The team designed a custom CPU from logic gates, adapted DOOM source code, and ran it on an FPGA at 0.7 FPS initially.
Major challenges were memory size (DDR3 vs BRAM) and speed; they implemented two caches (ICache and DCache) with a memory arbiter to handle DDR3 latency.
Porting DOOM revealed CPU bugs such as infinite loops, incorrect immediate calculations, and cache overwrite issues; debugging required isolating parts of the program.
Hardware issues included unstable DDR3 interface (solved with Clock Domain Crossings and FIFOs) and keyboard input problems (used UART forwarding from a laptop).
Optimizations like raising clock speed, fixing fetch stage inefficiency, adding multiplication instructions, pipelining cache reads, and enabling compiler optimization raised FPS from 0.7 to 15-20.
The team plans to implement out-of-order processing and better memory access to reach 30 FPS, and possibly port Quake 2 after building a basic GPU.
Max single-threaded CPUs at scale are designed for the agentic AI era, focusing on high per-core performance to speed up AI agent tasks.
Traditional data center CPUs prioritize core count and cost efficiency over single-threaded performance, which can slow down agent loops where each step depends on the previous result.
NVIDIA Vera exemplifies this new CPU category with its Olympus core, offering 50% higher instructions per cycle than NVIDIA Grace, high memory bandwidth, and predictable latency.
Vera enables faster agent loops, with 1.8x sustained per-core performance over x86 in loaded workloads, improving tasks like tool calls, code execution, and data processing.
Real-world testing by Perplexity shows Vera completes coding workflows 1.5x faster and starts concurrent sandboxes up to 1.9x faster than x86 CPUs.
Vera also accelerates CPU-side data workloads, with partners reporting 3x faster SQL analytics and up to 6x lower latency in real-time streaming compared to x86 servers.
The CPU supports diverse agent workloads, including tool execution, data processing, and reinforcement learning, all on a single architecture, integrating with NVIDIA GPUs and storage processors.
NVIDIA plans future advancements with the Rosa CPU and Rigel core, aiming for even higher per-core performance while maintaining efficiency for the agentic AI era.