EpochPOWER™ · The Disruptive Cluster Architecture
The network was never the answer.
Every cluster architecture built on TCP/IP inherits the same ceiling: protocol overhead, throughput collapse, and thousands of lines of code just to manage a network that shouldn't be in the critical path at all. EpochPOWER removes it — nodes talk over the PCIe OmniBus Matrix, not Ethernet.
Patented PCIe OmniBus Matrix architecture
The Incast problem
Request data from seven servers at once. Watch your bandwidth collapse to a fraction of what the link can carry.
When data is striped across networked storage nodes, a client reading fragments from multiple sources simultaneously can overload switch buffers — a TCP throughput collapse of one or two orders of magnitude below link capacity. This is Incast, and it's structural to network-based clustering, not a tuning problem.
At 64 storage nodes with a 1024KB switch buffer, observed throughput reaches roughly 65% of link capacity — and larger buffers only delay the collapse. A switch large enough to avoid it outright has run past $500K.
Where traditional clusters struggle
Cluster software exists almost entirely to compensate for the network underneath it.
Data segmentation — Metadata directories track where fragments live, requiring client-request redirection that consumes network capacity and diminishes throughput on every read.
Node-to-node spillage — When a node fills, data spills to the adjacent node — a problem requiring thousands of lines of code and a serious performance hit each time it happens.
Forced replication — Dedicated per-node storage means a node failure makes that data unavailable — so multiple copies become a requirement, not a choice, at real storage and power cost.
File-size limitations — Clusters not designed for files exceeding a single node's capacity create hard problems for large-file workloads — petroleum exploration, weather modeling, anything that doesn't fit neatly.
The solution
Stop connecting nodes through a network. Connect them through the Matrix.
A GateStor PCIe OmniBus Host Bus Adapter extends the OmniBUS® Matrix across every node in the cluster. Any processor can see the collective memory and adapters of any other node — the cluster behaves less like a network of separate machines and more like a single parallel supercomputer.
Traditional Cluster — switch-mediated: Every request passes through a shared switch buffer — the same fixed resource that causes Incast the moment enough nodes respond at once.
EpochPOWER — Matrix-unified: Nodes share one PCIe OmniBus Matrix directly — the lowest level of I/O interconnection, with no network protocol and no switch buffer to overload.
What this eliminates
Six problems that traditional clusters manage in software, gone at the hardware level.
Network protocol overhead — All nodes connect to the PCIe OmniBus Matrix instead of a network — no TCP/IP, no protocol tax on every transfer.
Data segmentation — Any node can access any other node's storage. It becomes a common pool rather than a set of assigned, isolated volumes.
Node-to-node spillage — Data grows across the shared pool with no requirement to fit on a single node's storage.
Forced duplication — Since any node reaches any data, redundancy becomes optional — a performance choice, not a survival requirement. Substantial savings in storage and power follow directly.
Availability & load balancing — Any node can take over any host connection dynamically, with no prior setup — multiple simultaneous failures and the cluster keeps operating.
Incast timeouts — Dynamic sub-clusters can serve a defined number of hosts while still reaching all data in the larger cluster — something traditional systems can't support at all.
PCIe OmniBus broadcast transfers
One transfer, every node — instead of hours of copying between them.
A patented GateStor technology moves data in a single transfer across every node on the Matrix, eliminating the double and triple transfer operations (to memory, then to the RAID controller) inherent in conventional storage systems. In a cluster, this avoids the duplicate transfers that on other systems can take hours — sometimes days.
Scale
Point-to-point for small clusters. A stackable switch fabric for a thousand nodes.
A 32-port PCIe OmniBus switch handles node and host server connections and can be stacked to accommodate clusters up to 1,024 nodes, with multiple switches providing redundant paths.
An interconnect standard, not a workaround.
EpochPOWER reflects the direction infrastructure is already taking — PCIe-connected NVMe, dynamic resource allocation, and clusters that pivot between compute and data roles as demand changes. By putting the whole cluster's memory and adapters on one Matrix, it lets AI train and infer across the collective in place — the model for the intelligence infrastructure the next generation of workloads will run on.
EpochPOWER™ · Patented PCIe OmniBus Matrix Cluster Architecture
Explore all platforms