Breaking

Blackwell Posts 20x Power Efficiency Over Hopper on New AgentPerf Benchmark

June 15, 2026 at 11:27 EDT

Infra & Chips
AI Agents
Research & Papers

NVIDIA says its latest Blackwell-based systems achieved up to 20x the "agents per megawatt" (the number of concurrent agents running per megawatt of power) over the prior-generation Hopper on AgentPerf, the first agent-focused infrastructure benchmark from Artificial Analysis.

June 1, 2026 · NVIDIA — GTC Taipei

NVIDIA RTX Spark: a superchip to make the Windows PC an AI-agent machine

Built with Microsoft, RTX Spark fuses a Blackwell RTX GPU with a 20-core Grace CPU on one chip — pitched to run autonomous personal AI agents locally, around the clock, no cloud required.

1 PF

Up to 1 petaflop (FP4) of AI performance

128GB

Unified LPDDR5X memory, shared CPU + GPU

70B

~70 billion transistors on TSMC 3nm

One chip, the whole stack

A single piece of silicon stacks every layer NVIDIA wants to own — from CPU cores to up to 1 PF of AI compute.

20-core CPU

Grace · Arm

6,144 cores

Blackwell RTX GPU

1 petaflop

FP4 Tensor Cores

CPU and GPU share one 128GB memory pool over NVLink-C2C — up to 273 GB/s bandwidth.

SHIPPING FALL 2026

Thin 14mm, ~3-pound Windows-on-Arm laptops (14–16") and compact desktops from Dell, Microsoft Surface and others. Price undisclosed. Microsoft aims to evolve Windows into an "agentic AI OS."

THE PROMISE

Praised power efficiency on DGX Spark siblings
Local fine-tuning of 70B-class models via unified memory
Native full CUDA/RTX stack: DLSS, Reflex, G-SYNC, ray tracing

THE CAUTION

No independent benchmarks until shipping
Windows-on-Arm app compatibility unproven
Value vs likely multi-thousand-dollar price still open

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.