What We Know About Nvidia’s Upcoming Blackwell GPUs and Their Power Surge

Background

Nvidia is preparing its next‑generation GeForce RTX 50 series, codenamed Blackwell, to compete with the best graphics cards on the market. Leaked information from a well‑known hardware source suggests that the memory interface of the Blackwell series will not differ dramatically from the Ada Lovelace line, retaining a 384‑bit memory bus for the flagship GB202 GPU.

Memory Interface and GDDR7

Blackwell may be Nvidia’s first GPU family to support GDDR7 memory. The first‑generation GDDR7 SGRAM IC is expected to deliver 32 GT/s per pin, which means a 384‑bit subsystem could provide roughly 1,536 GB/s of bandwidth, making a 512‑bit bus unnecessary. Micron has announced 16 Gb and 24 Gb GDDR7 chips slated for 2025, but it remains unclear whether the initial RTX 50 cards will use the 16 Gb or 24 Gb configuration.

Historically, Nvidia’s top consumer GPUs (AD102, GA102) have used a 384‑bit bus, which offers the best performance‑to‑cost ratio. Lower‑tier models use 320‑bit, 256‑bit, 192‑bit, or 128‑bit buses depending on the product segment.

According to the same leaker, the Blackwell family will likely omit a GB204 GPU, and the rumored GB205 may not directly inherit the AD104 design.

Power Consumption and New Accelerators

While Nvidia’s current 700 W H100 AI accelerator is already power‑hungry, Dell’s chief operating officer Jeff Clarke has indicated that the upcoming B100 GPU and a further B200 accelerator could reach 1,000 W TDP without requiring liquid cooling. Clarke appears to be referring to the GB200 “Superchip,” which combines Nvidia’s Grace CPU with a B100‑class GPU, similar to the GH200 architecture.

Based on Grace‑CPU power estimates, the GB200’s TDP could be around 1,300 W, roughly 30 % higher than its predecessor. Dell also plans to showcase advanced liquid‑cooling expertise as an alternative cooling solution for such high‑density chips.

Other Roadmap Items

Beyond GPUs, Nvidia’s roadmap calls for faster InfiniBand and Ethernet NICs by the end of the year, with port speeds up to 800 Gb/s. Analysts warn that even with a projected three‑fold increase in 2024 GPU shipments, supply constraints are likely to persist.

Overall, the Blackwell architecture is expected to bring higher memory bandwidth via GDDR7, significantly higher power envelopes, and tighter integration with Nvidia’s Grace CPU, shaping the next wave of AI‑focused accelerators.