Sub-millisecond interconnect

Raw compute tuned
for frontier scale.

On-demand GPU clusters for training, inference, and large-scale compute.

View benchmarks
Clusters Cluster igs-prod-01

Cluster igs-prod-01

Running
GPU cluster · 512 × H100 · us-west-2
Node utilization i
96% at 14:27
Nodes 512
GPUs 512
GPU utilization 96%
GPU memory 712 TB
Interconnect 0.6 ms
Cluster inspector
Cluster ID igs-prod-01
Status Running
Region us-west-2
Instance H100 (80 GB)
Nodes 512
GPUs 512
Total memory 40.0 TB
Interconnect 0.6 ms
Created Jan 26, 2025, 10:14 UTC
Node ID Rack / Leaf GPU Util Thermal Fabric RTT State
igs-w2-n001 rack-a01 / ib-04 97.4% 63°C 0.58 ms Healthy
igs-w2-n019 rack-a03 / ib-04 96.1% 64°C 0.61 ms Healthy
igs-w2-n142 rack-b12 / ib-08 95.8% 65°C 0.59 ms Healthy
igs-w2-n388 rack-d06 / ib-14 96.5% 62°C 0.60 ms Healthy
Job ID Workload Nodes Step Time Status
job-9401 frontier-70b-pretrain-s4 384 412 ms Running
job-9398 rlhf-policy-eval-v12 96 118 ms Running
job-9391 kv-cache-shard-bench 32 24 ms Running
Fabric & compute benchmarks
AllReduce (8 GB)364 GB/s
MFU sustained61.8%
Checkpoint write1.8 TB/s
Thermal variance±1.4°C
BENCHNCCLRing+Tree auto-tuned over 8x NDR 400Gb/s rails — zero packet drops (0 / 4.2e11)
BENCHNVLINKIntra-node bidirectional peak 900 GB/s across all 512 active H100 GPUs
BENCHCLOCKZero thermal throttling detected over 14-day continuous training window
14:27:08SCHEDGang allocation verified: 512/512 bare-metal nodes synchronized in 184ms
14:27:12FABRICInfiniBand NDR leaf health nominal; p99 tail latency 0.60ms across us-west-2
14:27:19CKPTAsync snapshot #418 flushed to NVMe tier (1.82 TB/s, 0 stalled steps)
14:27:25TELEMPeak aggregate GPU utilization reached 96.0% (avg core temp 64°C)
SSH & API endpoints
SSHssh ops@igs-prod-01.us-west-2.ignis.internal
APIhttps://api.ignis.compute/v1/clusters/igs-prod-01
VPNWireGuard mesh active · 14 authorized engineering keys
RBACSOC2 Type II enforced · hardware-backed FIDO2 tokens
Job dispatched to igs-prod-01
Velt
Norra
Kinet
Sable
Orin
Mero
Architecture

Engineered for sustained
peak throughput.

A purpose built cluster architecture that keeps training, inference, and large scale compute at peak.

Zero throttling

Sustained peak for demanding workloads

1,980 MHz locked · 64°C avg

Direct fabric

0.2ms
0.6ms
3.2 TB/s

Low latency across the entire cluster

Scale clusters without
hidden bottlenecks.

Get more from your infrastructure with high-performance, efficient training at scale.

Bare-metal
nodes
RDMA fabric
Instant
checkpoints
Live thermals
Cluster load 96%
Sculptural matte-black hexagonal basalt columns illuminated by a single molten ember seam
Bare-metal silicon · Liquid-cooled

Single-tenant physical racks wired directly to non-blocking NDR fabric.

Hardware & fabric

Bare-metal silicon,
measured cleanly.

Compare node memory, collective throughput, and fabric latency across dedicated Ignis configurations.

Specification H100 SXM5 B200 HGX NDR RDMA
Memory per node 640 GB HBM3 1.44 TB HBM3e GPUDirect
Peak dense compute 15.8 PFLOPS FP8 72.0 PFLOPS FP4 In-network SHARP
Intra-node link 900 GB/s NVLink 1.8 TB/s NVLink 3.2 Tbps / node
All-reduce latency 0.62 ms 0.48 ms 0.39 ms rail-opt
NVMe scratch tier 30.7 TB Gen5 61.4 TB Gen5 120 GB/s Lustre
Thermal envelope Direct-to-chip Closed-loop hydro Non-blocking 1:1
Verified across 512-GPU continuous pre-training runs with zero thermal throttling. SOC 2 Type II · ISO 27001

Built for unbroken
training runs.

Live bare-metal telemetry across every rack and NDR rail.

Collective throughput 512 × H100 SXM5 · 8× NDR 400Gb/s
Sustained MFU 62.4%
AllReduce BW 368 GB/s
Step jitter 0.14 ms
368 GB/s
14-day continuous pre-training · 0 step stalls us-west-2 · live
Pricing

Transparent compute,
built for frontier runs.

Zero egress penalties, bare-metal isolation, and deterministic InfiniBand pricing across every tier.

12-month reservation — save 22% vs hourly rates

On-Demand

Instant bare-metal nodes for evaluations, fine-tuning, and burst inference.

$2.45 / GPU · hr
8 to 64 GPUs · billed per minute
  • H100 SXM5 bare-metal nodes
  • 3.2 Tbps RoCEv2 & InfiniBand
  • Instant NVMe checkpoint snapshots
  • Live telemetry & SSH root access

Dedicated Pod

Isolated private superclusters with custom topology and hydro cooling.

$1.54 / GPU · hr
1,024+ GPUs · private hall & SRE
  • Physically air-gapped GPU suites
  • Custom Slurm / Kubernetes images
  • Embedded fabric engineers on call
  • 100% geothermal Iceland or US-West

Reserve dedicated
compute in minutes.

Private GPU clusters for training, inference, and production.