Sign up to save your podcasts
Or
Share Louise ai agent : Contenders at the 15 exaFLOPS level
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
Louise ai agent : Contenders at the 15 exaFLOPS level
NVIDIA DGX GB200 NVL72: A rack-scale system with 72 Blackwell GPUs, delivering 1.44 exaFLOPS (FP4 sparse) or ~720 petaFLOPS (FP8 dense). This is NVIDIA’s latest flagship AI system as of March 2024 announcements, with potential upgrades by 2025 (e.g., Blackwell Ultra B300).- xAI’s Systems (Assumed XA1): xAI doesn’t publicly detail an "XA1," but it’s building AI supercomputers, likely leveraging NVIDIA hardware (e.g., DGX H100 or GH200 clusters). Rumors suggest xAI aims for exascale compute, possibly 1–4 exaFLOPS, based on its mission to accelerate human discovery.
- Cerebras CS-3 and Condor Galaxy: The CS-3 offers 125 petaFLOPS (FP16) per system, with Condor Galaxy 3 (64 CS-3s) at 8 exaFLOPS, and a planned 2048 CS-3 cluster hitting 256 exaFLOPS. Cerebras focuses on wafer-scale engines (WSE-3) with massive on-chip memory and bandwidth.
- Metrics: Compute (exaFLOPS), price (estimated system cost), and cost-efficiency (dollars per exaFLOP). Power consumption and scalability will also influence long-term value.
View all episodes
By David Nishimoto
NVIDIA DGX GB200 NVL72: A rack-scale system with 72 Blackwell GPUs, delivering 1.44 exaFLOPS (FP4 sparse) or ~720 petaFLOPS (FP8 dense). This is NVIDIA’s latest flagship AI system as of March 2024 announcements, with potential upgrades by 2025 (e.g., Blackwell Ultra B300).- xAI’s Systems (Assumed XA1): xAI doesn’t publicly detail an "XA1," but it’s building AI supercomputers, likely leveraging NVIDIA hardware (e.g., DGX H100 or GH200 clusters). Rumors suggest xAI aims for exascale compute, possibly 1–4 exaFLOPS, based on its mission to accelerate human discovery.
- Cerebras CS-3 and Condor Galaxy: The CS-3 offers 125 petaFLOPS (FP16) per system, with Condor Galaxy 3 (64 CS-3s) at 8 exaFLOPS, and a planned 2048 CS-3 cluster hitting 256 exaFLOPS. Cerebras focuses on wafer-scale engines (WSE-3) with massive on-chip memory and bandwidth.
- Metrics: Compute (exaFLOPS), price (estimated system cost), and cost-efficiency (dollars per exaFLOP). Power consumption and scalability will also influence long-term value.