On August 13, 2026, the 5th Dishui Lake China RISC-V Industry Forum was successfully held at the InterContinental Shanghai Dishui Lake.
Lu Xingqiang, Marketing Director of Shenzhen Geesemi Co., Ltd., was invited to attend and introduced the company's new DSP product based on RISC-V and its self-developed instruction set—the GS32MT8800.
Photo | Lu Xingqiang, Marketing Director of Shenzhen Geesemi Co., Ltd.
Established four years ago, Geesemi's product line currently covers digital power, server power, and high-speed rail transit, among other fields. Robotics is a new direction that has only begun mass shipping in the past year.
Two 400MHz Cores, Current Loop Runs Within 500ns
The GS32MT8800 uses Geesemi's self-developed DSP300 core, which is customized and optimized based on the N900 base core, featuring dual cores each running at 400MHz.
A relatively intuitive metric in the field test is the current loop response—professionals in the motor control field know the importance of this indicator. Currently, this chip can achieve loop control within 500ns.
Regarding this, Lu Xingqiang provided a comparative reference: under the same conditions, the loop performance of the GS32MT8800 can roughly achieve 1.5 to 2 times that of similar chips on the market. In terms of actual processing capability, this is roughly equivalent to an ARM core running at 1.6GHz.
Single Chip Controls 25 Degrees of Freedom, Solving the Space Issue
Dexterous hands are a relatively typical application scenario in current humanoid robots. A single hand has 22 to 25 degrees of freedom. The traditional approach uses three or even more chips for coordinated control, which brings problems such as numerous communication harnesses, poor coordination, and large control board footprint. The interior of robot joints is already very cramped; stuffing a few more boards inside would either exceed the weight limit or fail to fit in the available volume.
Therefore, the approach of the GS32MT8800 is to converge all control within 25 degrees of freedom into a single chip, relying on the chip's internal scalable architecture and a sufficient number of encoder interface resources.
Lu Xingqiang pointed out that processing the sampling, calculation, and command dispatch for all joints with a single chip not only eliminates the latency of inter-chip communication but also saves harnesses and board area. Currently, this level of integration is relatively high among publicly disclosed solutions.
Achieving 200MHz Performance at 100MHz by Cutting Latency
Simply increasing the clock frequency does not equate to good control real-time performance; this is a consensus in the motor control field.
Regarding this, Lu Xingqiang revealed that the primary optimization for the GS32MT8800 is not computing power, but the latency of the entire chain—from current sampling to loop calculation and then to control parameter dispatch. In traditional methods, hardware-software interaction consumes a large portion of the time.
What Geesemi has done with this chip is to pipeline the post-sampling processing, calculation, and control parameter dispatch, reducing intermediate handshakes and waiting. The final result is that at the same 100MHz clock frequency, the actual control performance can approach the level of competitors running at 200MHz. This difference is quite noticeable in scenarios with a 100kHz control frequency, and 100kHz currently basically covers the dynamic response requirements of humanoid robots for the next two to three years.
Self-Built Toolchain and Algorithm Library, Ready for Direct Customer Evaluation
The general ecosystem for RISC-V is still slowly developing. Geesemi's choice is to develop a full set of toolchains in-house, covering everything from math libraries and OS encapsulation to debugging simulation and mass production burning. For robotics applications, they simultaneously provide reference solutions for single-axis servo, multi-axis collaborative arms, and dexterous hands, delivering both the hardware baseboard and algorithm library together so that customers can directly run evaluations on the board.
Currently, customers who have already received shipments of the GS32MT8800 are concentrated in the directions of humanoid joint servo, multi-axis control for collaborative robots, AGV chassis, and dexterous hands.
Feedback from Geesemi's market side indicates that the most frequently asked question from customers currently is not whether the performance is sufficient, but when they can get more samples and more mature reference code.