Nvidia Reveals Next-Gen AI Chips, Roadmap Through 2028 (cnbc.com)
(Tuesday March 18, 2025 @06:40PM (msmash)
from the pushing-the-limits dept.)
- Reference: 0176764877
- News link: https://tech.slashdot.org/story/25/03/18/201213/nvidia-reveals-next-gen-ai-chips-roadmap-through-2028
- Source link: https://www.cnbc.com/2025/03/18/nvidia-announces-blackwell-ultra-and-vera-rubin-ai-chips-.html
Nvidia unveiled its next wave of AI processors at GTC on Tuesday, [1]announcing Blackwell Ultra chips that will ship in the second half of 2025, followed by the Vera Rubin architecture in 2026. CEO Jensen Huang also revealed that its 2028 chips will be named after physicist Richard Feynman.
The Blackwell Ultra maintains the same 20 petaflops of AI performance as standard Blackwell chips but increases memory from 192GB to 288GB of HBM3e. Nvidia claims these chips can process 1,000 tokens per second -- ten times faster than its 2022 hardware -- enabling AI reasoning tasks like running DeepSeek-R1 models with 10-second response times versus 1.5 minutes on H100 chips.
Vera Rubin will deliver a substantial leap to 50 petaflops in 2026, featuring Nvidia's first custom Arm-based CPU design called Olympus. Nvidia is also changing how it counts GPUs -- Rubin itself contains two dies working as one chip. The annual release cadence represents a strategic shift for Nvidia, which previously introduced new architectures every two years before the AI boom transformed its business.
[1] https://www.cnbc.com/2025/03/18/nvidia-announces-blackwell-ultra-and-vera-rubin-ai-chips-.html
The Blackwell Ultra maintains the same 20 petaflops of AI performance as standard Blackwell chips but increases memory from 192GB to 288GB of HBM3e. Nvidia claims these chips can process 1,000 tokens per second -- ten times faster than its 2022 hardware -- enabling AI reasoning tasks like running DeepSeek-R1 models with 10-second response times versus 1.5 minutes on H100 chips.
Vera Rubin will deliver a substantial leap to 50 petaflops in 2026, featuring Nvidia's first custom Arm-based CPU design called Olympus. Nvidia is also changing how it counts GPUs -- Rubin itself contains two dies working as one chip. The annual release cadence represents a strategic shift for Nvidia, which previously introduced new architectures every two years before the AI boom transformed its business.
[1] https://www.cnbc.com/2025/03/18/nvidia-announces-blackwell-ultra-and-vera-rubin-ai-chips-.html
PC gaming is dying (Score:2)
by rsilvergun ( 571051 )
We just can't get graphics cards. Intel put out a pretty nice chip that's supposed to retail for 250 bucks but the cheapest you're going to find it unless you get lucky or you live near a microcenter is 350. I've heard tails of people who paid MSRP for one but I'm pretty convinced you're more likely to see Bigfoot than a MSRP GPU right now.
I suspect years from now if we ever have antitrust law enforcement again here in America then it'll get revealed that there was collusion going on. Like when lcd mon
marketing perf again? (Score:2)
> Nvidia claims these chips can process 1,000 tokens per second -- ten times faster than its 2022 hardware
So if we go by Nvidia's bullshit performance metrics from their 50 series graphics cards we can expect around about a 10% perf boost?