A report shared by memory manufacturer, Micron has confirmed that NVIDIA’s Ampere flagship, the GeForce RTX 3090 will use 24GB of GDDR6X memory running at a whopping 21Gbps. The document discusses the need for high-speed, relatively low-cost memory solutions demanded by the GPU market, both in consumer and Data Centern segments:

The RTX 3090, like the Titan RTX is categorized as an AI Inference Accelerator despite the fact that both GPUs are not compute cards, rather actual graphics cards with graphics pipelines. The RTX 3090, AKA Ampere Titan will feature 12 stacks of GDDR6X memory connected via a 384-bit bus, running at 21Gbps, resulting in a massive bandwidth of around 1000GB/s.
This is exactly as I had predicted a week earlier. You may read that post here:
Here are the expected specifications for NVIDIA’s RTX 30 series GPUs:
GPU | Die | Cores | Memory | Memory Speed | Bus Width |
---|---|---|---|---|---|
RTX Titan | GA-102-400 | 5376 | 24GB GDDR6X | 21Gbps | 384-bit |
RTX 3080 Ti | GA102-300 | 5,248 | 20GB GDDR6X | 19Gbps? | 352-bit |
RTX 3080 | GA102-200 | 4,352 | 10GB GDDR6X | 18Gbps? | 320-bit |
RTX 3070 | GA104? | 2,944 | 8GB GDDR6 | 16Gbps? | 256-bit |
RTX 3070 Ti | GA104? | 3,072 | 8GB GDDR6X | 16Gbps? | 256-bit |