GeForce RTX 3080 Ti: Polnischer Händler listet neues MSI-OC-Modell
Die Hinweise darauf, dass MSI ein übertaktetes Modell der GeForce RTX 3080 Ti auf den Markt bringen wird, verdichten sich. Ein polnischer Händler hat die neue Grafikkarte in seinem Shop gelistet.
Nachdem MSI kürzlich offenbar erste Exemplare einer bislang unveröffentlichten übertakteten GeForce RTX 3080 Ti in die USA geliefert hat, nahm jetzt ein polnischer Händler die MSI GeForce RTX 3080 Ti Ventus 3X OC in seinen eigenen Shop auf. Der Anbieter mit dem Namen X-Kom nimmt bislang aber noch keine Vorbestellungen auf und gibt auch keinen Preis an. Allerdings können Kunden sich über die Verfügbarkeit der Karte informieren lassen.
MSI GeForce RTX 3080 Ti OC mit 12GB Speicher
Dass der Händler die neue und bislang noch nicht offiziell vorgestellte Grafikkarte in den eigenen Shop aufgenommen hat, bestätigt die Gerüchte, dass der Launch der GeForce RTX 3080 Ti in Kürze bevorsteht. Wie Videocardz berichtet, soll der der Betreiber von X-Kom direkte Verbindungen zu den Herstellern besitzen und Informationen zu neuer Hardware im Voraus erhalten.
Die Listung im Shop von X-Kom lässt bislang kaum Rückschlüsse auf die Spezifikationen der übertakteten Custom-GPU von MSI zu, nur so viel: Die Karte wird voraussichtlich mit einem 2.5-Slot und drei Lüftern sowie HDMI- und Display-Ports ausgestattet sein. Außerdem nutzt sie laut Angaben der Redaktion von Wccftech einen Dual-8-Pin-Stromanschluss. Die GeForce 3080 Ti selbst soll über 12GB GDDR6X-Speicher bei einer Bandbreite von 912GB/s verfügen. Weitere Details zu der neuen Nvidia-Grafikkarte lesen Sie in unserer ausführlichen News zur Leistung und Verfügbarkeit der Hardware.
Angeblich soll der Verkauf der Nvidia GeForce RTX 3080 Ti am 26. Mai 2021 beginnen. Custom-Modelle könnten kurz darauf in den Handel kommen beziehungsweise zur Vorbestellung verfügbar sein.
Quelle: Wccftech / Videocardz
https://lmgtfy.app/?q=sam...
Jo, genau, ich verstehe nicht was "Memory Multiplier techniques" bedeuten soll. Auch "better memory efficiency" und "makes it look like your GPU has x times more memory than what's physically available" kann ich nicht übersetzen. Bin zu dumm...
Ich glaube eher Du verstehst nicht die Zusammenhänge. Hier noch mal eine sehr ausführliche Erklärung:
How this works
The Xbox Series X has a moderately fast SSD, with read speeds of 2.4GB/s, and decompression hardware that should effectively double that speed for textures using a custom algorithm they call BCPack. New techniques are needed to use this bandwidth effectively.
Sampler Feedback Streaming uses the results of the previous rendered frame to determine which textures to read into memory. The idea is that most textures are only kept in memory at a lower detail, and as the game needs higher resolution textures, those are loaded in for the next frame.
Each texture is made of mipmaps, which is a chain of progressively smaller versions of the same texture. Developers should also use tiled resources, which splits large textures into small squares, each of which can be loaded individually.
A 1024x1024 texture might be made up of 64 tiles of size 128x128. Its mipmap would include a 512x512 texture made of 16 tiles of size 128x128, and an even smaller 256x256 texture made of 4 tiles, and even smaller textures for even further away views, each no larger than a tile.
Each tile can either be loaded into memory, or simply reserved without using up memory. When a GPU renders a scene, it looks at specific tiles on specific mipmaps. The Sampler Feedback extension means that every time the GPU loads a tile, it can mark in a buffer saying that the resource was needed, and at what level of detail. If the tile it needs is not loaded into memory, the GPU will fall back to a lower-resolution version of the texture, but it will still mark in the buffer that it wanted that mipmap.
This means that after rendering the CPU has a bunch of information on exactly what tiles are needed. It can quickly compare this data to the tiles it has loaded, and will end up with a list of tiles that the GPU wanted, but didn't have. The CPU can also keep track of which tiles have not been requested recently, since those can probably be discarded.
What this means
Two things determine how fast an SSD can read: the size of the request (how much data you want), and the queue depth (how many things you are asking for in parallel). For an SSD, 4kiB reads at queue depths of 1 or 2 is a typical ‘random’ workload, and is generally fairly slow. In contrast, 128kiB reads at queue depths of 16 should utilize most of the speed of most SSDs.
With 128x128 tiles, and 4 bytes of data per texel, a tile is 64kiB. Multiple textures, like the color and normal, will always be needed at the same time, and so can be loaded and unloaded together. After compression, each request might be 64kiB or more. Because the CPU knows all the textures it needs at the same time, it can queue all the reads as fast as it wants to. This means that you should be able to use close to the full speed of the SSD for texture streaming.
If the speed of texture loading, after decompression, is around 3.8 GiB/s, that would allow loading 64 MiB per frame at 60fps. This is fast enough to load 1000 tiles each frame, about the equivalent of a single 4096x4096 texture.
For objects approaching from afar, textures will be loaded in when the GPU changes which mipmap it is rendering from, with only a frame or two's delay, so loading should be effectively invisible. If you look at an uncached object up close for the first time, it might render using lower resolution textures for a frame or two. How low that is depends on how much memory the developer wanted to save.
I believe this API requires specific GPU hardware to support, but it is not exclusive to consoles.