source license, you can't really get much more open or free.

Awesome news! That should help it become rapidly adopted as an industry standard in the PC space which should in turn drive the use of GPU decompression in PC games, because why not? It's seemingly all win.
 
Back
Top