What Nvidia’s RTX 50 Series and Blackwell Architecture Mean for GPUs and Data Centers

Nvidia unveiled its RTX 50 series built on the new Blackwell architecture at CES, introducing four consumer GPUs – RTX 5090, RTX 5080, RTX 5070 Ti and RTX 5070. The flagship RTX 5090 packs 92 billion transistors, delivers 3 352 TOPS of AI compute, and is priced at $1 999, offering roughly double the performance of the RTX 4090 thanks to Blackwell and DLSS 4. The RTX 5070, priced at $549, claims performance comparable to the RTX 4090 ($1 599) despite a much lower cost.

For data‑center workloads, Nvidia showcased the Grace Blackwell NVLink72 super‑chip, which can combine 72 or 144 Blackwell GPUs to achieve 1.4 ExaFLOPS, house 130 trillion transistors, and deliver four‑fold compute gains that would surpass the world’s fastest supercomputers.

On the personal‑AI front, Nvidia announced Project DIGITS, a pocket‑sized AI supercomputer slated for May release. It uses the GB10 Grace Blackwell chip, operates at FP4 precision, and can reach 1 PFLOPS while running on a standard power outlet.

Leaked specifications suggest the RTX 5090 will feature 32 GB of GDDR7 memory on a 512‑bit bus (≈1 792 GB/s bandwidth), a 3.5‑slot thickness, and a TDP of 575 W. Rumors also cite 21 760 CUDA cores and a price range of $2 000–$2 500. The RTX 5080 is expected to launch first (around Jan 21), equipped with 16 GB GDDR7, PCIe 5.0, a GB203‑400 Blackwell GPU, and 10 752 CUDA cores.

Reference links: https://videocardz.com/newz/exclusive-first-look-at-geforce-rtx-5090-with-32gb-gddr7-memory ; https://www.techradar.com/computing/gpu/nvidia-geforce-rtx-5080-appears-to-leak-ahead-of-possible-ces-2025-announcement .