One of the performance issues that professionals who use their graphics card for content creation on a daily basis face is the limitation of VRAM, suddenly if the data is too big, it goes to system SSD or other. It’s worse, than a conventional hard drive.
The solution? Place NVMe memory inside the graphics card, either through an M.2 drive slot or with NVMe memory soldered to the card.
NVMe memory to increase the amount of VRAM
One of the peculiarities of graphics cards used in professional environments is that they generally use VRAM in a so-called “clamshell” configuration which is typical of DDR memory, in which two GDDR memory chips are connected to a common memory interface. But since the connection is with two chips instead of just one as usual, what is done is split the data bus between the two, so that each memory chip has half the bandwidth that ‘only one, but for each memory controller interface
But with the advent of the DirectStorage API and support for SSD memory in GPUs, it is now possible to give NVMe memory access to a GPU, which helps to create an additional level in the hierarchy. GPU, in which to store data you can use GPU to access it faster.
For example we can have a scene to render which occupies several tens of gigabytes and which we generally need to access the hard disk, thanks to the access speed of NVMe it is possible to access this information very quickly. But wouldn’t it be better if the GPU had such a memory?
There is an SSD on my graphics card
A few years ago, AMD presented the Radeon SSG, this card was an AMD Vega with a peculiarity and it was that it had up to 2 TB of NVMe storage in the form of 4 M.2 type SSDs at the interior. This concept did not have much success on the market and the Radeon SSG remained an anecdote.
But this idea seems to have ended up appealing to Microsoft, which decided to standardize it in the DirectStorage API and asked AMD to integrate it into the Xbox Series X SoC. Those of Lisa Su, however, did not hesitate to bringing the technology to SONY and the result is a new generation of consoles not only based on SSD but in which “VRAM” has a greater capacity.
The concept is very simple, the GPU sees both the VRAM and the memory of the SSD as if it were indeed the same memory, but via an advanced memory controller, when a request is made to a corresponding address to the SSD to which it is copied a part of the VRAM itself which serves as a data cache for the SSD, therefore thanks to the high speed of the PCI Express interface, the data is quickly transferred from the NVMe memory to the RAM completely transparent.
How does the graphics access the SSD on the card?
Each GPU has two DMA units which allow you to access not only the system RAM, but the latest GPUs launched in the market and compatible with DirectStorage, what has been done is to give access to the system SSD by giving access at PCI Express Ports.
But installing an SSD in the GPU is different, since this memory would be before the PCI Express port, for that it would be necessary to add 4 PCI Express lanes to place an SSD, which AMD did with the Radeon SSG, but in the case from the AMD graphics card, they added another 16-lane PCIe interface, resulting in 4 M.2 drives.
While it may seem silly to put one or more SSD interfaces on a graphics card or, failing that, to solder an SSD with NVMe memory, in reality it is not at all if we take into account the enormous advantages that this implies, especially for the GPU when searching for large volumes of data.
Key to video editing and professional ray tracing
One of the utilities of GPUs is for encoding and decoding video in real time, although there are units specializing in this area, when it comes to transferring video from one format to another. , especially if we are talking about a large format. size, we can see that the information ends up not going into VRAM and is dumped to a hard drive.
Having an SSD allows content to be dumped into the SSD once it disappears from VRAM, imagine editing a huge video of tens or hundreds of Gigabytes on your PC and finding that the graphics have a direct access. This is extremely useful for what is called video encoding, which involves converting a video file from one format and resolution to another format and resolution in real time to be transmitted via streaming. .
The fact that the video CODEC located on the GPU has access to an SSD located on the graphics card means to speed up the performance, since the search for data is much shorter than the search for an SSD installed in the system. This means being able to access streaming video data as quickly as possible, ideal for video editing at 4K resolutions.
Another key utility will be professional ray tracing, with the aim of completely storing the entire spatial data structure of an extremely complex scene, this will help speed up the rendering process and prevent systems from spending a long time generating a scene. or a succession of them.
It can be concluded that adding NVMe or SSD drives in a graphics card is something that greatly contributes to the performance of a GPU in professional environments and that it will be a spearhead in the hardware market for them. professional content creators.