Data center
Recover core capacity currently spent moving packets and return it to workloads that generate revenue. Higher user density on the same footprint, on existing hardware.
Next-Generation Networking on x86 Hardware
High-Performance Data Acceleration
GTX2 improves infrastructure economics by accelerating data movement across existing and next-generation enterprise systems.
GateSpeed increases the productive capacity of servers, storage and networks while improving utilization of valuable compute and AI accelerators. The GTX2 platform makes these gains deployable, manageable, and scalable across the infrastructure.
01 — The problem
Throughput Limits Limit Business.
Every workload that matters now — AI training and inference, streaming, storage replication, security inspection — is bounded by how fast data moves between the network and the application. When transport is the bottleneck, the cost shows up everywhere else: cores burned on packet handling instead of revenue work, user density capped well below what the hardware could carry, and roadmaps shaped around what the network can survive rather than what the business wants to build.
Three approaches dominate today: 1) Dedicated hardware, 2) Specialized fabrics, and 3) Kernel bypass. Each delivers real throughput. Each carries real limitations. Numbered one to three, each is set out in turn — what it does well, and what it costs you.
How it's done today
1Dedicated hardware
Fast, but fixed
Purpose-built appliances and ASIC-based systems deliver throughput, and for one well-defined function they are hard to beat. But capacity is fixed at purchase, lead times are long, and cost and power scale with every additional box.
Refresh cycles run in years. The capacity you bought is the capacity you have until the next one, and adapting to a new workload means new hardware rather than new configuration.
2Specialized fabrics
Different fabrics for different workloads
InfiniBand excels in tightly coupled AI training and HPC clusters, where ultra-low latency and high-bandwidth node-to-node communication are critical. Ethernet + TCP/IP is better suited to applications, clients, storage, inference, and distributed infrastructure.
The two are complementary. InfiniBand earns its cost inside the training cluster, where latency between nodes is the binding constraint. GTX2 raises throughput across the Ethernet estate around it — applications, clients, storage, inference — on the hardware already in the rack.
3Kernel bypass
Closing the Integration Gap
DPDK and similar frameworks provide a fast path from the network interface to user space, laying the foundation for high-performance networking. But packet I/O is only the beginning. ARP, IP and TCP processing, routing, NAT, filtering, and TLS still have to be implemented and integrated above it.
That work is where projects stall, and where the cost keeps living long after the benchmark is met — every layer rebuilt by hand is a layer your team maintains.
02 — The platform
Ultra-high performance on flexible, low-cost hardware.
Through the GateSpeed platform, GTX2 delivers throughput associated with specialized silicon — on commodity x86 servers you can already buy, from low-cost Atom-class CPUs to server-class Xeons.
Scales with cores, not with budget
Throughput scales linearly as cores are added, across CPU classes. Capacity planning becomes a question of how many cores to assign rather than which specialized appliance to buy and how long it takes to arrive.
What the platform provides
A complete platform, not just kernel bypass
Built on an optimized DPDK foundation, GateSpeed provides a high-performance user-space network and transport stack, standalone network functions, and direct application integration — all without the Linux networking stack in the packet-processing data path.
Network functions that stand on their own
Deployment-ready functions including routing, firewall, load balancing, and VPN can operate independently and scale across CPU cores without requiring application integration.
A network stack that integrates with applications
Applications such as NGINX and Apache can integrate directly with GateSpeed's user-space network and transport stack through a shared-object socket library, bypassing the Linux networking stack for packet processing.
Network functions that extend the application
GateSpeed network functions can also be incorporated into the application data path, processing traffic before or after the application while extending it with integrated networking, security, and traffic-management capabilities.
03 — Architecture
Your Application The NIC. Nothing in Between.
GateSpeed creates a high-performance user-space data path to the NIC, bypassing the Linux kernel networking stack for packet processing.
The platform includes a suite of high-performance network functions and a shared-object network and transport stack for direct integration with applications such as NGINX. Selected network functions can be incorporated into the application data path, enabling services such as packet filtering and traffic management before traffic reaches the application.
Packet-processing capacity scales linearly with each additional CPU core.
Fig. 1 — The kernel networking stack is not in the data path.
04 — Validation: Measured, Not Modeled
GTX2: GateSpeed + Intel Labs
GTX2 partners with Intel Labs and Intel Network Builders to test the GateSpeed platform. Intel's 2024 white paper documents GateSpeed's throughput and concurrent-user capabilities, and the companion webinar walks through the architecture.
As the work extends to AI workloads, Intel's April 2026 partner brief outlines GateSpeed's role in improving performance in edge AI and private cloud environments.
200 Gbps
Single-core L3 forwarding
Layer 3 packet forwarding at 200 Gbps on one Intel Xeon Scalable core across 512–1518 byte packets, with 80% processing headroom remaining at 1518 bytes.
82–90 Gbps
NGINX web serving, HTTP
Near line rate on a 100 GbE link using four or fewer cores, holding steady from 5,000 to 40,000 concurrent connections. HTTPS with Intel QAT reached 61–74 Gbps on four cores.
100 Gbps
Site-to-site IPsec VPN
Near line rate on six cores with software-only AES-128-CBC-HMAC-SHA1 encryption. Adding two Intel QAT instances reached the same throughput on four cores total.
8-to-1 reduction
In RAM buffer requirements
Intel reports field test results that per-core efficiency translates to up to an 8-to-1 reduction in RAM buffer requirements in certain scenarios.
Test conditions. Intel Labs testbeds, Rio Rancho. Server A: 2× Intel Xeon Gold 6230N (20-core, 2.3 GHz), 384 GB DDR4. Server B: 2× Intel Xeon Gold 6342 (24-core, 2.8 GHz), 512 GB DDR4, Intel QuickAssist Adapter 8970. Intel Ethernet E810 100 GbE adapters throughout. DPDK 22.07, Intel QAT 4.20, NGINX 1.18, Ubuntu 20.04, kernels 5.8 and 5.15. Hyper-threading on, Turbo off. Published by Intel, January 2024. Performance varies by use, configuration and other factors.
05 — Deployment Options
Deploy it. Integrate it. Extend it.
GateSpeed provides deployment-ready network functions and pre-integrated applications, while the shared-object interface enables additional applications to connect directly to the GateSpeed network and transport stack. Network functions can also be incorporated into the application data path to extend applications with networking, security, and traffic-management capabilities.
High-throughput standalone network functions
- Router
- Load balancer
- Firewall
- IPSec and client-server VPNs
Pre-integrated applications
- NGINX web server
- Apache web server
Integrate your applications
Link the GateSpeed shared object using standard GNU tools. Applications call familiar network socket functions while gaining direct access to the GateSpeed user-space network and transport stack, with RESTful APIs for management. Applications which already call standard socket functions and use shared object socket libraries simply re-link to the GateSpeed shared object.
Extend applications with network functions
GateSpeed network functions can be incorporated directly into the application data path, processing traffic before or after application processing and adding capabilities such as firewall, load balancer, traffic management, and other network services.
06 — Use cases
Four places the stack moves the needle.
Private and hybrid cloud
Consistent, predictable per-core throughput across on-premises and cloud environments — the same stack and the same integration path on both sides of the boundary.
Private and sovereign AI
Training and inference are bounded by how fast data reaches the accelerators. High-throughput transport on commodity servers keeps that pipeline fed inside your own perimeter.
Corporate networks
Firewall, VPN, and web-server acceleration on standard x86 servers, at throughputs that would otherwise call for dedicated appliances or custom silicon.
07 — Team
The GTX2 Team
Deep experience in high-performance networking and infrastructure, custom network infrastructure development and deployment, high-speed applications, and systems integration worldwide.
Jayson White
President
Entrepreneur and researcher at Harvard Kennedy School; studied at Columbia. Author of The Next American City, co-founder of Knotel, WeWork alum.
Julian Lin
CTO
30+ years in RF, satellite, and embedded systems. Three degrees from Rice. Telecom instrument in the Smithsonian collection.
David Giannini
Chairman
30 years in managed network services and biotech. Founded, scaled and exited companies across North America, Europe and Asia. JD, California State Bar.
Tuan Wang
Consulting Engineer
Network performance, DPDK infrastructure, and distributed systems in Python and C/C++. BS Computer Science; MS candidate in Artificial Intelligence at Santa Clara. CCNA certified.
Luke Giannini
Consulting Engineer
Deployment automation, management tooling, and DPDK network configuration. BA in Computer Science, Colorado Boulder; MS candidate at Santa Clara.
Next step