AMD Radeon PRO V620 vs NVIDIA A100 SXM4 40 GB Comparison
AMD Radeon PRO V620
A100 SXM4 40 GB
PERFORMANCE BENCHMARKS
Analysis: AMD Radeon PRO V620 vs NVIDIA A100 SXM4 40 GB
The benchmark data is unequivocal: the NVIDIA A100 SXM4 40 GB outperforms the AMD Radeon PRO V620 in every recorded test, with a commanding lead in compute workloads and a substantial edge in graphics API performance. Across the two head-to-head comparisons, the A100 secures 2 wins to 0, with an average benchmark score of 187,147 versus 136,472 for the Radeon PRO V620 — a gap that places the NVIDIA part in the 98th percentile of all GPUs, while the AMD card sits in the 96th.
Head-to-Head Benchmarks
The most decisive margin comes in Geekbench OpenCL, where the A100 SXM4 40 GB scores 201,096 against the Radeon PRO V620’s 128,580. That is a 56.4% advantage, the single largest delta between the two cards in any test. This result aligns with the architectural focus of the A100: its 19.49 TFLOPS FP32 throughput and 77.97 TFLOPS FP16 (4:1) performance are paired with 432 tensor cores, making it a compute-first accelerator. The Radeon PRO V620, by contrast, delivers 20.28 TFLOPS FP32 and 40.55 TFLOPS FP16 (2:1) — its FP32 figure is actually slightly higher than the A100’s, yet its OpenCL score trails by more than half, suggesting that raw shader count alone does not translate into compute benchmark dominance. The A100’s 54,200 million transistors on an 826 mm² die, versus 26,800 million on 520 mm² for the AMD chip, points to a far more complex memory hierarchy and specialized hardware that the OpenCL workload clearly exploits.
In Geekbench Vulkan, the margin narrows considerably but still favors NVIDIA. The A100 scores 173,198 versus 144,364 for the Radeon PRO V620, a 20% lead. This is notable because Vulkan is a lower-level graphics API where AMD’s RDNA 2.0 architecture typically performs well, and the Radeon PRO V620 does have 72 ray accelerators and DirectX 12 Ultimate support, which the A100 lacks entirely (its DirectX field is null). Yet the A100’s 160 ROPs and 432 TMUs, combined with a 5120-bit memory bus delivering 1.56 TB/s of bandwidth, overcome the AMD card’s 128 ROPs and 288 TMUs. The Radeon PRO V620’s memory bandwidth of 512.0 GB/s is exactly one-third of the A100’s, and that disparity likely explains why the NVIDIA card still wins even in a graphics-centric API test.
Contextualizing these scores against nearest rivals reinforces the A100’s position. The A100’s average of 187,147 is 1.3% ahead of the NVIDIA RTX 5000 Ada Generation (184,664) and 1.9% ahead of the A100 SXM4 80 GB (183,725). It trails only the Tesla V100S PCIe 32 GB (194,415) by 3.7%. The Radeon PRO V620’s 136,472 average is nearly identical to its closest competitors — 0.5% above the Radeon Pro W6800X Duo (135,774), 0.8% above the Radeon PRO W6800 (135,396), and 0.9% above both the NVIDIA A10M (135,230) and RTX 4000 Ada Generation (135,218). In other words, the A100 is competing at the top tier of server accelerators, while the Radeon PRO V620 sits in a crowded mid-range cluster.
Where Each One Wins
The A100 SXM4 40 GB wins every benchmark in this comparison, so the "where each one wins" split is asymmetric. In OpenCL, the A100’s 56.4% lead makes it the clear choice for general-purpose compute, machine learning training, and scientific simulation workloads that rely on FP16 or FP32 throughput with tensor core acceleration. Its 40 GB of HBM2e memory with 1.56 TB/s bandwidth is a massive advantage for large datasets that must reside on the GPU — the Radeon PRO V620’s 32 GB of GDDR6 at 512.0 GB/s cannot keep pace in memory-bound tasks.
For Vulkan workloads, the A100 still wins, but the 20% margin suggests the Radeon PRO V620 is more competitive in real-time graphics rendering, particularly given its ray tracing capabilities (72 RT cores) and full DirectX 12 Ultimate support. The Radeon PRO V620 also has higher peak clocks — a 2200 MHz boost versus 1410 MHz for the A100 — and a higher pixel rate (281.6 GPixel/s versus 225.6 GPixel/s). If a workload is purely rasterization-heavy and relies on clock speed rather than memory bandwidth, the AMD card narrows the gap. However, the A100’s texture rate (609.1 GTexel/s versus 633.6 GTexel/s) is actually slightly lower, yet it still wins in Vulkan, indicating that the NVIDIA driver stack and memory subsystem compensate for any raw throughput deficit.
The Radeon PRO V620 does have one practical advantage: it is a dual-slot PCIe card with two 8-pin power connectors and a 300 W TDP, whereas the A100 is an SXM module with no power connectors and a 400 W TDP. This means the AMD card can be installed in standard server chassis without specialized power delivery, while the A100 requires an SXM-compatible motherboard. The Radeon PRO V620 also has physical dimensions (267 mm length, 120 mm height, 50 mm width) that make it a drop-in replacement for many existing GPU slots, while the A100’s slot width is listed as "SXM Module" with no standard PCIe form factor.
The Verdict
The data supports only one conclusion for compute performance: the NVIDIA A100 SXM4 40 GB is the superior accelerator. With a 56.4% OpenCL lead and a 20% Vulkan lead, it wins outright in both categories, and its 98th percentile ranking versus the Radeon PRO V620’s 96th confirms its higher standing. Anyone running FP16 training workloads, large-scale inference, or HPC simulations should choose the A100 without hesitation — the 40 GB HBM2e pool and tensor cores are simply in a different class.
The Radeon PRO V620 is not a bad card; it is a capable RDNA 2.0 part with 32 GB of memory and full graphics API support. But its benchmark results place it in the same performance tier as the RTX 4000 Ada Generation and Radeon PRO W6800, not the A100’s tier. If a system cannot accommodate an SXM module, or if the workload is primarily Vulkan-based rendering where the 20% deficit is acceptable, the Radeon PRO V620 becomes a reasonable fallback. However, for any task where raw compute throughput matters, the A100’s 187,147 average score versus 136,472 is a 37% overall advantage that no amount of clock speed or ray tracing hardware can overcome.
FAQ
Q: Which GPU has a higher average benchmark score?
A: The NVIDIA A100 SXM4 40 GB has an average score of 187,147, while the AMD Radeon PRO V620 averages 136,472 — a difference of 50,675 points in favor of NVIDIA.
Q: How large is the OpenCL performance gap?
A: The A100 scores 201,096 in Geekbench OpenCL versus 128,580 for the Radeon PRO V620, giving NVIDIA a 56.4% lead.
Q: Does the AMD card win in any benchmark?
A: No. The head-to-head results show the A100 winning both Geekbench OpenCL and Geekbench Vulkan, with the Radeon PRO V620 recording zero wins.
Q: What memory configurations do the two cards use?
A: The A100 uses 40 GB of HBM2e with a 5120-bit bus and 1.56 TB/s bandwidth. The Radeon PRO V620 uses 32 GB of GDDR6 with a 256-bit bus and 512.0 GB/s bandwidth.
Q: Which card has more shading units?
A: The A100 has 6912 shading units, while the Radeon PRO V620 has 4608. The AMD card does have 72 ray accelerators, which the A100 lacks.
Q: Are both cards still in production?
A: No. Both are marked as end-of-life, with the A100 released on 2020-05-13 and the Radeon PRO V620 released on 2021-11-03.
Architecture Differences
The NVIDIA A100 is built on the Ampere architecture using the GA100 chip, fabricated on a 7 nm process at TSMC. It packs 54,200 million transistors into an 826 mm² die, yielding a transistor density of 65.6 million per square millimeter. The Radeon PRO V620 uses AMD’s RDNA 2.0 architecture with the Navi 21 chip, also on 7 nm TSMC, but with 26,800 million transistors on a 520 mm² die — a density of 51.5 million per square millimeter. This means the A100 has roughly double the transistor count and a 37% larger die, which explains its superior memory bandwidth and tensor core capabilities.
The A100 features 432 tensor cores, which the Radeon PRO V620 entirely lacks, while the AMD card counters with 72 ray accelerators that the A100 does not have. Both support PCIe 4.0 x16, but the NVIDIA part is an SXM module requiring specialized mounting, whereas the AMD card is a standard dual-slot PCIe design. The A100 has no display outputs, and neither does the Radeon PRO V620. For APIs, the A100 lists null values for DirectX, OpenGL, and Vulkan, while the Radeon PRO V620 supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4.
Memory architecture differs fundamentally: the A100 uses HBM2e with a 5120-bit interface and 1.56 TB/s bandwidth, while the Radeon PRO V620 uses GDDR6 with a 256-bit interface and 512.0 GB/s bandwidth. The A100’s memory clock is 1215 MHz (2.4 Gbps effective), whereas the AMD card runs at 2000 MHz (16 Gbps effective) — the GDDR6 transfers faster per pin, but the HBM2e’s massive bus width gives NVIDIA a 3x bandwidth advantage.
Specification Differences
| Specification | NVIDIA A100 SXM4 40 GB | AMD Radeon PRO V620 |
|---|---|---|
| Chip | GA100 | Navi 21 |
| Architecture | Ampere | RDNA 2.0 |
| Transistors | 54,200 million | 26,800 million |
| Die Size | 826 mm² | 520 mm² |
| Base Clock | 1095 MHz | 1825 MHz |
| Boost Clock | 1410 MHz | 2200 MHz |
| Memory Size | 40 GB | 32 GB |
| Memory Type | HBM2e | GDDR6 |
| Memory Bus | 5120 bit | 256 bit |
| Memory Bandwidth | 1.56 TB/s | 512.0 GB/s |
| Shading Units | 6912 | 4608 |
| TMUs | 432 | 288 |
| ROPs | 160 | 128 |
| Tensor Cores | 432 | None |
| Ray Accelerators | None | 72 |
| FP32 Performance | 19.49 TFLOPS | 20.28 TFLOPS |
| FP16 Performance | 77.97 TFLOPS (4:1) | 40.55 TFLOPS (2:1) |
| Pixel Rate | 225.6 GPixel/s | 281.6 GPixel/s |
| Texture Rate | 609.1 GTexel/s | 633.6 GTexel/s |
| TDP | 400 W | 300 W |
| Slot Width | SXM Module | Dual-slot |
| Power Connectors | None | 2x 8-pin |
| Suggested PSU | 800 W | 700 W |
| Release Date | 2020-05-13 | 2021-11-03 |
The A100 leads in transistor count, die size, memory size, memory bus width, memory bandwidth, shading units, TMUs, ROPs, tensor cores, and FP16 performance. The Radeon PRO V620 counters with higher clocks (1825 MHz base, 2200 MHz boost), slightly higher FP32 throughput (20.28 versus 19.49 TFLOPS), higher pixel and texture rates, ray accelerators, and a lower 300 W TDP. Both share the same 7 nm TSMC process, PCIe 4.0 x16 interface, and end-of-life production status. The A100’s 40 GB of HBM2e at 1.56 TB/s dwarfs the 32 GB of GDDR6 at 512.0 GB/s, making memory-bound workloads decisively NVIDIA’s territory.