The first-gen GDDR7-based graphics cards will use 16Gb dies, 2GB minimum, and 32Gbps speeds

The first generation of GPUs based on GDDR7 memory will use 16Gb dies and 2GB VRAM capacities, at 32Gbps bandwidth, says hardware leaker.

The first-gen GDDR7-based graphics cards will use 16Gb dies, 2GB minimum, and 32Gbps speeds
Comment IconFacebook IconX IconReddit Icon
Gaming Editor
Published
Updated
3-minute read time

NVIDIA will be the first to use next-gen GDDR7 memory on its next-gen GeForce RTX 50 series "Blackwell" GPUs, where we could see the first-generation GPUs using GDDR7 featuring 16Gb GDDR7 modules, which means the minimum VRAM capacity for next-gen graphics card will begin at 2GB.

The first-gen GDDR7-based graphics cards will use 16Gb dies, 2GB minimum, and 32Gbps speeds 604
2

GDDR6 started off with 8Gb dies and 1GB capacity per module, but now, hardware leaker "kopite7kimi" has posted on X, saying, "We will use 16Gbit first." I don't think we're going to get some beast GeForce RTX 5090 with double the VRAM -- so, 48GB of GDDR7, at first -- but it means that we'll get similar, or the same VRAM capacity but at wicked bandwidth levels compared to current-gen cards.

For example, lower-end graphics cards using 8GB of 32Gbps GDDR7 memory on a 192-bit memory bus will feature 512GB/sec of memory bandwidth. Meanwhile, a 192-bit memory bus ramps up to 768GB/sec, and a mid-range 256-bit memory bus will feature 1TB/sec of memory bandwidth (that's as much as the RTX 4090).

  • 128-bit @ 32 Gbps: 512 GB/s (8 GB)
  • 192-bit @ 32 Gbps: 768 GB/s (12 GB)
  • 256-bit @ 32 Gbps: 1024 GB/s (16 GB)
  • 320-bit @ 32 Gbps: 1280 GB/s (20 GB)
  • 384-bit @ 32 Gbps: 1536 GB/s (24 GB)
  • 512-bit @ 32 Gbps: 2048 GB/s (32 GB)

But things start getting juicy at 384-bit memory bus levels with 1.5TB/sec of memory bandwidth (and 24GB of GDDR7; this would be a great configuration of GDDR7 memory on the RTX 5090). NVIDIA or AMD can utilize the 512-bit memory bus with 32GB of GDDR7 memory on board, rocking a huge 2TB/sec of memory bandwidth (double, yeah, double the memory bandwidth of the RTX 4090).

I wrote back in early February that NVIDIA is reportedly the only one buying GDDR7 memory from the likes of Micron, Samsung, and SK hynix... AMD isn't buying GDDR7, and has been buying oodles of faster GDDR6 memory modules at 20Gbps for its next-gen RDNA 4-based graphics cards.

NVIDIA is expected to roll out its new Blackwell GPU architecture at its GPU Technology Conference on March 18, where we'll see the new B100 AI GPU... but Blackwell-based GeForce RTX 50 series GPUs are expected later in the year. They will be all the rage in 2025 and beyond, all with GDDR7 memory.