AMD Instinct MI100 vs NVIDIA CMP 90HX Comparison

FAQ

Q: Which GPU scores higher in the Geekbench OpenCL benchmark?

A: The AMD Instinct MI100 records a score of 139035, while the NVIDIA CMP 90HX scores 69000. The MI100 wins this benchmark by a margin of 101.5%.

Q: How does the MI100 compare to its closest rivals in OpenCL performance?

A: The MI100 sits 0.7% above the NVIDIA Tesla V100 PCIe 16 GB (score 138063) and 0.9% above the Tesla V100 SXM2 32 GB (score 137731). It also leads the AMD Radeon PRO V620 by 1.9% and the Radeon Pro W6800X Duo by 2.4%.

Q: Where does the CMP 90HX rank relative to its own competitor set?

A: The CMP 90HX is 0.3% ahead of the Intel Arc A770 (score 68809) and 0.6% ahead of the AMD Radeon Instinct MI25 (score 68562). However, it trails the AMD Radeon Pro WX 8200 by 1.2% and the NVIDIA Quadro P6000 by 1.4%.

Q: What is the percentile ranking for each card in the full GPU database?

A: The MI100 places in the 96th percentile of all GPUs, while the CMP 90HX lands in the 90th percentile. The MI100's overall standing is notably higher despite both cards being end-of-life products.

Q: Do both cards have display outputs?

A: No. Neither the AMD Instinct MI100 nor the NVIDIA CMP 90HX has any display outputs. Both are compute or mining oriented cards with no video connections.

Q: What PCIe interface does each card use?

A: The MI100 uses PCIe 4.0 x16, while the CMP 90HX uses an older PCIe 1.0 x4 interface. This difference can affect data transfer rates in systems that rely on PCIe bandwidth.

Architecture Differences

The AMD Instinct MI100 is built on the CDNA 1.0 architecture with the Arcturus chip, fabricated on a 7 nm process at TSMC. It packs 25,600 million transistors onto a 750 mm² die, yielding a transistor density of 34.1 million per mm². The NVIDIA CMP 90HX uses the Ampere architecture with the GA102 chip, manufactured on an 8 nm node at Samsung. It contains 28,300 million transistors on a 628 mm² die, achieving a higher density of 45.1 million per mm². The MI100 has the larger die, but the CMP 90HX packs more transistors into a smaller area.

The MI100 is a compute-focused accelerator with 7680 shading units, 480 texture mapping units, and 64 raster output units. It has no dedicated ray tracing cores or tensor cores listed in the database. The CMP 90HX, by contrast, includes 6400 shading units, 200 TMUs, and 80 ROPs, but also features 50 ray tracing cores and 200 tensor cores. This makes the CMP 90HX a more feature-rich GPU in terms of specialized hardware, despite having fewer raw shading units.

Memory architecture differs substantially. The MI100 uses 32 GB of HBM2 on a 4096-bit bus, delivering 1.23 TB/s of bandwidth. The CMP 90HX has 10 GB of GDDR6X on a 320-bit bus with 760.3 GB/s. The MI100's memory system offers roughly 62% more bandwidth, which is typical for a data center oriented part. The CMP 90HX's smaller memory pool and narrower bus reflect its mining-oriented design.

Clock behavior also differs. The MI100 runs at a 1000 MHz base clock with a 1502 MHz boost. The CMP 90HX starts higher at 1500 MHz base and boosts to 1710 MHz. Memory clocks are similar in ratio: the MI100's HBM2 runs at 1200 MHz (2.4 Gbps effective), while the CMP 90HX's GDDR6X runs at 1188 MHz (19 Gbps effective). The GDDR6X uses a much higher effective data rate per pin.

API support is another clear differentiator. The MI100 has no DirectX, OpenGL, or Vulkan support listed. The CMP 90HX supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. This means the MI100 is strictly a compute device, whereas the CMP 90HX retains graphics API compatibility even though it has no display outputs.

Head-to-Head Benchmarks

The only recorded benchmark comparison in the database is the Geekbench OpenCL test. Here, the AMD Instinct MI100 scores 139035 against the NVIDIA CMP 90HX's 69000. The delta is 101.5%, meaning the MI100 more than doubles the CMP 90HX's score. This is a decisive win for AMD in raw compute throughput as measured by OpenCL.

Looking at the rival sets reinforces this gap. The MI100's nearest rivals all score within 2.4% of its result, with the Tesla V100 variants at 138063 and 137731, the Radeon PRO V620 at 136472, and the Radeon Pro W6800X Duo at 135774. The CMP 90HX, meanwhile, sits in a much lower performance band, with rivals clustered around 68562 to 69986. The CMP 90HX's score is closer to the Intel Arc A770 (68809) and the Radeon Instinct MI25 (68562) than to any high-end compute card.

The performance gap in the head-to-head test is consistent with the architectural positioning. The MI100's 32 GB HBM2 memory and 4096-bit bus provide enormous bandwidth that is well suited to compute workloads. The CMP 90HX, with its 10 GB GDDR6X and 320-bit bus, is not designed for the same class of compute tasks. Even though the CMP 90HX has higher clock speeds and more specialized tensor and ray tracing cores, those features do not translate into OpenCL compute advantage over the MI100.

The percentile data also illustrates the separation. The MI100 sits in the 96th percentile of all GPUs, while the CMP 90HX sits in the 90th. That six-percentile gap may sound modest, but the raw score difference of 70,035 points is substantial. In practical terms, the MI100 is operating in a different performance tier altogether.

Specification Differences

| Specification | AMD Instinct MI100 | NVIDIA CMP 90HX |

| --- | --- | --- |

| Process Node | 7 nm | 8 nm |

| Foundry | TSMC | Samsung |

| Transistors | 25,600 million | 28,300 million |

| Die Size | 750 mm² | 628 mm² |

| Base Clock | 1000 MHz | 1500 MHz |

| Boost Clock | 1502 MHz | 1710 MHz |

| Memory Size | 32 GB | 10 GB |

| Memory Type | HBM2 | GDDR6X |

| Memory Bus Width | 4096 bit | 320 bit |

| Memory Bandwidth | 1.23 TB/s | 760.3 GB/s |

| Shading Units | 7680 | 6400 |

| TMUs | 480 | 200 |

| ROPs | 64 | 80 |

| RT Cores | None | 50 |

| Tensor Cores | None | 200 |

| FP32 Performance | 23.07 TFLOPS | 21.89 TFLOPS |

| FP16 Performance | 46.14 TFLOPS (2:1) | 21.89 TFLOPS (1:1) |

| TDP | 300 W | 320 W |

| Bus Interface | PCIe 4.0 x16 | PCIe 1.0 x4 |

| DirectX Support | N/A | 12 Ultimate (12_2) |

| OpenGL Support | N/A | 4.6 |

| Vulkan Support | N/A | 1.4 |

| Release Date | November 2020 | July 2021 |

The MI100 leads in memory capacity, bandwidth, shading units, texture units, and FP32 throughput. The CMP 90HX leads in clock speeds, ROP count, and includes ray tracing and tensor cores that the MI100 lacks. The MI100 also has a lower TDP of 300 W versus 320 W, and both cards require a 700 W suggested power supply with dual 8-pin connectors. Both are dual-slot cards with no display outputs. Physical dimensions are similar: the MI100 is 267 mm long and 111 mm tall, while the CMP 90HX is 285 mm long and 112 mm tall.

The Verdict

The data points to a clear split in intended use. The AMD Instinct MI100 is a high-bandwidth compute accelerator designed for data center workloads. Its 101.5% OpenCL advantage over the CMP 90HX, combined with its 96th percentile ranking, makes it the stronger choice for any application that relies on OpenCL compute performance. The 32 GB HBM2 memory and 1.23 TB/s bandwidth provide the kind of memory subsystem that large compute jobs require.

The NVIDIA CMP 90HX, on the other hand, is a mining GPU with a lower performance ceiling. Its 90th percentile ranking and 69000 OpenCL score place it in a different class entirely. However, it does have API support for DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4, which the MI100 completely lacks. This makes the CMP 90HX more versatile for tasks that need graphics API compatibility, even though it has no display outputs.

For compute-focused buyers, the MI100 is the obvious pick from the benchmark data. For anyone needing API support or specialized tensor and ray tracing hardware, the CMP 90HX offers capabilities the MI100 does not. The MI100's higher FP16 throughput (46.14 TFLOPS versus 21.89 TFLOPS) also matters for workloads that can use reduced precision. The CMP 90HX's FP16 is at parity with FP32, which is less flexible for mixed-precision compute.

Where Each One Wins

The AMD Instinct MI100 wins in raw OpenCL performance, memory bandwidth, memory capacity, and FP32 compute throughput. Its 23.07 TFLOPS FP32 output is 5.4% higher than the CMP 90HX's 21.89 TFLOPS. The MI100 also doubles the FP16 throughput at 46.14 TFLOPS versus 21.89 TFLOPS, making it better suited for workloads that leverage half-precision arithmetic. Its 1.23 TB/s memory bandwidth is a massive advantage for data movement heavy tasks, and the 4096-bit HBM2 interface is designed for sustained memory access patterns typical in scientific computing and machine learning training.

The NVIDIA CMP 90HX wins in clock speeds, with a 1500 MHz base and 1710 MHz boost compared to the MI100's 1000 MHz base and 1502 MHz boost. It also has more ROPs (80 versus 64), giving it a higher pixel rate of 136.8 GPixel/s compared to 96.13 GPixel/s. The CMP 90HX also includes 50 ray tracing cores and 200 tensor cores, which the MI100 lacks entirely. For workloads that use these specialized units, the CMP 90HX is the only option of the two. Its API support for DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4 adds further flexibility for software stacks that require graphics interfaces.

In terms of physical and electrical design, the CMP 90HX uses a slightly older PCIe 1.0 x4 interface, which is a notable bottleneck for data transfer to the host system. The MI100's PCIe 4.0 x16 interface is far more capable for moving data in and out of the card. The CMP 90HX is also slightly larger at 285 mm versus 267 mm for the MI100, and it draws 20 W more power at 320 W versus 300 W. Both cards share the same dual-slot footprint, dual 8-pin power requirement, and 700 W suggested power supply.

The benchmark results indicate that the MI100 is the superior compute card, while the CMP 90HX is a niche product for specific mining or API-dependent tasks. The recorded data shows no benchmark where the CMP 90HX wins, and the only head-to-head test is a 101.5% victory for the MI100. That said, the CMP 90HX's tensor and ray tracing hardware, along with its graphics API support, give it a functional edge in scenarios where those features are required. For general compute, the MI100 is the clear winner.