Accelerating Single-Cell Tumor Classification with AMD Radeon™ AI PRO R9700S GPUs, ROCm™ Software, and HIP

Nov 09, 2026

Glowing blue DNA double helix formed by digital particles against a dark background with layered black geometric arrows.

Dayhoff Health accelerated a four-region single-cell tumor classification workflow on AMD Radeon AI PRO R9700S GPUs in Rackspace Technology infrastructure. Combining AMD RDNA™ 4 architecture, 32 GB of VRAM per GPU, multi-GPU scalability, and the open AMD ROCm and HIP software stack, the tested platform reduced the complete workflow from 76.1 minutes in a direct CPU paired run to a three-run median of 11.6 minutes on two GPUs.

A faster path from sequencing data to research insight

Dayhoff Health selected the AMD Radeon AI PRO R9700S as the hardware foundation for this accelerated workflow. Designed for enterprise AI deployments, the R9700S combines AMD RDNA 4 architecture, 32 GB of VRAM per GPU, multi-GPU scalability, and support for the open AMD ROCm software platform. In the Rackspace Technology-hosted environment, one or two R9700S GPUs (gfx1201) worked with AMD ROCm™ 7.x software, as documented in the Dayhoff Health whitepaper, and HIP on a dual-socket AMD EPYC™ 7003 Series server CPUs host with 24 cores per socket and 48 vCPUs. That combination gave Dayhoff Health the compute capacity, memory footprint, and programming environment needed to optimize parallel stages of the single-cell analysis workflow.

On this tested workload, Dayhoff Health reduced end-to-end runtime for the full four-region workflow from 76.1 minutes in a direct CPU paired run to a three-run median of 11.6 minutes on two Radeon AI PRO R9700S GPUs, a 6.59x speedup. The result demonstrates how the R9700S platform can support memory-intensive, highly parallel AI and scientific computing workloads when paired with application-level optimization. The Dayhoff Health whitepaper, “Accelerating Single-Cell Tumor Classification with ROCm™ software and HIP” (https://dayhoffhealth.com/tumor-classification/), explains how 10x Genomics Cell Ranger 10.0.0 converts FASTQ reads to a count matrix and CopyKAT 1.2.5 classifies cells as aneuploid/putative tumor or diploid/normal. It also evaluates agreement with CPU results within clearly stated correctness and clinical-use boundaries.

From raw reads to putative tumor and normal classifications

The benchmark centers on one archival breast-tumor FFPE Flex dataset with four multiplexed regions. In total, the input comprised 1.03 billion reads and 12,375 nuclei. This workload exercises both major stages: Cell Ranger converts FASTQ reads to a count matrix, and CopyKAT classifies cells as aneuploid/putative tumor or diploid/normal.

Within that benchmark environment, 2 TB of local NVMe storage supported the CPU and GPU runs. The R9700S GPU provided 32 GB of VRAM per GPU for demanding data and model workloads, while ROCm™ 7.x software and HIP gave Dayhoff Health an open programming foundation for developing and tuning its GPU implementation. Together, the hardware and software stack supported GPU-resident processing, kernel-level parallelism, and scaling across one- and two-GPU configurations. The results reflect Dayhoff Health’s implementation and the full tested system, not hardware in isolation.

Measured gains across the complete workflow

Measured times were specific to this four-region workload. Cell Ranger complete output was 3900.548 seconds (65.0 minutes) on CPU, compared with GPU medians of 608.550 seconds (10.1 minutes, 6.41x) on one GPU and 528.965 seconds (8.8 minutes, 7.37x) on two GPUs. For CopyKAT across all four regions, the CPU median was 645.382 seconds (10.8 minutes), compared with medians of 279.608 seconds (4.7 minutes, 2.31x) on one GPU and 148.355 seconds (2.5 minutes, 4.35x) on two GPUs.

Workflow stage CPU time 1 GPU median Speedup 2 GPU median Speedup
Cell Ranger complete output 3900.548 s / 65.0 min 608.550 s / 10.1 min 6.41x 528.965 s / 8.8 min 7.37x
CopyKAT, four regions 645.382 s median / 10.8 min 279.608 s / 4.7 min 2.31x 148.355 s / 2.5 min 4.35x
Complete workflow 4567.887 s / 76.1 min direct paired run 898.070 s / 15.0 min 5.09x 693.628 s / 11.6 min 6.59x

How the acceleration works

For the complete workflow, the direct CPU paired run was 4567.887 seconds (76.1 minutes). The three-run GPU medians were 898.070 seconds (15.0 minutes) on one Radeon AI PRO R9700S GPU, a 5.09x speedup, and 693.628 seconds (11.6 minutes) on two R9700S GPUs, a 6.59x speedup. These results show the value of pairing a high-memory, scalable enterprise AI GPU platform with software engineered for the workload. Dayhoff Health targeted specific sources of delay and parallel work across the pipeline rather than simply moving an unchanged application to a different processor.

First, an I/O-less GPU-resident count path removes disk roundtrips, reducing data movement between storage and compute. Second, repetitive operations are mapped across many GPU threads to support parallel processing in this workflow.

The workflow also takes advantage of the R9700S platform’s ability to scale beyond a single GPU. Lane-local Cell Ranger work can be divided across devices, while independent CopyKAT regions can run concurrently. In this tested implementation, two-GPU scaling followed these natural workload boundaries, allowing Dayhoff Health to expand available compute and memory resources without forcing every operation into one shared execution path. The approach illustrates a pattern other teams can evaluate on the R9700S GPU: identify independent work units, limit data movement, and use HIP to express parallel kernels across AMD GPUs.

Performance with carefully defined correctness

Speed matters only when researchers understand how outputs compare. For Cell Ranger, the GPU-produced matrix achieved a per-gene correlation of 1.00000 with the CPU result and approximately 99% element-level parity, rather than full byte identity. The accepted two-GPU output matched the reference across 20 HDF5, 121 CSV, and 30 compressed-matrix files. Dayhoff Health also validated the GPU pipeline against official Cell Ranger outputs supplied by Single Cell Discoveries on two small multiplexed datasets, an independent reference that tests reproduction of the provider’s result rather than agreement with Dayhoff Health’s own CPU run. In deterministic CopyKAT runs, predictions were byte-identical across CPU, one-GPU, and two-GPU execution.

Standard-mode CopyKAT timing runs were not bit-reproducible, so they were compared statistically rather than byte-for-byte. In biological validation of the evaluated regions, positive aneuploid calls showed high precision, while sensitivity and abstention were region dependent. That distinction is important for technical teams evaluating accelerated scientific software. Reproducibility should be assessed at the level appropriate to the algorithm, with deterministic paths compared exactly and statistical paths evaluated using documented agreement and sensitivity measures.

The scope is also specific. This workflow performs analytical cell screening in an already obtained tumor specimen. It is not population screening, exhaustive tumor detection, or a clinically validated diagnosis. CopyKAT classification is not a diagnosis, and routine clinical use would require separate biological and regulatory validation. Performance results are hardware- and dataset-specific and come from one primary patient tumor; they do not establish performance across other datasets, cancer types, GPU configurations, or life-science workloads.

A practical research infrastructure pattern

For organizations evaluating compute-intensive research pipelines, the Radeon™ AI PRO R9700S GPU offers a practical enterprise AI foundation that combines substantial per-GPU memory, multi-GPU deployment options, and an open software ecosystem. In this test, ROCm™ software and HIP supported kernel-level parallelism and multi-GPU scheduling on the R9700S GPU platform. By shortening GPU-accelerated runtimes, research teams can run more development cycles in the same period, test pipeline changes sooner, compare analytical settings more frequently, and refine new sequencing pipelines faster. That faster iteration can support faster innovation in pipeline design, subject to each team’s own data and validation requirements.

The tested environment was hosted by Rackspace Technology, which Dayhoff Health confirmed may be acknowledged. That deployment context reinforces an important consideration for technical decision-makers: performance depends on the full system, including compute, memory, storage, software versions, and workload structure. The whitepaper provides a concrete result, not a universal guarantee, and offers an implementation pattern that other teams can evaluate against their own data and validation requirements.

Turning acceleration into shorter research cycles

Dayhoff Health’s results show how targeted GPU engineering can unlock the capabilities of the AMD Radeon™ AI PRO R9700S GPU for demanding AI and scientific computing workflows. By combining 32 GB of VRAM per GPU, multi-GPU scalability, AMD ROCm™ software and HIP, GPU-resident counting, and workload-aware parallelization, the implementation reduced the tested workflow from a 76.1-minute direct CPU paired run to three-run medians of 15.0 minutes on one R9700S GPU and 11.6 minutes on two.

For researchers and developers, the Radeon™ AI PRO R9700S provides a scalable platform for iterating, investigating, and validating GPU-accelerated workflows. Shorter GPU-accelerated runtimes can compress the research loop, allowing teams to test more pipeline variants, evaluate results sooner, and refine new sequencing pipelines more quickly. For technical leaders, Dayhoff Health’s work offers a concrete example of how high-capacity GPU memory, multi-GPU deployment, and the open ROCm™ software and HIP software stack can support faster research iteration and faster innovation in pipeline development while preserving clearly defined correctness criteria. As with any benchmark, teams should assess the approach against their own data, system configurations, and validation requirements.

Read the Dayhoff Health whitepaper: https://dayhoffhealth.com/tumor-classification/

Learn more about the Radeon™ AI PRO R9700S GPU: https://www.amd.com/en/products/graphics/workstations/radeon-ai-pro/ai-9000-series/amd-radeon-ai-pro-r9700s.html

Endnotes:

All performance and/or cost savings claims are provided by the 3rd party organization featured herein and have not been independently verified by AMD. Performance and cost benefits are impacted by a variety of variables. Results herein are specific to such 3rd party organization and may not be typical. GD-181a.

Share:

Article By


Prosumer Marketing Lead, AMD Ryzen Threadripper & Radeon AI PRO

Business Development | AMD Radeon AI Pro

Related Blogs