CiMBA Speeds Genome Sequencing with On-Device Basecalling
As genome sequencing is finding utility in a wide variety of domains beyond the confines of traditional medical settings, its computational pipeline faces two significant challenges.
Content & Liability Disclaimer
This article and its accompanying video are automated summaries derived from the original research paper by Unknown authors. The original research was conducted solely by the paper's authors; PDFdigest did not conduct any of the research and makes no claims of ownership over the underlying scientific work.
The video narration is generated by artificial intelligence and references the paper's authors for attribution. The video is not narrated by any of the paper's authors. This content may contain inaccuracies, omissions, or misinterpretations of the original research. First-person language (e.g., "we found", "our results") reflects the original authors' voice, not PDFdigest's. Always read the original paper for accurate, verified information before making any decisions based on this content.
This content is provided "as is" without any warranties, express or implied. Simulated systems OÜ, its officers, directors, employees, and agents shall not be liable for any direct, indirect, incidental, special, consequential, or punitive damages arising from your use of, reliance on, or access to this content, including but not limited to errors, omissions, or misinterpretations of the original research. This disclaimer applies to the fullest extent permitted by applicable law.
- 1 As genome sequencing is finding utility in a wide variety of domains beyond the confines of traditional medical settings, its computational pipeline faces two significant challenges.
Vi. Methodology
VI. METHODOLOGY To comprehensively explore CiMBA/AL-Dorado’s feasibility as a real-time basecaller, we analyze both CiMBA’s runtime basecalling characteristics, as well as study the impact of CiM noise on AL-Dorado in terms of single inference as well as downstream analysis accuracy.
Vii. Results
VII. RESULTS This section details analysis results on both the CiMBA architecture and the AL-Dorado DNN basecaller. We analyze performance both in terms of the throughput and power consumption of CiMBA, as well as the accuracy implications of performing basecalling on a future CiMBA prototype.
Ii. Background
II. BACKGROUND Oxford Nanopore Technology’s (ONT) SotA sequencing flow for generating ultra-long reads (up to 2.2 million bases) enables analysis of human genome regions inaccessible by other sequencing technologies [52] , and recovery of highlycontiguous, even nearly complete, microbial genomes [53] . ONT’s portable, handheld sequencing machines, the MinION and MinION Mk1C, illustrated in Figure 1 , enable diverse biomedical applications ranging from clinical diagnostics [2] , [3] to environmental monitoring [9] , [54] . ONT sequencing uses flow cells composed of nanoscale channel arrays containing nanopores. When a DNA.
How PDFdigest Helps You Understand Research
Instant Paper Analysis
Get structured summaries and key findings from dense PDFs in seconds.
Visual Explanations
Turn complex methods, figures, and results into clearer visual breakdowns.
AI-Powered Q&A
Ask focused questions and get answers grounded in the paper.
C. Enabling efficient, parallel DNN operations on CiM tiles
C. Enabling efficient, parallel DNN operations on CiM tiles Crystalline (ordered) Amorphous (disordered) G1 + G2 + G1 – G2 – t1 t2 V V w1∝(G1 + -G1 -) x1∝Vt1 w2∝(G2 + -G2 -) x2∝Vt2 (x1w1+x2w2)∝(Vt1(G1 + -G1 -)) + Vt2(G2 + -G2 -)) 1 1 2 2 4 4 4 4 (a) (b) (c) 3 3 3 3 512×512=262,144 Simultaneous MACs CiM Tile 2 2 Pulse-width modulator Inp. Activations 1 1 I=VC ADC Converter Post-Processing x512 TABLE I COMMUNICATION/STORAGE OVERHEAD FOR 9 DATASETS [80] IS REDUCED BY 43.7×/4.37× VIA.
A. Real-time genome analysis
A. Real-time genome analysis Portable sequencing enabled a wide range of applications beyond clinical scenarios [7] – [9] , driven in large part by ONT’s introduction of the MinION Mk1C, a sequencing device with an onboard Jetson TX2 embedded GPU [81] , pictured in Figure 1 -a. However, even at the current level of flow cell technology, the TX2 has trouble maintaining realtime throughput [13] , and further improvements to flow cell technology will soon require more compute power than is offered by the embedded GPU [16] , [43] .
D. CiM-amenable model architectures
D. CiM-amenable model architectures Not all network architectures are amenable for analog acceleration. For instance, depth-wise-based bottleneck layers [87] results in poor array utilization (¡0.1%) [88] , addressed in [89] by alterations in the model architecture or in [88] by a hybrid digital-analog acceleration. Furthermore, layers with uneven row/column aspect ratios or tiny kernels may result in under-utilization of the CiM arrays [67] . Modelarchitecture co-design is therefore critical to fully benefit from CiM acceleration.
B. 2D mesh based architecture
B. 2D mesh based architecture As specified in Section II-C, a high bandwidth interconnect is needed to support the CiM tiles’ low latency and energy efficiency on the system-level. Moreover, the interconnect should allow flexibility on mapping workloads to the heterogeneous fabric, as neural network model architectures may undergo rapid changes. For CiMBA, we adopt a 2D mesh as described in [67] and illustrated in Figure 5 -a, to move activations between nodes. This mesh comprises of multiple sets of parallel wires, running in the X and Y direction and.
B. AL-Dorado model architecture
B. AL-Dorado model architecture Upon the Dorado Fast architecture, we explored a range of architectural modifications to optimize the network for CiMBA. The AL-Dorado networks were developed through several design/experiment iterations, which will be elaborated on in Section VII. For brevity, only one studied AL-Dorado network is presented here, illustrated in Figure 7 . Specifically, the LSTM size is boosted from dimensions of 96 to 128 for layers 1-3, and to 256 for layers 4-5 to account for the heterogeneous layer response to analog conversion detailed in Section VII-D. We.
C. AL-Dorado training HW/SW environment
C. AL-Dorado training HW/SW environment Training is performed using A100 GPUs via distributed data parallel training. We use the Bonito software repository as a base for developing the AL-Dorado model [58] rather than the newer Dorado repository as its Python implementation enables more expedient development/experiment iterations, however, the latest Dorado-Fast model is ported from the Dorado repository. The network is trained until validation accuracy saturates, a total of 30 epochs. To study the impact of analog noise sources described in Section III-C and develop the mitigation strategies detailed in Section.
D. PCM hardware validation
D. PCM hardware validation In order to define AIHWKIT’s tile configuration, we also have at our disposal a physical PCM memory array consisting of >1 million PCM cells allowing single device read/writes, pictured in Figure 9 [97] . We characterize the read, write, and drift characteristics of this chip and configure AIHWKIT to simulate these characteristics. Table III indicates array key characterization parameters. The exact functionality of AIH-WKIT is beyond the scope of this work, and we recommend the reader to [98] for a detailed description of the implemented methodologies.
E. Training/validation/bacterial datasets and limitations
E. Training/validation/bacterial datasets and limitations Training and validation DNA datasets collected on R9.4.1 flow cells are downloaded from the Bonito repository [58] . They consist of 65k/1000 full reads for training/validation, respectively, split into chunks of 4000. The validation dataset is used to analyze single chunk accuracy of Dorado-Fast and AL-Dorado in floating point and on CiMBA, as each validation chunk comes with a reference sequence. To verify accuracy beyond inference accuracy, we also perform post-basecalling analysis on a set of reads generated using a MinION R9.4.1 flowcell. Table I.
F. Post-basecalling analysis flow
F. Post-basecalling analysis flow We evaluate AL-Dorado performance using aligned basecalling accuracy, i.e., the total number of exactly matched bases between a read and the reference genome divided by the total alignment length including insertions and deletions. We basecall each read set, producing either a FASTQ or FASTA file suitable for downstream analysis. We align each basecalled read to its corresponding reference genome of the same species using the state-of-the-art read mapper, minimap2 [100]. We use Rebaler [101] to generate a consensus sequence from each basecalled read set before polishing.
A. CiMBA performance analysis
A. CiMBA performance analysis Figure 10 illustrates CiMBA’s performance against the SotA baselines described in Section VI-A. As can be seen, CiMBA’s throughput outperforms all devices except the A100, as expected when comparing against a data-center level GPU. However, when throughput is balanced against power and area footprint, CiMBA outperforms all other embedded devices by at least 16.5×/27× in terms of bps/W and bps/mm 2 , respectively. At 25mm 2 and with an average power consumption of 1.17W, CiMBA favorably compares to the MinION Mk1C’s 322mm 2 embedded GPU in.
B. CiMBA runtime analysis
B. CiMBA runtime analysis Figure 11 highlights a subset of interesting results gained from our system level simulations of CiMBA. Specifically, we are able to see in (a) the breakdown of runtime into different op categories. VMMs refer to the LSTM matrix multiplication operations, while LSTM Ops account for all operations required to calculate the input, forget, cell, and output gates, and the hidden state. Other includes CNN, clamp, batchnorm, and LA decoder operations. Resource contention accounts for any time an operation must wait for a preceding operation to release.
D. Model sensitivity to analog nonidealities
D. Model sensitivity to analog nonidealities To address PCM drift, we measure individual layers’ sensitivity to analog noise by maintaining portions of the network in digital while converting the rest of the network to analog, illustrated in Figure 13 . As can be seen, layers are not impacted equally by analog execution. Namely, it is clear that the CNN layers, particularly the first layer, are highly sensitive to analog noise. This can be explained by the layer’s 1×5 kernel; as only 5 PCM cells contribute to the analog VMM, these.
G. Downstream analysis
G. Downstream analysis Figure 16 illustrates the downstream analysis of the 9 microbial datasets listed in Table I . The network’s varied performance across datasets aligns with previous research on the same dataset [39] , [103] . We observe that the accuracy loss between Dorado Fast in floating point and its analog equivalent, along with the loss for AL-Dorado with the aforementioned LA decoder parameters, is consistent with the values reported in Sections VII-E and VII-F. This demonstrates that AL-Dorado’s CiM aware retraining and optimization strategies generalize beyond the training/validation.
Frequently Asked Questions
As genome sequencing is finding utility in a wide variety of domains beyond the confines of traditional medical settings, its computational pipeline faces two significant challenges.
VI. METHODOLOGY To comprehensively explore CiMBA/AL-Dorado’s feasibility as a real-time basecaller, we analyze both CiMBA’s runtime basecalling characteristics, as well as study the impact of CiM noise on AL-Dorado in terms of.
VII. RESULTS This section details analysis results on both the CiMBA architecture and the AL-Dorado DNN basecaller. We analyze performance both in terms of the throughput and power consumption of CiMBA, as.
Yes. PDFDigest can turn this paper into a structured explanation, key takeaways, visual summaries, and a narrated video when available.