Nvidia's Vera CPU: 88 Custom Olympus Cores, 176 Threads
TL;DR
- Vera's Olympus core targets IPC over core count, inverting how server CPU vendors have competed on parallelism for the past decade.
- Vera's monolithic die with 164 MB L3 and 3.4 TB/s coherency fabric avoids chiplet topology fragmentation that AMD EPYC must manage at scale.
- NVIDIA Spatial Multithreading partitions core resources per thread, suppressing noisy-neighbor latency degradation that undermines SLA guarantees in multi-tenant AI factories.
Nvidia's next server CPU finally has architectural meat on the bone, and the interesting part is not the core count. Vera pairs 88 custom Olympus cores with 176 threads (via what Nvidia calls spatial multithreading) and a monolithic die feeding up to 1.5TB of LPDDR5X at 1.2TB/s of memory bandwidth. That is a real step from Grace's 72 threads and 480GB memory ceiling, but the bet is on IPC and memory patterns, not raw throughput.
The Olympus core, as Ryan Smith walks through at ServeTheHome, is deliberately wide: 18 total execution pipes, an instruction fetch unit that can feed as many as 16 instructions per cycle, a decode queue that outputs up to 10 fused instructions, and a neural branch predictor resolving two branches per cycle. Nvidia is also leaning on value prediction and a graph prefetcher tuned for pointer-heavy access patterns, both explicitly aimed at the kind of graph traversal that shows up in agentic AI workloads.
The published SPEC CPU 2026 numbers are more modest than the architectural pitch suggests. Vera posts a SPECrate2026_int_base of 925 against 898 for AMD's EPYC 9755, only a 3% multi-threaded lead. Nvidia's single-thread sweep across the integer suite is more decisive, and the company's own claims include 1.9x to 2.4x instruction-fetch gains and 29.3x scaling on a graph traversal benchmark at 32 cores versus 10x for AMD's system. Take those as vendor-supplied, not independent.
The honest caveat is that Nvidia picked the EPYC 9755 as its comparison, which Smith flags as not AMD's strongest single-thread part, and no floating-point SPEC data has been published yet. The bigger issue is timing. Systems are expected later in 2026, which leaves Nvidia a narrow window before AMD's Venice (Zen 6, up to 256 cores per socket) starts shipping and reframes the throughput conversation. For hyperscalers picking a CPU to sit beside the accelerators, Vera's real pitch is the 1.8TB/s NVLink-C2C into Nvidia's own GPU stack, not raw SPEC leadership.
What others are reporting
-
NVIDIA Newsroom Read →
Official launch with ecosystem breadth: names 30-plus manufacturing and hyperscaler partners, includes Redpanda 5.5x latency reduction as a real-world deployment benchmark.
The CPU is no longer simply supporting the model; it's driving it. With breakthrough performance and energy efficiency, Vera unlocks AI systems that think faster and scale further.
-
Forbes Read →
Frames Nvidia's edge as system-level co-design of CPU, GPU, and networking, arguing AMD and Intel optimize individual chips while Nvidia controls the full AI factory stack.
AI agents continuously shift between GPU inference and CPU-driven work such as code execution, database lookups, tool invocation and orchestration.
-
NVIDIA Technical Blog Read →
Architectural depth on monolithic die, 164 MB L3, 3.4 TB/s fabric, Spatial Multithreading design, and a graph prefetcher tuned for pointer-heavy agent runtimes.
Agentic AI shifts more of the critical execution path onto the CPU. Agents operate in sandboxes to execute code, invoke tools.
-
NVIDIA Technical Blog Read →
Quantifies RL training impact at 85% environment completions vs 45% on baseline; adds KV-cache eviction analysis showing how CPU speed prevents costly context recomputation.
The CPU is on the critical path for reasoning, response time, and learning because it executes the work between model steps.
-
NVIDIA Blog Read →
Deployment confirmation from May 2026: Anthropic, OpenAI, and SpaceXAI received first units, with OCI committing to hundreds of thousands of CPUs by year-end.
Agentic AI is creating a new CPU moment in the AI factory — as models move from answering to acting, Vera is purpose-built to keep that work moving at scale.
-
NVIDIA Technical Blog Read →
Frames the buyer rationale: Nvidia is redefining server CPU evaluation from cores-per-dollar to tokens-per-dollar, challenging how data center operators size purchases.
CPU time compounds across the request. This makes the CPU part of the critical path.
Originally reported by servethehome.com
Read the original article →Original headline: Nvidia Details Vera CPU: 88 Custom Olympus Cores, 176 Threads and 1.2 TB/s LPDDR5X for Agentic AI