CN-122001390-A - TSV perception self-adaptive hybrid fault-tolerant coding method for three-dimensional integrated circuit

CN122001390ACN 122001390 ACN122001390 ACN 122001390ACN-122001390-A

Abstract

The invention discloses a TSV perception self-adaptive hybrid fault-tolerant coding method oriented to a three-dimensional integrated circuit, and belongs to the technical field of calculation, calculation or counting. The coding method comprises the steps of firstly executing type identification and dynamic characteristic evaluation on data needing interlayer transmission through TSVs, then selecting mixed coding such as error correction, difference and dynamic near zero suppression according to data types according to coding parameter sets, completing bus alignment at a sending end, sequentially completing de-interlacing, alignment unpacking, corresponding error correction decoding, difference and near zero reconstruction after the data are transmitted to a receiving end through the TSVs, simultaneously estimating an error rate, and adaptively updating the coding parameter sets according to the error rate and eye pattern characteristics of the receiving end so as to improve the transmission quality of a next channel. The invention improves the reliability and signal integrity of interlayer data transmission under the high coupling noise scene, simultaneously effectively reduces hardware overhead and power consumption, and is suitable for cross-layer data transmission of high-speed and high-bandwidth TSV interconnection.

Inventors

LIU BO
SONG YUAN
LI NINGYUAN
LIU DENGKE
YANG JIALIN
XIA CHENJIE
CAI HAO
YANG JUN

Assignees

东南大学

Dates

Publication Date: 20260508
Application Date: 20251215

Claims (10)

1. The TSV perception self-adaptive hybrid fault-tolerant coding method for the three-dimensional integrated circuit is characterized by comprising the following steps of: Step 1, performing type identification and dynamic characteristic evaluation on transmission data between target TSV channels; Step 2, performing hybrid coding according to the data type and dynamic characteristic evaluation result identified in the step 1 according to the current coding parameter set, wherein the coding parameter set comprises a multi-level coding strategy Code rate Threshold of dynamic near zero suppression mechanism Differentially encoded threshold Differentially encoded window size ; Step 3, after the data buses after mixed coding are aligned, the data buses are transmitted to a receiving end through a target TSV channel, and then decoding and reconstruction are carried out; step 4, calculating the error rate according to the data obtained after decoding and reconstruction; and 5, adaptively updating the coding parameter set transmitted next time according to the error rate calculated in the step 4 and the eye pattern characteristics measured by the receiving end.
2. The method for three-dimensional integrated circuit-oriented TSV perception adaptive hybrid fault-tolerant encoding according to claim 1, wherein the data types identified in the step 1 comprise semi-precision floating point data, fixed point weight data and data with local correlation.
3. The method for three-dimensional integrated circuit-oriented TSV perception self-adaptive hybrid fault-tolerant coding is characterized in that the multi-level coding strategy comprises coding strategy protection levels of single-error correction double-error detection double-adjacent error correction codes, single-error correction double-error detection codes, parity check and direct connection in sequence from high to low.
4. The TSV aware adaptive hybrid fault tolerant coding method for a three-dimensional integrated circuit according to claim 3, wherein step 2 performs hybrid coding according to the data type and dynamic characteristic evaluation result identified in step 1 according to the current coding parameter set, specifically: For half-precision floating point data, selecting sign bits and exponent bits as key bits and encoding according to multilevel coding strategy Protection by threshold value for non-critical bits A controlled dynamic near zero suppression mechanism; for fixed point weight data, packaging to form a data packet, and encoding key bits in the data packet according to a multilevel coding strategy Protecting, and keeping the rest bits to the original value; for data with local correlation, window size in differential encoding And (5) performing window self-adaptive differential coding.
5. The method for three-dimensional integrated circuit-oriented TSV aware adaptive hybrid fault-tolerant coding according to claim 4, wherein in step 3, a fixed word length packet is constructed for the hybrid coded data, and the packet header information position is adjusted or the packet is merged so that the packet to be transmitted is aligned with the bus.
6. The method for three-dimensional integrated circuit-oriented TSV perception adaptive hybrid fault-tolerant coding according to claim 5, wherein in step 3, the data bus after hybrid coding is aligned and then transmitted to the receiving end through the target TSV channel, and then decoding and reconstruction are performed, specifically: Firstly, carrying out de-interlacing and alignment unpacking processing on data transmitted to a receiving end through a target TSV channel; Then, the key bits are subjected to multi-level coding strategy Decoding the corresponding error correction code, performing error positioning and correction, complementing non-key bits of the semi-precision floating point data according to the low order bits of the last valid sample, and performing differential encoding on the window self-adaptive differential encoding data segment according to the window size of differential encoding And performing differential decomposition code reconstruction.
7. The method for TSV perception adaptive hybrid fault-tolerant encoding for three-dimensional integrated circuits according to claim 6, wherein in step 4, the bit error rate is calculated according to the data obtained after decoding and reconstruction, specifically: , wherein, In order to achieve a bit error rate, To count the average value of the received "1/0" level at the moment when the receiving end decides to be optimal, To count the standard deviation of the received "1/0" level at the moment when the receiver decides optimal, For the a priori probability of "1/0" at the transmitting end, As a complementary error function.
8. The method for adaptively updating the coding parameter set of the next transmission according to the bit error rate calculated in the step 4 and the eye pattern feature measured by the receiving end in the step 5, specifically, when the bit error rate of the current transmission is increased, the protection level of the coding strategy is improved, and the method is reduced according to the following steps 、 And is reduced in size Trend update of (a) 、、 When the error rate of the current transmission is reduced, the protection level of the coding strategy is reduced according to relaxation 、 And increase Trend update of (a) 、、 Is a value of (a).
9. The three-dimensional integrated circuit-oriented TSV aware adaptive hybrid fault tolerant coding method of claim 8 wherein the updating is performed by 、、 The values of (a) are specifically as follows: , , , Wherein, the And Values before and after threshold updating for the dynamic near zero suppression mechanism, And For the adjustment factor of the dynamic near zero suppression mechanism threshold, And The values before and after the threshold update for the differential encoding, And For the adjustment coefficient of the differentially encoded threshold value, And The values before and after the updating for the differential encoding window size, 、 And For the adjustment factor of the differential encoding window size, For a target bit error rate preset according to the TSV channel signal-to-noise ratio, For the reference eye-diagram height, For the reference eye-diagram width, For average error rate, for The height of the eye pattern, Is the width of the eye pattern, In order to perform the rounding operation, 、、、、、、 The TSV channel model is preset through experiments.
10. The three-dimensional integrated circuit-oriented TSV aware adaptive hybrid fault tolerant coding method according to claim 9, characterized in that 、、、、、、 The method comprises the steps of obtaining insertion loss under different TSV parameters based on TSV channel models through experimental presetting, dividing at least two insertion loss grades according to the obtained value range of the insertion loss, establishing a mapping relation between the insertion loss grades and adjustment coefficient configuration, and selecting corresponding adjustment coefficient configuration from the mapping relation according to the grade of the insertion loss of a target TSV channel.

Description

TSV perception self-adaptive hybrid fault-tolerant coding method for three-dimensional integrated circuit Technical Field The invention relates to a three-dimensional integrated circuit, interlayer interconnection and fault-tolerant coding engineering technology, in particular discloses a TSV perception self-adaptive hybrid fault-tolerant coding method for a three-dimensional integrated circuit, which is suitable for reliable and high-efficiency transmission of weight, activation and intermediate results in a 3D neural network accelerator, and belongs to the technical field of calculation, calculation or counting. Background The Three-dimensional integrated circuit (3D IC) takes a Through-Silicon Via (TSV) as a core to realize chip stacking and interconnection in the vertical direction, can obviously improve interconnection bandwidth in a limited area, reduce cross-layer communication time delay, and has obvious advantages in the aspects of system level energy efficiency and heterogeneous integration. The technology can obtain stable mechanical and electrical connection by introducing high-density and low-impedance interconnection channels in the vertical direction and relying on a bonding process, and inhibit signal loss caused by parasitic effect, and is more beneficial to supporting the construction of a large-scale parallel and high-energy-efficiency system compared with the traditional welded interconnection. TSVs have been widely used in high bandwidth Memory, near Memory Computing (NMC), and other scenarios, and have driven the evolution of system morphology from two-dimensional to three-dimensional between "processor-Memory". However, the dense TSV array may still cause bit flipping and signal integrity degradation of interlayer transmission under the effects of coupling crosstalk, process deviation, high-speed switching, and the like, so as to put higher reliability requirements on upper layer coding and transmission mechanisms. In high-performance scenes such as a neural network accelerator, a heterogeneous computing platform and the like, TSVs become key bearing of an interlayer data path by virtue of the advantages of high bandwidth, low time delay and high density interconnection, and are responsible for high-speed and stable transmission of multi-type data such as weights, activation, intermediate results and the like, wherein the data are often represented by adopting mixed precision and multiple bits. The neural network has certain fault tolerance capability, is derived from the statistical average effect caused by parameter redundancy and interlayer accumulation and robustness formed in the training and quantization processes, and the related research compares the bit level sensitivity of the hybrid precision network, so that the neural network has stronger fault tolerance space for the least significant bit disturbance and is more sensitive to the most significant bit, and the rule has consistency in different models and data types. These facts indicate that the combination of data type, bit importance and local correlation in the inter-layer transmission implements differential protection and cooperates with the physical mapping and transmission organization of TSVs, helping to achieve a better balance between reliability and resource overhead. The interlayer transmission facing the neural network distinguishes the protection intensity according to the data characteristics and combines codes such as difference, and the like, so that the precision is maintained, and meanwhile, the load of hardware and energy consumption is reduced, thereby improving the overall efficiency of the three-dimensional integrated system. The prior method mainly improves the availability and the robustness of the TSV path from two paths, namely one type suppresses interference through a process at a physical level, and the other type adopts error correction coding, redundant transmission or simple compression and bus coding with uniform intensity at a circuit and architecture level so as to reduce switching activity. The method can relieve error codes under specific conditions, but has common limitations that firstly, the protection intensity is static configuration and lacks perception of fluctuation of channel quality along with changes of environment and workload, secondly, the method is insufficient in utilization of data type difference, bit level importance and local correlation, and is easy to be subjected to over protection of non-key bits or insufficient protection of key bits, thirdly, the coding parameters and transmission organization strategies are lack of cooperative optimization, and transmission quality and energy efficiency are difficult to be stably maintained under different network models and different data stream forms. In summary, the present invention is directed to a TSV-aware adaptive hybrid fault-tolerant encoding method for three-dimensional integrated circuits to overcome the above-men