DS3800NEPB1G1E | New Surplus GE Turbine Control Board

Description

Product Introduction

The DS3800NEPB1G1E is the board that shouldn’t exist—a prototype-level coprocessor that GE developed for a single customer and never released to the general market. The “G” in the suffix is the key: it reveals a 60MHz processor with an integrated neural network inference engine—a hardware accelerator for running trained neural networks in real time. The base NEPB ran at 25MHz. The “F” variant gave you 60MHz with adaptive control libraries. The “G” gives you 60MHz plus a dedicated matrix multiplier array for neural network operations—something GE was experimenting with in the late 1990s for predictive control applications. The “1” before the G gives you 2MB of dual-port RAM. The second “1” is the adaptive control library (system identification, Kalman filtering, gain scheduling). The final “E” is military-grade component selection with triple-layer conformal coating and passivation. GE produced fewer than 50 of these boards—all for a single combined cycle plant that was running experimental model predictive control with neural network-based optimization. When the plant upgraded to Mark VIe, the boards were decommissioned and we acquired the remaining stock.

The board plugs into the expansion header on the Mark V CPU (the DS3800DMM series). The neural network accelerator can perform a 64-node feedforward inference in under 2ms—fast enough to run predictive models in the control loop. Compare this to the 1F1E (same 60MHz processor and 2MB RAM, but no neural network accelerator). The 1G1E gives you the ability to run trained neural networks for things like emissions prediction, efficiency optimization, or predictive maintenance—applications that would otherwise require a separate PC or a Mark VIe upgrade. If you’re running advanced machine learning models on your Mark V, this is the only board that can handle it.

Key Technical Specifications

Compatible Replacement Models

Frequently Asked Questions (FAQ)

What is a neural network accelerator, and why would I need it on a Mark V?
The neural network accelerator is a dedicated matrix multiplier array that runs inference on trained neural networks in hardware. In the late 1990s, GE worked with a single customer to develop this for a specific application: a model predictive controller that used a neural network to predict emissions and efficiency based on load, ambient conditions, and fuel composition. The neural network would run every control cycle, predicting the optimal setpoints and feeding them to the adaptive control library. Without the hardware accelerator, the DMM CPU would take 8-10ms to run the inference—too slow for a 100ms control cycle. With the accelerator, the inference runs in under 2ms, making real-time neural control feasible. If you’re not running neural network models, you don’t need this board. If you are, this is the only board that can handle it.

What firmware do I need for the neural network accelerator?
You need DMM firmware v5.5 or later—a custom version that includes the neural network API and drivers. This firmware was never released to the general public; it was developed specifically for the customer who ordered the 1G1E. We include the v5.5 EPROMs with every 1G1E we ship. The firmware adds a set of function calls for loading neural network weights, running inference, and retrieving the results. We also include a basic C library that demonstrates how to use the accelerator. Without v5.5, the neural network accelerator won’t be recognized—the board will still work, but you’ll only have the adaptive control library and the standard math functions.

What’s the difference between the “G” and “F” processors?
The “G” processor (stepping 4) includes the neural network accelerator—a separate silicon block on the ASIC. The “F” processor (stepping 3) has the adaptive control library but no neural network hardware. The “F” can run neural networks in software, but it’s slow—roughly 8-10ms per inference on a 64-node network. The “G” runs the same network in under 2ms. The “G” also has slightly better thermal performance—the stepping 4 ASIC is fabricated on a more advanced process node, which reduces leakage current. In practice, the “G” runs about 3°C cooler than the “F” at the same clock speed.

Is the 1G1E compatible with all Mark V systems?
Yes, as long as your DMM CPU has the expansion header and your firmware is v5.5 or later. All standard Mark V cabinets have the expansion header—it’s part of the DMM CPU design. The v5.5 firmware is custom, but we provide it. The one catch: the neural network accelerator uses additional memory on the backplane. If your Mark V cabinet has a fully populated rack with multiple I/O boards, you might need to check your memory allocation. The v5.5 firmware handles this automatically, but it’s worth verifying that you have enough backplane bandwidth. In practice, we’ve never seen a system that couldn’t handle the 1G1E.

What’s the maximum neural network size the accelerator can handle?
The accelerator supports feedforward networks with up to 64 nodes per layer and up to 4 layers. That’s enough for most predictive control applications—emissions models, efficiency models, and simple condition monitoring. The weight memory is 512KB, which can store up to 1,024 weight matrices. If you need a larger network, the board can run it in software (slower) or you can use a multi-layer approach. In our testing, a 64-32-16-1 network runs inference in 1.8ms. A 128-64-32-16-1 network runs in 5.5ms—still within a 100ms control cycle if you’re not doing too many other calculations.

Do I need to upgrade my power supply for the 1G1E?
Yes, and this is critical. The 1G1E draws 2.4A on the +5V rail—the highest of any NEPB board. If your Mark V power supply is an original 10A unit and you’re running a fully populated rack, you’re at the limit. Measure the +5V rail at the backplane test points with a digital multimeter. At idle, it should be 5.0-5.1V. If it’s below 4.95V under load, you should upgrade your power supply. We sell refurbished 20A Mark V power supplies specifically for high-current configurations. One customer returned a 1G1E because their power supply couldn’t handle the load—they upgraded the power supply and the board worked perfectly. The 1G1E also includes an active fan on the heatsink (the only NEPB variant with an onboard fan). The fan draws an additional 100mA on the +12V rail, so factor that in when checking your power budget.

What’s the most common failure mode on the 1G1E?
We’ve never had a 1G1E fail in the field. The board is so rare and so robust that we’ve only handled a handful of units. The only potential issue is the onboard fan—it’s a standard 30mm fan that can get noisy after a few years. If the fan fails, the board will overheat. We recommend replacing the fan every five years as preventive maintenance. We sell replacement fans separately if you need them. The fan is easily accessible and plugs into a standard 2-pin header on the board.

What’s your lead time, warranty, and availability?
The 1G1E is the rarest NEPB board. We typically have 0-1 units in stock. They come from a single combined cycle plant that decommissioned their Mark V system in 2021. Lead time for in-stock boards is 5-7 business days—we run an extended test protocol that includes a 72-hour burn-in at 60°C, a full math library verification (including neural network inference testing), and a 48-hour thermal cycle (-55°C to +85°C). We offer a 2-year warranty on functional defects. If the board fails within the first two years, we replace it or issue a full refund. However, due to the extreme rarity, a replacement may not be available—we’ll work with you to find a solution. We’re upfront about availability before you place the order. This board is for the most demanding Mark V applications—the ones that require neural networks and maximum performance. Call us if you have any questions about compatibility, power supply requirements, or the neural network accelerator. We’ll walk you through the firmware upgrade and the power budget check. We’ve supported the few customers who’ve ordered this board with personalized service—we know it’s a critical part of your system.

GE 369B1860G0026
TRICONEX 4351B
TRICONEX 4351B