Our Products
Lorem ipsum dolor sit amet, consectetur adipiscing elit ipsum lorem sit amet consectetur dolot amet adipscing amet.
What Drives the AI Visibility Gap?
H3 placeholder for section subtitle, lorem ipsum dolor sit amet, consectetur adipiscing elit, quisque venenatis eleifend mattis
Problem Statement: Four drivers of lack of visibility
Messy Configs (FA) | Real-time (FM) | Correlation (WM) | Attribution (FM + WM)
Addressing the Visibility Gap: Clockwork Fleet Audit,
Fleet Monitoring, Workload Monitoring
From clean starts to continuous uptime: end-to-end AI fleet assurance
Deploy a reproducible, known good baseline fleet to run AI workloads
Keep the fleet healthy, performant and cost-effective while AI jobs run
Clockwork Solution
H3 placeholder for section subtitle, lorem ipsum dolor sit amet, consectetur adipiscing elit, quisque venenatis eleifend mattis
Fleet Audit
Body text placeholder. Lorem ipsum dolor sit amet, consectetur adipiscing elit ipsum lorem sit amet consectetur dolot amet adipscing amet.
- Software checks
- Node checks
- Front-end network
- Back-end GPU network validation
“I want to make sure my cluster is configured correctly before I run a week-long training job.”
“I found (i) 3 through cabling checks; (ii) 7 through cross-cluster ping6 test; BUT (iii) 3 are unique that I would not have found!”
Fleet Monitoring
Body text placeholder. Lorem ipsum dolor sit amet, consectetur adipiscing elit ipsum lorem sit amet consectetur dolot amet adipscing amet.
- Runtime link failures/flaps
- Runtime fabric topology
- Runtime fabric performance
- Congestion and contention monitoring
“We want to detect network failures/link flap as soon as they happen, and not when our jobs stall!”
“We want to be sure that replacement GPUs in the cloud are meeting topology/latency SLAs?”
“We’d like to track latency continuously and get alerted when it goes above our set thresholds.”
Workload Monitoring
Body text placeholder. Lorem ipsum dolor sit amet, consectetur adipiscing elit ipsum lorem sit amet consectetur dolot amet adipscing amet.
- Deep workload visibility
- Correlation of data path performance with network metrics to identify root cause of job performance
“I ran all_reduce_perf workload twice, 1st run ~360Gbps, 2nd run only ~190Gbps. What could be the problem?”
“We saw a sudden slowdown in job performance, could it be network-related?”
Why High Precision Attribution Matters
H3 placeholder for section subtitle, lorem ipsum dolor sit amet, consectetur adipiscing elit, quisque venenatis eleifend mattis
Out-of-band and in-band Qpair one-way-delays are very different. The workload was mistakenly configured to use RoCEv1 instead of RoCEv2.
AI clusters are very complex configurations with a Cambrian explosion in the number of configuration options and AI reliable runs are very sensitive to these. The ability to have an early warning indicator of failures and slowdowns down to nanosecond accuracy – aka – high precision attribution – becomes a significant time-to-market advantage.
Accelerate AI Jobs with Real-time Resilience
Transform Your Compute Efficiency
Body text placeholder. Lorem ipsum dolor sit amet, consectetur adipiscing elit ipsum lorem sit amet consectetur dolot amet adipscing amet.
Body text placeholder. Lorem ipsum dolor sit amet, consectetur adipiscing elit ipsum lorem sit amet consectetur dolot amet adipscing amet.
Disruptive Network Failures and Link Flaps
Are Common and Expensive
H3 placeholder for section subtitle, lorem ipsum dolor sit amet, consectetur adipiscing elit, quisque venenatis eleifend mattis
One of the most common problems encountered is Infiniband/RoCE link failure. Even if each NIC-to-leaf switch link had a mean to failure rate of 5 years, due to the high number of transceivers, it would only take 26.28 minutes for the first job failure.
lost per incident
Source: Falcon: Pinpointing and Mitigating Stragglers for Large-Scale Hybrid-Parallel Training, 2024; The Llama 3 Herd of Models, 2024; “Alibaba HPN: A Data Center Network for Large Language Model Training”, ACM SIGCOMM ’24; Gemini: Fast Failure Recovery in Distributed Training with In-Memory Checkpoints, 2023
Clockwork’s Workload Failover Provides
Resilience To Link Flaps
H3 placeholder for section subtitle, lorem ipsum dolor sit amet, consectetur adipiscing elit, quisque venenatis eleifend mattis
Body text placeholder. Lorem ipsum dolor sit amet, consectetur adipiscing elit ipsum lorem sit amet consectetur dolot amet adipscing amet.
Link/NIC flapping
-
Body text placeholder
-
Lorem ipsum dolor sit amet, consectetur adipiscing elit ipsum lorem sit amet
Node Inbound
Infiniband Throughput
-
Body text placeholder
-
Lorem ipsum dolor sit amet, consectetur adipiscing elit ipsum lorem sit amet
Body text placeholder. Lorem ipsum dolor sit amet, consectetur adipiscing elit ipsum lorem sit amet consectetur dolot amet adipscing amet.
Detecting & Eliminating Contention
H3 placeholder for section subtitle, lorem ipsum dolor sit amet, consectetur adipiscing elit, quisque venenatis eleifend mattis
Body text placeholder. Lorem ipsum dolor sit amet, consectetur adipiscing elit ipsum lorem sit amet consectetur dolot amet adipscing amet.
Workload Acceleration
Proven Throughput Gains Across Real-World AI Workloads
Hyperscaler with Clockwork vs Dynamic Load Balancing (DLB)
2 all-to-all jobs
The hyperscaler with Clockwork enabled has 33% more outbound throughput vs. DLB
Large Social Media Company with Clockwork vs. ECMP
2 all-to-all jobs
The large social media company with Clockwork enabled has 29% more throughput vs. ECMP
Learn More
Stop wasting GPU cycles. Start scaling smarter.
Clusters must deliver high uptime while running at maximum efficiency.
Turn your GPU clusters into a competitive advantage—not a cost center.
