NVIDIA A10G vs NVIDIA CMP 90HX Comparison

# NVIDIA A10G vs NVIDIA CMP 90HX

The NVIDIA A10G and NVIDIA CMP 90HX are two Ampere-generation GA102 parts with entirely different missions. The A10G is a server-oriented accelerator with 24 GB of GDDR6 memory and a 150 W power envelope, while the CMP 90HX is a mining-focused card with 10 GB of GDDR6X and a 320 W power draw. The database records a single head-to-head benchmark, Geekbench OpenCL, where the A10G scores 158063 against the CMP 90HX's 69000. That is a 129.1% advantage for the A10G, a decisive gap that places the two cards in different performance tiers entirely.

Head-to-Head Benchmarks

The only direct comparison available in the database is the Geekbench OpenCL test, and the result is not close. The A10G delivers 158063 points, while the CMP 90HX manages 69000 points. The delta percentage of 129.1% means the A10G scores more than twice as high as the CMP 90HX in this compute workload. This is a comprehensive win for the A10G, reflecting its larger memory capacity, wider memory bus, and more compute resources.

Context from the nearest rivals underscores how wide this gap is. The A10G's average benchmark score of 151963 places it at the 97th percentile of all GPUs in the database. Its closest measured competitor, the NVIDIA Tesla V100 PCIe 32 GB, scores 150305, which is 1.1% behind the A10G, effectively a statistical tie. The AMD Radeon Pro W6800X scores 160671, putting it 5.4% ahead of the A10G, and the NVIDIA A100 PCIe 40 GB scores 162504, which is 6.5% ahead of the A10G. The AMD Instinct MI100 scores 139035, trailing the A10G by 9.3%. The A10G is solidly competitive with high-end server accelerators.

The CMP 90HX, by contrast, sits at the 90th percentile with an average benchmark score of 69000. Its nearest rivals cluster tightly around that figure: the Intel Arc A770 scores 68809, just 0.3% behind the CMP 90HX, and the AMD Radeon Instinct MI25 scores 68562, 0.6% behind it. On the other side, the AMD Radeon Pro WX 90HX's score is 1.2% ahead of the AMD Radeon Pro WX 8200, and the NVIDIA Quadro P6000 scores 69986, putting the CMP 90HX ahead by 1.4%. The CMP 90HX is essentially at parity with cards like the Arc A770 and Quadro P6000, not with server accelerators.

The implication is clear from the recorded data: the A10G is built for sustained compute throughput, and it crushes the CMP 90HX in OpenCL. The 129.1% delta is not a marginal difference; it is a fundamental separation in compute capability.

FAQ

Q: What is the Geekbench OpenCL score for the NVIDIA A10G?

A: The A10G scores 158063 points in Geekbench OpenCL, with an average benchmark score of 158063 across all recorded tests.

Q: How does the A10G compare to the CMP 90HX in the database's head-to-head benchmark?

A: The A10G wins the only head-to-head test, Geekbench OpenCL, scoring 158063 against the CMP 90HX's 69000, a delta of 129.1%.

Q: What percentile ranking does the CMP 90HX hold relative to all GPUs in the database?

A: The CMP 90HX sits at the 90th percentile of all GPUs, with an average benchmark score of 69000.

Q: Which rival cards are closest to the A10G in average benchmark score?

A: The NVIDIA Tesla V100 PCIe 32 GB is 1.1% behind the A10G, the AMD Radeon Pro W6800X is 5.4% ahead, the NVIDIA A100 PCIe 40 GB is 6.5% ahead, and the AMD Instinct MI100 is 9.3% behind the A10G.

Q: Which cards are closest to the CMP 90HX in average benchmark score?

A: The Intel Arc A770 is 0.3% behind the CMP 90HX, the AMD Radeon Instinct MI25 is 0.6% behind, the AMD Radeon Pro WX 8200 is 1.2% ahead, and the NVIDIA Quadro P6000 is 1.4% ahead.

Q: Does the CMP 90HX have any benchmark wins over the A10G?

A: No, the database records zero wins for the CMP 90HX and one win for the A10G in head-to-head testing.

The Verdict

The data positions these two cards in different leagues. The A10G is the clear choice for compute-heavy server workloads. Its 158063 OpenCL score places it 129.1% ahead of the CMP 90HX, and its 97th percentile ranking among all GPUs indicates that it pairs well against top-tier accelerators like the A100 PCIe 40 GB and Radeon Pro W6800X. The A10G's 24 GB of memory and 600.2 GB/s of bandwidth make it suitable for large datasets and memory-intensive tasks.

The CMP 90HX, with its 69000 OpenCL score and 90th percentile ranking, is not in the same compute tier. It trades at the level of the Intel Arc A770 and Quadro P6000, both of which are within 1.4% of its average score. For applications that rely on OpenCL compute, the CMP 90HX is outmatched. Its strengths lie elsewhere, though the database records no benchmark that reflects a typical mining workload.

The verdict from the recorded data is straightforward: the A10G is the superior compute accelerator. The CMP 90HX should only be considered if its specific characteristics, such as its 760.3 GB/s memory bandwidth from GDDR6X, are relevant to a particular use case. For general compute, the A10G wins decisively.

Specification Differences

The two cards differ across nearly every major specification. Memory capacity is the first major split: the A10G has 24 GB of GDDR6, while the CMP 90HX has 10 GB of GDDR6X. The bus width also differs, with the A10G using a 384-bit interface and the CMP 90HX a 320-bit interface. Memory bandwidth favors the CMP 90HX at 760.3 GB/s versus the A10G's 600.3 GB/s, a consequence of its faster 19 Gbps effective memory clock.

Compute units are heavily skewed toward the A10G. It has 9216 shading units, 288 TMUs, 96 ROPs, 72 RT cores, and 72 RT cores, and 288 tensor cores. The CMP 90HX has 6400 shading units, 200 TMUs, 80 ROPs, 50 RT cores, and 200 tensor cores. The A10G's FP32 throughput is 31.52 TFLOPS, while the CMP 90HX achieves 21.89 TFLOPS. The A10G's pixel rate is 164.2 GPixel/s and its texture rate is 492.5 GTexel/s; the CMP 90HX delivers 136.8 GPixel/s and 342.0 GTexel/s.

Power and physical specs also diverge. The A10G has a TDP of 150 W, is single-slot, uses an 8-pin EPS connector, and suggests a 450 W PSU. The CMP 90HX has a TDP of 320 W, is dual-slot, requires two 8-pin connectors, and suggests a 700 W PSU. The A10G is 267 mm long and 112 mm tall; the CMP 90HX is 285 mm long and 112 mm high. The A10G uses PCIe 4.0 x16, the CMP 90HX PCIe 4.0 x4, a notable difference in bus interface bandwidth for card-to-system communication.

The A10G released in April 2021, while the CMP 90HX arrived later in July 2021. The A10G lists its predecessor as Tesla Turing, while the CMP 90HX has no recorded predecessor or successor. The CMP 90HX's memory clock is 1188 MHz with 19 Gbps effective, while the A10G's memory clock is 1563 MHz with 12.3 Gbps effective, numbers that define their respective memory performance.

The CMP 90HX has a boost clock of 1710 MHz, which matches the boost clock of 1710 MHz, but its base clock of 1500 MHz is higher than the A10G's base clock of 1320 MHz. The A10G's memory clock of 1563 MHz is higher than the CMP 90HX's 1188 MHz, but the CMP 90HX's 19 Gbps effective rate is faster than the A10G's 12.3 Gbps effective, reflecting its GDDR6X design.

Both cards share the same GA102 chip, with 28,300 million transistors on a 628 mm² die, and both are built on an 8 nm Samsung process. Both have no display outputs. The A10G's average benchmark score is 151963, the CMP 90HX's is 69000. The A10G is at the 97th percentile, the CMP 90HX at the 90th.

Architecture Differences

The architecture is identical in foundation: both use the Ampere GA102 chip, fabricated on Samsung's 8 nm process with 28,300 million transistors on a 628 mm² die. Transistor density is 45.1M per mm² for both. The difference is in how NVIDIA configured the silicon.

The A10G enables more of the GA102 chip. It has 9216 shading units, 288 TMUs, 96 ROPs, 72 RT cores, and 288 tensor cores. The CMP 90HX disables portions of the chip, leaving 6400 shading units, 200 TMUs, 80 ROPs, 50 RT cores, and 200 tensor cores. This is a substantial reduction in compute capacity; the A10G has roughly 44% more shading units and 44% more tensor cores than the CMP 90HX, though the exact numbers are not directly stated as percentages in the database.

The memory architecture also differs. The A10G uses GDDR6 with a 384-bit bus and 600.3 GB/s bandwidth. The CMP 90HX uses GDDR6X with a 320-bit bus and 760.3 GB/s bandwidth. The CMP 90HX's memory is faster in raw bandwidth despite the narrower bus, due to its higher 19 Gbps effective clock versus the A10G's 12.3 Gbps effective clock. The A10G compensates with more capacity: 24 GB versus 10 GB.

The TDP gap reflects the intended use. The A10G is a 150 W card, optimized for server density and low power consumption. The CMP 90HX is a 320 W card, built for mining where power efficiency is secondary. The A10G uses a single-slot design and an 8-pin EPS connector; the CMP 90HX needs a dual-slot cooler and two 8-pin connectors. The A10G's suggested PSU is 450 W, while the CMP 90HX suggests 700 W.

The bus interface is another architectural divergence. The A10G uses PCIe 4.0 x16, providing a full 16-lane connection for host communication. The CMP 90HX uses PCIe 1.0 x4, a crippled interface that limits data transfer speeds. This makes sense for a mining card, which mostly performs local computation. For the A10G, the wide bus is essential for server workloads that stream data to and from the host.

Feature-wise, both cards support DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. Neither has display outputs, so neither is intended for direct visual output. The A10G's API support and compute resources make it a general-purpose accelerator. The CMP 90HX's API support exists but is less relevant for a card with no display output.

The release timeline shows the A10G launched in April 2021, with the CMP 90HX following in July 2021. Both cards are now end-of-life. The A10G has a successor listed as Server Ada, while the CMP 90HX has no recorded predecessor or successor, marking it as a one-off product line.

The recorded data makes the architectural intent clear: the A10G is a full-fidelity Ampere server chip with balanced memory capacity and bandwidth, while the CMP 90HX is a cut-down variant with faster GDDR6X memory but far fewer compute resources and a crippled PCIe interface. These are not interchangeable parts. They serve different workloads, and the benchmark results reflect that separation.