This post has been republished via RSS; it originally appeared at: New blog articles in Microsoft Community Hub.
Article contributed by Amirreza Rastegari, Jon Shelley, Jithin Jose, Anshul Jain, Jyothi Venkatesh, Joe Greenseid, Fanny Ou, and Evan Burness
Azure has announced new HBv4-series and HX-series virtual machines (VMs) for high performance computing (HPC). This blog provides in-depth technical and performance information about these new VMs.
These VMs are powered by the latest technologies, including:
- 4th Gen AMD EPYC CPUs (Genoa while in Preview, Genoa-X at General Availability in 1H2023)
- 800 GB/s of DDR5 memory bandwidth (STREAM TRIAD)
- 400 Gb/s NVIDIA Quantum-2 CX7 InfiniBand, the first on the public cloud
- 80 Gb/s Azure Accelerated Networking
- 3.6 TB local NVMe SSD providing 12 GB/s (read) and 7 GB/s (write) of storage bandwidth
HBv4 and HX – VM Size Details & Technical Specifications Overview
HBv4 and HX VMs are available in the following sizes with specifications as shown in Tables 1 and 2, respectively. Just like existing H- VMs, HBv4 and HX-series also include constrained cores VM sizes, enabling customers to choose a size along a spectrum of from maximum-performance-per-VM to maximum-performance-per-core.
HBv4-series VMs
|
VM Size |
176 CPU cores |
144 CPU cores |
96 CPU cores |
48 CPU cores |
24 CPU cores |
|
VM Name |
standard_HB176rs_v4 |
standard_HB176-144rs_v4 |
standard_HB176-96rs_v4 |
standard_HB176-48rs_v4 |
standard_HB176-24rs_v4 |
|
InfiniBand |
400 Gb/s Quantum-2 (NDR) |
||||
|
CPU |
AMD EPYC™ 7004-series (standard Genoa during Preview) |
||||
|
Peak CPU Frequency |
3.7 GHz * |
||||
|
RAM per VM |
688 GB |
||||
|
RAM per core |
4 GB |
5 GB |
7.5 GB |
15 GB |
30 GB |
|
Memory B/W per VM |
800 GB/s |
||||
|
Memory B/W per core |
4.5 GB/s |
5.6 GB/s |
8.3 GB/s |
16.6 GB/s |
33.3 GB/s |
|
L3 Cache per VM |
768 MB |
||||
|
L3 Cache per core |
4.4 MB |
5.3 MB |
8 MB |
16 MB |
32 MB |
|
SSD Perf per VM |
2 x 1.8 TB NVMe – total of 12 GB/s (Read) / 7 GB/s (Write) |
||||
Table 1: Technical specifications of HBv4-series VMs
HX-series VMs
|
VM Size |
176 CPU cores |
144 CPU cores |
96 CPU cores |
48 CPU cores |
24 CPU cores |
|
VM Name |
standard_HX176rs |
standard_HX176-144rs |
standard_HX176-96rs |
standard_HX176-48rs |
standard_HX176-24rs |
|
InfiniBand |
400 Gb/s NDR |
||||
|
CPU |
AMD EPYC™ 7004-series (Preview) |
||||
|
Peak CPU Frequency |
3.7 GHz * |
||||
|
RAM per VM |
1.4 TB |
||||
|
RAM per core |
8 GB |
10 GB |
15 GB |
29 GB |
59 GB |
|
Memory B/W per VM |
800 GB/s |
||||
|
Memory B/W per core |
4.5 GB/s |
5.6 GB/s |
8.3 GB/s |
16.6 GB/s |
33.3 GB/s |
|
L3 Cache per VM |
768 MB |
||||
|
L3 Cache per core |
4.4 MB |
5.3 MB |
8 MB |
16 MB |
32 MB |
|
SSD Perf per VM |
2 * 1.8 TB NVMe – total of 12 GB/s (Read) / 7 GB/s (Write) |
||||
Table 2: Technical specifications of HX-series VMs
*Clock frequencies are based on non-AVX workload scenarios and are based on measured frequency delivery for workloads as captured by the Azure HPC team with AMD EPYC 7004-series processors and corresponding system firmware. Experienced clock frequency by a customer is a function of a variety of factors, including the coding and usage of a given application. Frequencies indicated above are not necessarily indicative of final clock frequencies for EPYC 7004-series processors.
For more information see the official documentation for HBv4-series and HX-series VMs.
Microbenchmark Performance
This section focuses on microbenchmarks that characterize performance of the memory subsystem and the InfiniBand network of the HBv4-series and HX series VMs.
STREAM – Memory Performance
Below in Figure 1, we share the results of running We ran the industry standard STREAM benchmark on HBv4/HX VMs. The STREAM benchmark was run using the following:
sudo ./run_stream_dynamic.py -nt 30 -t 176 -oca 0-175 -m 20000 -thp madvis
This returned a result of ~770 GB/s bandwidth for STREAM-TRIAD, which is over 2x greater than that provided from DRAM on HBv3 VMs (~350 GB/s STREAM-TRIAD) as documented here.
Figure 1: STREAM-TRIAD measures 765.52GB/s Memory Bandwidth for HBv4/HX series VMs
InfiniBand Perftests – Network Performance
HBv4 and HX VMs are equipped with latest NVIDIA Quantum-2 CX7 InfiniBand (NDR) interconnect. We ran the industry standard IB perftests test across two (2) HBv4-series VMs featuring 400 Gb/s (NDR) InfiniBand links. The IB bandwidth test was run using the following:
Unidirectional bandwidth:
numactl -c 0 ib_send_bw -aF -q 2
Bi-directional bandwidth:
numactl -c 0 ib_send_bw -aF -q 2 -b
Results of these tests are depicted in Figures 2 and 3, below.
Figure 2: Unidirectional InfiniBand bandwidth measuring up to the expected peak bandwidth of 400 Gb/s
Figure 3: Bi-directional InfiniBand bandwidth measuring up to the expected peak bandwidth of 800 Gb/s
As depicted above, HBv4/HX-series VMs achieve line-rate bandwidth performance (99% of peak) for both unidirectional and bi-directional tests.
Application Performance
This section will focus on characterizing performance of HBv4 and HX VMs on commonly run HPC applications. Performance comparisons are also provided across other HPC VMs offered on Azure, including:
- Azure HBv4/HX with 176 cores of AMD EPYC “Genoa” (HBv4 full specifications, HX full specifications)
- Azure HBv3 with 120 cores AMD EPYC “Milan-X” (full specifications)
- Azure HBv2 with 120 cores AMD EPYC “Rome” processors (full specifications)
- Azure HC with Intel 44 cores of Xeon Platinum “” (full specifications)
Note: HC-series represents a highly customer relevant comparison as the majority of HPC workloads, market-wide, still run largely or exclusively in on-premises datacenters and on infrastructure that is operated for, on average, between 4-5 years. Thus, it is important to include performance information of HPC technology that aligns to the full age spectrum that customers may be accustomed to using on-premises. Azure HC-series VMs well-represent the older end of that spectrum and also feature highly performant technologies like EDR InfiniBand, 1DPC DDR4 2666 MT/s memory, and Xeon Platinum 1st Gen (“Skylake”) processors that dominated HPC customer investments and configuration choices during that period. As such, application performance comparisons below commonly use HC-series as a representative proxy for an approximately 4-year-old HPC optimized server.
Summary performance improvements with HBv4 and HX VMs compared to our most recent HPC VM offering, HBv3-series VMs are as follows:
- Up to 2.24x higher performance for CFD workloads
- Up to 5.3x higher performance for FEA workloads
- Up to 2.51x higher performance for weather simulation workloads
- Up to 2x higher performance for molecular dynamics workloads
- Up to 1.87x higher performance for rendering workloads
- Up to 2.45x higher performance for chemistry workloads
Computational Fluid Dynamics (CFD)
Ansys Fluent – version 2022 R2
Figure 4: On Ansys Fluent (Aircraft Wing 14M) HBv4/HX VMs provide a greater than 4x performance uplift compared to 4-year-old HPC server (represented by HC-series VMs) and 1.84x higher performance compared to the most recent Azure HPC VM, HBv3-series.
The absolute values for the benchmark represented in Figure 4 are shared below:
|
VM Type |
Average Solver Rating |
|
4 year-old HPC server |
729.77 |
|
HBv2 |
1314.27 |
|
HBv3 |
1764.80 |
|
HBv4/HX |
3247.70 |
Table 3: Ansys Fluent (aircraft wing 14M) absolute performance (average solver rating, higher = better).
In addition, we share here scale-up performance within a single VM:
Figure 5: On Ansys Fluent (Aircraft Wing 14M) performance increases an additional 38% from the 96-core VM size to the 176 core VM size, illustrating the tradeoff between per-core v. per-VM performance.
The absolute values for the benchmark represented in Figure 5 are shared below:
|
HBv4/HX VM size |
Average Solver Rating |
|
96 CPU cores |
2357.5 |
|
144 CPU cores |
2854.0 |
|
176 CPU cores |
3247.7 |
Table 4: Ansys Fluent (aircraft wing 14M) absolute performance (average solver rating, higher = better).
Siemens Simcenter STAR-CCM+ - version 17.04.008
Figure 6: On Siemens Simcenter STAR-CCM+(Civil) HBv4/HX VMs show a greater than 5x performance uplift compared to 4 year-old HPC server, and more than 2x compared to HBv3-series.
The absolute values for the benchmark represented in Figure 6 are shared below:
| VM Type |
Average Solver Rating |
|
HBv2 |
1314.267 |
|
HBv3 |
1764.8 |
|
HBv4/HX |
3247.70 |
|
HC |
729.77 |
Table 5: Siemens Simcenter STAR-CCM+(Civil) absolute performance (time elapsed, lower = better).
In addition, we share here scale-up performance within a single VM:
Figure 7: On Siemens Simcenter STAR-CCM+ (Civil) time to solution decreases by nearly 40% from the 96-core VM size to the 176 core VM size, illustrating the tradeoff between per-core v. per-VM performance.
The absolute values for the benchmark represented in Figure 7 are shared below:
|
HBv4/HX VM size |
Time Elapsed (sec) |
|
96 CPU cores |
1.81 |
|
144 CPU cores |
1.42 |
|
176 CPU cores |
1.29 |
Table 6: STAR-CCM+(Civil) absolute performance (time elapsed, lower = better) across HBv4/HX VM sizes.
As we can see from the scale-up performance figures for Ansys Fluent and Siemens Simcenter STAR-CCM+, respectively, Constrained Cores HBv4/HX VMs provide significant benefits for customer workloads that may require lower core count due to commercial software licensing constraints. For example, looking at Table 4 for Ansys Fluent, the 96-core HBv4/HX VM size provides 73% of the performance of the 176-core VM size while requiring only 55% as many software licensed cores.
OpenFOAM – version 2012
Figure 8: On OpenFOAM (Motorbike 28M) HBv4/HX VMs provide more than a 4x performance uplift compared to a 4 year-old HPC server, and more than 2x compared to the most recent Azure HPC VM, HBv3-series.
The absolute values for the benchmark represented in Figure 8 are shared below:
| VM Type |
Average Solver Rating |
|
HBv2 |
1314.267 |
|
HBv3 |
1764.8 |
|
HBv4/HX |
3247.70 |
|
HC |
729.77 |
Table 7: OpenFOAM (Motorbike 28M cells) absolute performance (execution time, lower = better).
Finite Element Analysis (FEA)
Altair RADIOSS – version 2022.1
Figure 9: On Altair Radioss (T10M) HBv4/HX VMs provide more than a 4x performance uplift compared to 4 year-old HPC server, and more than 2x compared to the most recent Azure HPC VM, HBv3-series.
The absolute values for the benchmark represented in Figure 9 are shared below:
| VM Type |
Average Solver Rating |
|
HBv2 |
1314.267 |
|
HBv3 |
1764.8 |
|
HBv4/HX |
3247.70 |
|
HC |
729.77 |
Table 8: Altair Radioss (T10M) absolute performance (execution time, lower = better).
MSC Nastran – version 2022.3
Note: for NASTRAN, the SOL108 medium benchmark was only tested on a HX-series VM because this VM type was created to support such large memory workloads. The larger memory footprint of HX-series (2x that of HBv4-series) allows the benchmark to run completely out of DRAM, which in turn provides additional performance speedup on top of that provided by the newer 4th Gen EPYC CPUs and faster memory subsystem. As such, it would not be accurate to characterize the performance depicted below as “HBv4/HX” and we have instead marked it simply as “HX.”
Figure 10: On MSC NASTRAN (SOL108 Medium) HX-series VMs provide more than a 8x performance uplift compared to 4 year-old HPC server, and more than 5x compared to the most recent Azure HPC VM, HBv3-series.
The absolute values for the benchmark represented in Figure 10 are shared below:
| VM Type |
Average Solver Rating |
|
HBv2 |
1314.267 |
|
HBv3 |
1764.8 |
|
HBv4/HX |
3247.70 |
|
HC |
729.77 |
Table 9: MSC NASTRAN absolute performance (Execution time: lower = better).
Weather Simulation
WRF – version 4.2.2
Figure 11: On WRF (Conus 2.5km) HBv4/HX VMs provide more than a 8x performance uplift compared to a 4 year-old HPC server, and more than 2x compared to the most recent Azure HPC VM, HBv3-series.
The absolute values for the benchmark represented in Figure 11 are shared below:
| VM Type |
Average Solver Rating |
|
HBv2 |
1314.267 |
|
HBv3 |
1764.8 |
|
HBv4/HX |
3247.70 |
|
HC |
729.77 |
Table 10: WRF (Conus 2.5km) absolute performance (time/time-step, lower = better).
Molecular Dynamics
NAMD – version 2.15
Figure 12: On NAMD (Apoa1 100K atoms) HBv4/HX VMs provide more than a 5x performance uplift compared to 4 year-old HPC server, and more than 2x compared to the most recent Azure HPC VM, HBv3-series.
The absolute values for the benchmark represented in Figure 12 are shared below:
|
VM Type |
Average Solver Rating |
|
HBv2 |
1314.267 |
|
HBv3 |
1764.8 |
|
HBv4/HX |
3247.70 |
|
HC |
729.77 |
Table 10: WRF (Conus 2.5km) absolute performance (time/time-step, lower = better).
Rendering
V-Ray – version 5.02.00
Figure 13: On V-Ray 5, HBv4/HX VMs provide more than a 4x performance uplift compared to 4-year-old HPC server, and 1.86x compared to the most recent Azure HPC VM, HBv3-series.
The absolute values for the benchmark represented in Figure 13 are shared below:
|
VM Type |
Average Solver Rating |
|
HBv2 |
1314.267 |
|
HBv3 |
1764.8 |
|
HBv4/HX |
3247.70 |
|
HC |
729.77 |
Table 12: Chaos V-ray 5 absolute performance (frames rendered, higher = better).
Chemistry
CP2K - version 9.1
Figure 14: On CP2K (H2O-DFT-LS), HBv4/HX VMs provide nearly a 5x performance uplift compared to 4-year-old HPC server, and nearly 2.5x compared to the most recent Azure HPC VM, HBv3-series.
The absolute values for the benchmark represented in Figure 14 are shared below:
|
VM Type |
Average Solver Rating |
|
HBv2 |
1314.267 |
|
HBv3 |
1764.8 |
|
HBv4/HX |
3247.70 |
|
HC |
729.77 |
Table 13: CP2K (H2O-DFT-LS) absolute performance (execution time, lower = better).
#AzureHPCAI