According to Igor’s Lab (a well-reputed outlet), NVIDIA’s RTX 3080 will come 10GB of GDDR6X memory and a power draw of 320W. This is a bit surprising considering that Ampere is supposed to be fabbed on TSMC’s 7nm process. We’re basically looking at an increase of nearly 100W over the vanilla RTX 2080 despite a 2x increase in transistor density. This falls in line with what our trusty Twitter sources have been claiming over the past few months. Igor doesn’t mention the core counts, but the memory width is included. Expect 4,352 cores paired with 10GB of GDDR6X memory over a 320-bit bus. Now, that’s a massive upgrade!
Part | PCB | Chip | Model | Extension | Memory | Interface | TBP | Connectors |
SKU10 | PG132 | GA102 | RTX 3090 | (Ti/Super)* | 24 GB GDDR6X (Double-Sided) | 384-bt | 350 W | 3x DP, HDMI NVLink |
SKU20 | PG132 | GA102 | RTX 3080 | (Ti/Super)* | 11 GB GDDR6X* | 352-bit* | 320 W | 3x DP, HDMI |
SKU30 | PG132 | GA102 | RTX 3080 | none | 10 GB GDDR6X | 320-bit | 320 W | 3x DP, HDMI |
The rumored RTX 3080 Ti/Super will come with slightly more memory (11GB GDDR6X) and a wider bus (352-bit), in addition to a higher core count (~5,000). Finally, the top-end RTX Titan or 3090 will come with twice as (24GB GDDR6X) and an even wider 384-bit bus. The core count will range in the 5,000-6,000 range, not quite as much as the A100’s 6,912.
The surprising part is that this time all three top-tier GPUs, the RTX 3080, the 3080 Ti as well as the Titan/3090 are based on the same GA102 die. Traditionally, the x80 and x70 the Gx102 die with the latter utilizing a cut-down variant and the former the full-fat variant. The Titan has been based on the xA102 die and the Ti uses a cut-down variant of the same. If this leak is indeed correct, then this marks a major shift in NVIDIA’s die policy regarding different consumer products.