GDDR7

GDDR7 uses PAM3 signaling to push per-pin rates to 28–32 Gbps in first-gen products (2025), with a path to 40+ Gbps. RTX 50 series adopts it: RTX 5090 hits 1.79 TB/s, RTX 5080 at 960 GB/s, RTX 5070 at 672 GB/s.

For local inference, GDDR7 is the largest single-generation bandwidth jump consumer cards have seen — 78% over the RTX 4090. That translates almost linearly into decode tok/s on memory-bandwidth-bound workloads.

Still not HBM territory (an H100 SXM does 3.35 TB/s) but closes the gap meaningfully for prosumer setups.

An operator upgrading from an RTX 4090 (1008 GB/s) to an RTX 5090 (1.79 TB/s) is deciding whether the price jump is worth it for decode-bound local inference. Since decode throughput scales with memory bandwidth, the RTX 5090's GDDR7 interface, running PAM3 signaling, delivers roughly 78% more bandwidth than the 4090 — the largest single-generation jump in a consumer GPU to date — which translates to close to a proportional tok/s gain on the same quantized 7B or 13B model. For someone running larger models that don't fit on either card, the calculus shifts toward VRAM capacity over bandwidth; but for anyone bandwidth-limited on models that already fit, GDDR7 is currently the single best lever available outside of full-on HBM data-center hardware.