GROMACS Performance on AMD Instinct MI355X

GROMACS Performance on AMD Instinct MI355X#

March 13, 2026 by David Björelind, Sebastian Remander, Paul Bauer.

5 min read. | 1143 total words.

Applications & models

Scientific Computing, HPC, OpenMP, AI/ML

HPC, Developers

Are you planning a hardware upgrade for your molecular dynamics workflows? In this blog, we benchmark GROMACS on AMD’s latest Instinct MI355X GPU and compare it head-to-head with the MI300X, demonstrating significant throughput improvements that accelerate time-to-results for life-science research. You will see exactly how much faster MI355X runs the standard ADH dodec benchmark across 1 to 8 GPUs. Use these results to make informed decisions about your next HPC deployment.

Prerequisites#

Molecular dynamics (MD) simulations are essential for understanding biomolecular systems at the atomic level, including docking and binding interactions key for drug discovery. As system sizes grow and simulation timescales extend, computational performance becomes critical.

Before diving into the benchmarks, here is what you should know:

GROMACS basics: Familiarity with running GROMACS simulations helps, but is not required. We provide context where needed.
Multidir runs: GROMACS multidir launches multiple independent simulation replicas in parallel. This approach suits ensemble MD, free-energy screening, and parameter sweeps.

For details on GPU compute partitioning and how to run GROMACS multidir workloads, see Applying GPU Compute Partitioning for GPU Workloads.

Hardware and Software Setup#

Test System#

We ran all benchmarks on a dual-socket server with the following configuration (see Table 1):

Component	Specification
CPUs	2x AMD EPYC 9654 96-Core Processors
GPUs	8x AMD Instinct MI300X or 8x AMD Instinct MI355X
Software	GROMACS with ROCm support
Table 1: Server configuration used for experiments

Benchmark Workload#

The ADH dodec (alcohol dehydrogenase dodecahedron) system serves as our test case. ADH is a crucial enzyme for metabolizing alcohol and other biological substances, making it a relevant model for drug interaction studies. This medium-sized biomolecular benchmark represents realistic production workloads and is widely used for GPU performance evaluation in the MD community.

Partitioning Modes#

SPX (Single Partition): All 8 XCDs on the GPU operate as one logical device. This is the default mode.
CPX (Compute Partitioned): Each XCD appears as an independent GPU, giving you up to 8 logical devices per physical card.

For multidir workloads that run many independent replicas, CPX mode often delivers higher aggregate throughput. See Applying GPU Compute Partitioning for GPU Workloads for partitioning instructions.

Results and Discussion#

Let’s examine the benchmark data. All performance figures are in ns/day (nanoseconds of simulation time per wall-clock day)—higher is better.

MI300X Throughput#

Table 2 shows MI300X performance across GPU counts, comparing SPX and CPX modes:

Compute Mode	1x MI300X	2x MI300X	4x MI300X	8x MI300X
SPX (Non-Partitioned)	647 ns/day	1,140 ns/day	2,435 ns/day	4,507 ns/day
CPX (Partitioned)	1,570 ns/day	2,822 ns/day	6,022 ns/day	8,026 ns/day
Speedup	2.43x	2.48x	2.47x	1.78x

Table 2: MI300X GROMACS performance with ADH dodec benchmark

Figure 1: MI300X multidir throughput in SPX (blue) vs. CPX (orange) mode. Higher ns/day is better.

As shown in Figure 1, partitioning delivers substantial gains—up to 2.47x at 4 GPUs. Even at 8 GPUs, where CPU resources start to limit scaling, you still see a 1.78x improvement.

MI355X Throughput#

Now let’s look at MI355X. Table 3 shows how the next-generation architecture delivers higher baseline performance:

Compute Mode	1x MI355X	2x MI355X	4x MI355X	8x MI355X
SPX (Non-Partitioned)	888 ns/day	1,321 ns/day	2,890 ns/day	5,952 ns/day
CPX (Partitioned)	2,065 ns/day	4,230 ns/day	8,350 ns/day	10,300 ns/day
Speedup	2.33x	3.20x	2.89x	1.73x

Table 3: MI355X GROMACS performance with ADH dodec benchmark

Figure 2: MI355X multidir throughput in SPX (blue) vs. CPX (orange) mode. Higher ns/day is better.

As seen in Figure 2, MI355X reaches 10,300 ns/day with 8 GPUs in CPX mode. The 2-GPU configuration shows a remarkable 3.20x speedup from partitioning, indicating excellent scaling efficiency at that scale.

Head-to-Head: MI355X vs MI300X#

How much faster is MI355X? Table 4 and Table 5 quantify the generational improvement with and without partitioning enabled:

Configuration	MI355X (CPX)	MI300X (CPX)	MI355X Advantage
1 GPU	2,065 ns/day	1,570 ns/day	+31.5%
2 GPUs	4,230 ns/day	2,822 ns/day	+49.9%
4 GPUs	8,350 ns/day	6,022 ns/day	+38.7%
8 GPUs	10,300 ns/day	8,026 ns/day	+28.3%

Table 4: MI355X advantage over MI300X with GPU partitioning enabled

For non-partitioned workloads, Table 5 shows MI355X still outperforms:

Configuration	MI355X (SPX)	MI300X (SPX)	MI355X Advantage
1 GPU	888 ns/day	647 ns/day	+37.2%
2 GPUs	1,321 ns/day	1,140 ns/day	+15.9%
4 GPUs	2,890 ns/day	2,435 ns/day	+18.7%
8 GPUs	5,952 ns/day	4,507 ns/day	+32.1%

Table 5: MI355X performance advantage over MI300X without GPU partitioning

Discussion#

Key Findings#

The benchmarks reveal several important insights:

MI355X delivers substantial gains over MI300X. In partitioned mode, MI355X achieves 28-50% higher throughput, depending on the GPU count (see Table 4). The largest gain (49.9%) occurs at the 2-GPU configuration, suggesting MI355X scales particularly well in this regime.

GPU partitioning benefits both architectures. Both MI300X and MI355X show 1.7x to 3.2x speedups when using CPX mode for multidir workloads (see Figure 1 and Figure 2). This technique is essential for maximizing throughput in ensemble MD scenarios.

Scaling efficiency varies with GPU count. At 8 GPUs with 128 replicas, speedups from partitioning are slightly lower (1.73-1.78x). At this scale, system resources may become a limiting factor as replica count increases.

Note

The ADH benchmarks are configured to offload all compute-intensive work to the GPU, keeping the CPU off the critical path. However, pair list regeneration — an infrequent but potentially expensive step — remains CPU-bound. Systems with more physical CPU cores can reduce the overhead of this step, particularly at high replica counts where many pair list updates may coincide.

Practical Implications#

Whether you are a researcher running production simulations or a developer optimizing HPC workflows, these results offer actionable guidance:

Leverage GPU partitioning on both MI300X and MI355X. Enabling CPX mode delivers 1.7-3.2x throughput gains for multidir workloads on either GPU generation. This is a straightforward configuration change that benefits any ensemble-style simulation campaign.
MI355X offers the highest throughput. For new deployments or hardware refreshes, MI355X provides 28%+ performance improvement over MI300X, enabling faster iteration on research questions and shorter queue times for batch jobs.
Balance compute resources. Tune the number of concurrent replicas for optimal performance, especially at very high replica counts.

Summary#

This benchmark comparison demonstrates that the AMD Instinct MI355X delivers significant performance improvements for GROMACS molecular dynamics simulations compared to the MI300X. Key takeaways:

MI355X achieves up to 10,300 ns/day throughput on ADH dodec with 8 GPUs in partitioned mode
MI355X performance advantage ranges from 28% to 50% over MI300X, depending on configuration
GPU compute partitioning (CPX mode) provides 1.7-3.2x speedups on both architectures
The best results require balancing GPU partitions and tuning replica counts for your specific system configuration

For life-science researchers running ensemble MD simulations, free-energy calculations, or parameter sweeps, upgrading to MI355X accelerates time-to-results and enables larger-scale computational campaigns.

Disclaimers#

Third-party content is licensed to you directly by the third party that owns the content and is not licensed to you by AMD. ALL LINKED THIRD-PARTY CONTENT IS PROVIDED “AS IS” WITHOUT A WARRANTY OF ANY KIND. USE OF SUCH THIRD-PARTY CONTENT IS DONE AT YOUR SOLE DISCRETION AND UNDER NO CIRCUMSTANCES WILL AMD BE LIABLE TO YOU FOR ANY THIRD-PARTY CONTENT. YOU ASSUME ALL RISK AND ARE SOLELY RESPONSIBLE FOR ANY DAMAGES THAT MAY ARISE FROM YOUR USE OF THIRD-PARTY CONTENT.

The benchmarks presented in this blog were conducted under specific test conditions. Performance may vary based on system configuration, software versions, workload characteristics, and other factors. The information contained herein is for informational purposes only.