Raw throughput
9.8 TFLOPS per socket from 512-bit vector units and a branch predictor trained on real HPC traces. Sustained clocks hold within 3% of peak across 72-hour runs — no thermal cliff, no silent throttling.
// NEXT-GEN SERVER PROCESSORS — COMPUTE 2026
128 cores of raw compute. Zero compromise.
The Nimbus X9 pairs 128 compute cores with 9.8 TFLOPS of vector throughput and a 5.0 GT/s coherent fabric — inside a 350 W envelope engineered for the AI-first data center.
- 128 coresCORES / SOCKET
- 9.8 TFLOPS peak vector throughputPEAK TFLOPS
- 5.0 gigatransfers per second fabric linkFABRIC LINK
- 2.1 times performance per wattPERF / WATT
// MODULES — X9 ARCHITECTURE
Engineered as modules, not monoliths
Every subsystem is isolated, instrumented, and hot-swappable in software — profile compute, power, and fabric independently.
Efficiency by design
Adaptive clock domains park idle cores in under 40 ns and re-boost them just as fast. At 40% load the X9 still runs at 82% efficiency — the same rack envelope now holds 2.1× the usable FLOPS.
Scale without limits
A 5.0 GT/s coherent fabric stitches 4,096 sockets into one system image with sub-microsecond latency. ECC covers every hop, and failed links re-route around faults faster than a scheduler can notice.
// PIPELINE — DEPLOYMENT
From pallet to petaflop in three steps
No cabling spreadsheets, no week-long burn-in. The fabric does the topology; you do the science.
Deploy
X9 sleds slide into standard 19-inch racks — tool-less, with auto-negotiating fabric links. Rack, power, and walk away.
Orchestrate
The Nimbus controller discovers every socket, maps coherence domains, and exposes the fleet as a single scheduler endpoint in under 60 seconds.
Scale
Add sleds under load. Thermal and workloads re-route across the fabric in real time — 2 to 2,048 nodes with zero reboots.
// FIELD REPORTS
Operators who moved to X9
“We cut inference cost per token by 38% in the first week on X9 clusters. The fabric is the real story — east-west latency simply disappears.”
“2,048 nodes, one system image, zero unplanned reboots in 14 months. Our previous fleet couldn’t survive a single firmware update.”
“Perf-per-watt closed the deal. We fit 40% more compute into the same power envelope, and the PUE curve followed it down.”
// PROCUREMENT
Pick your deployment class
Every tier ships with next-business-day replacement in year one. Volume terms available from 16 nodes up.
Single Node
$1,890 /SOCKET
per socket · 1-year NBD warranty
- 1× X9 socket — 128 cores
- 5.0 GT/s fabric port
- Standard firmware channel
- Community forum support
- Dedicated deployment engineer
- 24/7 priority SLA
Cluster
$7,900 /NODE
per dual-socket node · 3-year warranty
- 2× X9 sockets — 256 cores
- Redundant fabric links
- Nimbus orchestration controller
- 24/7 priority SLA
- Dedicated deployment engineer
- Custom TDP binning
Data Center
CUSTOM
volume licensing · negotiated SLA
- Fleet-wide fabric licensing
- Custom TDP binning
- On-site thermal audit
- Named solutions architect
- 99.99% uptime SLA
- Single-unit retail orders
// SPEC SHEET — Q&A
Frequently asked questions
How does the X9 compare to the previous generation?
On HPL and Llama-class inference benchmarks the X9 delivers 2.1× the performance per watt of the X7. Per-core thermal telemetry re-allocates the heat budget every 40 ns, so sustained clocks stay within 3% of peak across 72-hour runs.
Which socket, OS, and software stacks are supported?
The X9 uses the LGA-N9 socket. Linux kernels 6.x and newer are supported day one, with first-class integration in Kubernetes and Slurm schedulers. GPU and TPU accelerators attach over the 5.0 GT/s fabric without host-bus bottlenecks.
What are the power and cooling requirements?
TDP is 350 W per socket. Standard sleds are air-cooled up to a 35 °C inlet; liquid-cooled sleds unlock full boost in dense 1U racks. Per-core power telemetry streams to the orchestration controller in real time.
What support and warranty coverage is included?
All tiers include next-business-day replacement for the first year. Cluster and Data Center plans add a 24/7 priority SLA, signed firmware delta channels, and a named deployment engineer for the life of the fleet.
How quickly can we deploy a pilot cluster?
Sleds ship in 2–4 weeks. A 16-node pilot — racked, orchestrated, and running your first workload — is typically online in under a single business day.
// EARLY ACCESS — WAVE 03
Put the X9 to work.
Join the 2026 early-access program. Silicon, sleds, and orchestration — bundled for pilot clusters, deployed in a day.
NO SPAM · UNSUBSCRIBE ANYTIME · HPC TEAMS ONLY