HTX301
Trillion-Parameter-Class AI Inference, On-Premise.
HTX301 is the world’s first decode-first chip purpose-built for on-premise AI inference.
Built on Skymizer’s HyperThought™ LPU IP and LISA™ v3 ISA, HTX301 combines scalable model memory, inference-native silicon, and an integrated software stack to bring large and ultra-large model inference inside enterprise-controlled infrastructure.
- Up to 512 GBDRAM per high-capacity accelerator configuration
- Up to ~1TTrillion-parameter-class model capacity
- Decode-FirstPurpose-built for large-model serving economics
- PCIe-BasedFrom enterprise systems to Mini Data Center scale
Maximum memory, model capacity, power, performance, and form factor depend on hardware configuration, model architecture, precision / quantization, software release, and system design.
Why HTX301
Large-Model Inference Is a Memory and Infrastructure Problem
Running frontier-scale AI inside the enterprise is difficult not only because models require compute, but because serving them continuously requires large model-memory capacity, efficient data movement, KV-cache management, predictable latency, and manageable power.
Conventional GPU infrastructure often solves these requirements by scaling across multiple accelerators, high-speed fabrics, and complex rack-level systems. HTX301 takes a different approach: HyperThought is architected around the economics of serving large language models, with particular focus on the memory bandwidth and capacity that dominate the decode phase.
The result is a simpler path to private large-model inference for supported workloads.
Core benefits
Built Around the Workload That Runs Every Day
Ultra-Large Model Capacity
Scale to high-capacity accelerator configurations with up to 512 GB of DRAM and trillion-parameter-class model capacity.
Decode-First Architecture
Prioritize sustained token generation rather than repurposing silicon primarily designed for training.
Single-Card Simplicity
For supported workloads, reduce the need to distribute one model across complex GPU rack-level infrastructure.
Enterprise Data Sovereignty
Keep models, prompts, enterprise data, outputs, and operating policies under direct organizational control.
Predictable, Owned Cost
Own inference capacity instead of relying entirely on metered per-token cloud consumption.
Scale with Demand
Start with a practical deployment and expand from accelerator to rack to Mini Data Center as capacity requirements grow.
Configurations
One Architecture. Right-Sized for the Deployment.
HTX301 uses a modular compute-memory architecture. An “S” denotes the number of HTX301 compute-memory sets in the configuration — it does not define a fixed memory capacity by itself. Actual memory capacity depends on the selected memory implementation and product version.
HTX301-1S
Standard ConfigurationDesigned for evaluation, development, model enablement, AI Box, edge-server, and targeted enterprise inference applications.
- 1 compute-memory set
- Selected versions may start from 32 GB
- Lower system integration, power, and cooling requirements
- Suitable for software validation and targeted production workloads
- Practical PCIe integration path
HTX301-4S
Standard Commercial ConfigurationThe primary practical configuration for enterprise and system-integration deployments.
- 4 compute-memory sets
- Memory capacity depends on selected implementation
- High-capacity variants may scale up to 256 GB
- Balanced model capacity, system integration, power, and cooling
- Suitable for enterprise private AI and large-model deployment
HTX301-8S
Custom High-Capacity ConfigurationThe highest-capacity implementation of the current HTX301 platform, for projects that need greater model capacity than the standard configurations.
- 8 compute-memory sets
- Up to 512 GB depending on memory implementation
- Designed for ultra-large and trillion-parameter-class capacity targets
- Requires purpose-designed casing / chassis, cooling, power, and mechanical integration
- Intended for strategic customer, OEM, ODM, and platform-partner projects
The public HTX301 platform capability scales up to eight inference chips per accelerator. Commercial configuration, memory, physical form factor, thermal design, power delivery, supported models, and availability vary by partner platform and project.
Architecture
Purpose-Built for Decode
Decode-First LPU
HTX301 is designed around the sustained token-generation phase of large-model inference — the model-memory access, bandwidth, data movement, and execution behavior that dominate decode — instead of maximizing general-purpose training throughput.
LISA™ v3 ISA
HTX301 ships on Skymizer’s LISA™ v3 language instruction set architecture, providing a common architectural model between inference hardware, compiler, runtime, and model execution.
Compiler-Hardware Co-Design
Skymizer’s compiler is designed with knowledge of the underlying LPU execution pipeline, memory architecture, and data movement — enabling model-level optimization rather than treating the accelerator as independent kernels.
Scalable Compute-Memory Architecture
HTX301 can scale from single-device configurations to multi-device accelerator designs while retaining the same underlying HyperThought software and architectural foundation.
Prefill / decode
Different Phases. Different Infrastructure Economics.
Prefill is typically compute-intensive. Decode repeatedly generates tokens and is increasingly constrained by memory bandwidth, capacity, and data movement. HyperThought treats these phases independently — HTX301 can run as dedicated inference infrastructure, or as the decode-oriented component of a heterogeneous system in which GPUs or other resources handle prefill.
- User Request
- Prefill Resource
- KV Transfer / Handoff
- HTX301 Decode
- Token Stream
Heterogeneous prefill/decode is one supported architecture option, not the only deployment mode.
Deployment scale
Start with One System. Scale to a Mini Data Center.
Edge AI / AI Box
Compact HTX301 deployments for local model inference and specialized applications.
Enterprise On-Premise Server
Standard configurations integrated into enterprise-controlled infrastructure.
On-Premise Rack
Multiple accelerators and systems serving higher concurrency and larger workloads.
Mini Data Center
Multiple heterogeneous nodes connected through system-level scheduling and orchestration.
Grow inference capacity without changing the underlying HyperThought architecture.
Economics
Put Inference Economics Back Under Enterprise Control
Cloud inference is easy to start but can become difficult to predict as token volume, agent activity, context size, and user concurrency grow. HTX301 gives enterprises the option to own inference capacity directly.
Data Sovereignty
Run where enterprise data already lives.
Predictable TCO
Plan around owned hardware capacity and utilization instead of relying entirely on variable per-token pricing.
Lower Infrastructure Complexity
For validated workloads, reduce dependence on large GPU clusters, proprietary high-speed fabrics, and rack-scale cooling.
Incremental Growth
Scale capacity as demand grows instead of re-architecting the entire deployment.
Software
A Hardware-Software Inference Platform
HTX301 is not silicon alone. A validated HTX301 software release can include:
- Firmware
- Linux BSP
- PCIe / kernel driver
- Runtime
- SDK
- Compiler-generated model artifacts
- Containerized deployment environment
- Inference-engine integrations
- HTX301 CLI
- Diagnostics & validation tools
- Model profiles
- Model Zoo entries
Models and engines are marked Validated / Early Access / Preview / In Development / Plannedonly after they pass release-specific validation. Each supported model in the HTX301 Model Zoo is paired with a recommended configuration, memory requirement, software release, inference engine, model format, quantization, and validated execution path.
Use cases
Built for Private Enterprise AI
- Ultra-Large Private LLMs
- Run high-capacity language models inside infrastructure the enterprise owns and controls.
- Enterprise AI Agents
- Support persistent agentic workflows that generate continuous model calls across enterprise systems.
- Retrieval-Augmented Generation
- Private enterprise search, knowledge management, customer support, and regulated-information workflows.
- Coding / Engineering Assistants
- Keep proprietary source code, RTL, design information, and engineering context inside controlled infrastructure.
- Multimodal AI
- Support applications combining text, images, documents, audio, video, and structured data.
- Sovereign / Regulated AI
- Infrastructure for sectors where data location, model control, auditability, and operational independence matter.
Delivery model
From HTX301 Silicon to Partner-Built AI Systems
Skymizer is a platform enabler. HTX301 silicon and the associated HyperThought software stack can be integrated by board partners, OEMs, ODMs, server manufacturers, system integrators, distributors, and solution providers to build accelerator cards, AI Boxes, enterprise servers, private AI appliances, industry-specific AI platforms, custom high-capacity systems, and Mini Data Center solutions.
Not every final system configuration is a standard Skymizer-branded retail card.
Roadmap
Proven Silicon. A Platform Built to Evolve.
HTX301 ships on LISA™ v3 today. LISA™ v4 is in development to extend HyperThought toward the next generation of inference workloads:
- Mixture-of-Experts
- Multimodal inference
- Agentic inference
- Higher memory bandwidth
- KV-cache efficiency
- 2-bit-class quantization
- Sparsity
- Advanced process-node scaling
Roadmap technologies and future process implementations are not current HTX301 shipping specifications.
Evaluate
Evaluate HTX301 Against Your Model
Tell us your model, size, precision, target context length, expected concurrency, latency requirement, and deployment environment. Skymizer can help identify the appropriate HTX301 configuration, memory requirement, software status, model compatibility, and evaluation path.
- Appropriate HTX301 configuration & memory requirement
- Software status & model compatibility
- Server / chassis requirements & evaluation path
Prefer email? sales@skymizer.ai