Microsoft recently released another update for Direct Storage, bringing the version number up to a hardy 1.1. Big Blue claims that game loading times can be decreased by as much as 200%!
Wow… that’s a lot of precents! What does that mean for gamers, however? Let’s dig into this a bit.
What is Direct Storage?
Direct Storage was first introduced with the Xbox Series X and S. It’s a Direct X API that assists in transferring assets from a storage device to the GPU. In theory, it should allow games to load much faster.
The concept of Direct Storage was first introduced wide scale by Sony. Other than beefier hardware, Sony announced that the PS5 would include a direct link between the fast built-in NVME storage drive and the PlayStation’s GPU. Sony demoed the concept by demonstrating how the game Ratchet and Clank could instantly load new worlds while the game’s characters traveled between them.
Microsoft later announced that the Xbox Series X and S would have a similar feature called Direct Storage. Microsoft then updated the Direct X API and ported Direct Storage to Windows. Direct Storage was officially launched for PCs in March of 2022.
The update referenced above that increases load times by as much as 200% is part of the 1.1 update for Direct Storage. That update was released on November 8th, 2022.
How does Direct Storage work?
Direct Storage is an interesting beast. First, it’s important to know that because it’s a Direct X API, it’s easy for developers to integrate it into games. However, that doesn’t mean that Direct Storage will actually improve anything unless the game itself is optimized to utilize it properly.
Transferring assets from a storage device to a GPU has always been a bit of a headache for game devs. Traditionally, assets needed to be loaded in large chunks. This is/was due to how GPUs could access storage devices, or rather, how they couldn’t.
Before resizable bar was introduced, data needed to be transferred between computer components through its bus. That bus was limited, and only small chunks of data could be transferred at any time.
This is why large chunks of game data need to load before a game can be played. The GPU needs any data it may need to be transferred to its onboard memory first. Game engines can’t depend on assets being transferred on-the-fly during gameplay.
Resizable Bar changed how much data could be transferred at once. This was step 1.
Direct Storage builds on top of those concepts. First, game developers use a special compression method to compress the living poop out of game assets. Games are shipped with their assets pre-compressed. When a game loads, those compressed assets are decompressed by the CPU and then transferred directly to the GPU memory. This decreases load times.
Needing to pass assets through the CPU is a bottleneck, however. As it turns out, graphics cards are great at processing decompression algorithms. Direct Storage 1.1 swaps out the CPU in favor of the GPU. So, assets are transferred directly from storage to the GPU, and the graphics card does the heavy decompression lifting.
So, what does this mean for gamers?
The sad reality is that Direct Storage doesn’t mean anything for gamers yet, and it won’t for a few years. As many have pointed out online, game development takes years. While Direct Storage can be added to games after they are released, or in the middle of development, those games won’t truly take advantage of it. Load times may be decreased, but it won’t be a good representation of what Direct Storage is capable of.
However, that doesn’t mean we shouldn’t be excited about Direct Storage.
By the time game engines and game studios can fully integrate Direct Storage into their engines and learn how to utilize it properly, graphics cards will be beasts. While we can snark at how much power an Nvidia 4090 uses, the simple fact is that both Nvidia and AMD have produced some amazingly powerful GPUS for the Lovelace and RDNA3 generation.
That has a couple of interesting implications.
Intelligence Included
It’s not a secret that some game developers use neural networks to help build games. Machine learning can be helpful for things like programming enemy AI. However, machine learning is limited in scope for how useful it can be for creating games currently.
In the future, once GPUs are more powerful and can access data much faster, it may be possible to build more intelligence into games. GPUs are great at running machine learning algorithms and processing AI models. If they become a bit more powerful, they’ll be beefy enough to process both light machine-learning models and beautiful graphics at the same time. Arena shooters might become way more interesting when bots can emulate humans, learning player habits and reacting to them (in the moment) in new and novel ways.
Memory Constraints
Direct Storage becomes much more interesting when referring to GPU memory, however. The amount of onboard memory included with GPUs has always been a point of contention. Adding more RAM makes graphics cards more expensive (both in hardware and engineering costs). Adding too little RAM poses performance and fidelity issues.
The GeForce Nvidia 1050 Ti was an excellent example of a lower-end graphics card limited by GPU memory. The GeForce 1050 Ti wasn’t a graphical powerhouse, but it had enough hardware to process 1080p graphics well enough.
The problem is that as the 1050 Ti aged, it couldn’t hold the higher-resolution texture packs and assets in memory. It only had 4GB of GPU memory. Games like the Resident Evil 2 Remake could quickly flood the GPU memory with assets. RE2 fans using a 1050 Ti were forced to lower graphics settings even though the GPU technically had enough processing power to play the game at high settings at a good-enough framerate.
When Direct Storage is fully realized, this may become a thing of the past. Remember that graphics cards only need as much memory as they do because that memory needs to be filled with the game’s assets before it is played. If game engines can swap assets in near-real-time (a la’ Rachet and Clank: Rift Apart), GPU memory suddenly becomes much less critical.
NO! YOU’RE WRONG!
“But wait… The Radeon 7900XTX includes 24GBs of RAM. That kills your entire theory!”
But… does it? I previously wrote that the GeForce 4090 and Radeon 7900XTX aren’t meant for gamers. Both GPUs are SMB (small to medium business) cards in gamers’ clothing. That 24GB of memory is intended for other things – like machine learning and content creation.
In the future, flagship GPUs will still have tons of memory. Things like ML and 3D modeling will always need tons of GPU memory. That need grows as AI becomes more advanced and 3D models are rendered at higher resolutions with more polygons.
However, it’s entirely possible that graphics cards truly meant for gamers will level off at 16GB of memory or less. The only reason we need 16GB of memory now is to hold 4K-targeted assets.
Game engines don’t take up much memory in comparison. They’ve primarily leveled off in size for packaged games, and engines are not growing exponentially like assets targeting higher resolutions.
Frankly, if this prediction comes true, I’d rather have less RAM that’s blazingly fast than more RAM that’s slower for a gaming-specific graphics card. A faster connection between GPU cores and GPU memory will improve FPS.
Be the first to comment