[xg15]

> Recent additions where, add GPU possibility to read data from system disks, using mentioned before bus master, but those additions are not already wide implemented.

My impression is that high-end graphics cards (Nvidia RTX 30x and professional equivalents) more and more replicate parts of the PC architecture and become sort of mini-computers within a computer.

Following that logic, I wonder when we'll see the first card with its own dedicated flash memory - or why not a PCIe controller, so you can hook op an SSD...

[simne]

Some years amd produced enterprise gpu accelerator with ssd interface (or nvme, I don't find too many details). For Nvidia cannot remember. Currently this addition doesn't become commodity, but this is possible in nearest future.

[xg15]

I mean, for gaming cards, it would even make some sense.

When the user installs a new game, the installer could just pass on all assets of the game right to the GPU. The GPU could precompile the shaders and also store textures, models etc in whatever way would be optimal for that particular card.

During play, loading should be considerably faster as the GPU could just load the data from storage into graphics memory, without touching the main system at all.

A funny side-effect could be that the game install would also probably appear significantly smaller as it would take up less space on the filesystem.

Finally, Intellectual Property people would probably be happy as well, as the precious game assets wouldn't touch any user-accessible storage at all...

[simne]

> During play, loading should be considerably faster as the GPU could just load the data from storage into graphics memory, without touching the main system at all.

This could not affect performance much, because pcie speed far surpass speed of best ssd/nvme.

Currently only problem, that for all transactions ssd<->gpu, involved main cpu, so it have much load only for accept transactions. But this is already solved by new technology, which appear recently and allows gpu to access ssd directly, unfortunately not much details known.

As for size I'm not sure, because gpu hardware is too special for deal with filesystems, so I think, there will be special fs. Also possible that this part of disk system will be read only for gpu (all writes and maintenance will do host cpu), to simplify gpu software part.

[xg15]

> This could not affect performance much, because pcie speed far surpass speed of best ssd/nvme.

Ah, I wasn't aware of that. Thanks for the info.

I would think there could still be some speed advantage if the GPU does not have to share storage with CPU, coordinate access, etc.

> As for size I'm not sure, because gpu hardware is too special for deal with filesystems, so I think, there will be special fs. Also possible that this part of disk system will be read only for gpu (all writes and maintenance will do host cpu), to simplify gpu software part.

That was sort of what I meant. In that scenario, a game would take up less space in the "ordinary" filesystem, i.e. storage of the main system - because part of the game data would be moved to GPU storage instead. Hence appear to be smaller: If you added the space taken up in main storage and GPU storage back together, you'd arrive back at the old size. It's more a sleight of hand, not an actual reduction in size.

[simne]

> I would think there could still be some speed advantage if the GPU does not have to share storage with CPU

In enterprise advantage really big - because they typically use RAID disk arrays and enterprise systems usually have lot of disks, 5-10 disks normal, and exists machines with 64 disks or even more than 100. And also important thing, server hardware usually use much lower bus frequencies than gaming PC, so they just don't have opportunity to see max speed of bus.

That's why amd first created such card for enterprise - speed is really pain for enterprise computing.

> a game would take up less space in the "ordinary" filesystem, i.e. storage of the main system - because part of the game data would be moved to GPU storage instead

I mean, because of need less gpu load for access fs, there should be used simplified filesystems, or huge optimized for speed of access, and sucj FSs will be much less size effective than ordinary.

For example, read Knuth about search and sorting - you will know, that search/sort by using index is much faster, many magnitudes faster, but index size may be magnitudes larger than data.

[simne]

btw https://arstechnica.com/gadgets/2022/03/nvidia-wants-to-spee...