NVIDIA A10M vs NVIDIA CMP 90HX Comparison

The NVIDIA A10M wins this matchup decisively. In the one head-to-head benchmark recorded in the database, Geekbench OpenCL, the A10M scored 135230 against the CMP 90HX's 69000, a 96% advantage. Both cards share the same GA102 silicon, built on Samsung's 8 nm process, yet the recorded data shows they land in entirely different performance tiers: the A10M sits in the 96th percentile of all GPUs, while the CMP 90HX sits in the 90th. Same die, same transistor count of 28,300 million, same 628 mm² chip, and a near two-fold gap in measured compute throughput.

Head-to-Head Benchmarks

The database contains a single head-to-head test between these two cards, and it is not close.

Geekbench OpenCL: NVIDIA A10M 135230, NVIDIA CMP 90HX 69000. Winner: A10M, by 96%.

That margin is enormous, and it is consistent with where each card places among its own nearest rivals. The A10M's score of 135230 is statistically indistinguishable from the NVIDIA RTX 4000 Ada Generation, which averages 135218, a 0% delta. It sits just 0.1% behind the AMD Radeon PRO W6800 at 135396, 0.4% behind the AMD Radeon Pro W6800X Duo at 135774, and 0.9% behind the AMD Radeon PRO V620 at 136472. In other words, the A10M trades blows with modern professional workstation hardware from both vendors.

The CMP 90HX's 69000 puts it in a different neighborhood entirely. Its nearest rival is the Intel Arc A770 at 68809, which the CMP 90HX beats by 0.3%. It leads the AMD Radeon Instinct MI25 by 0.6% and trails the AMD Radeon Pro WX 8200 by 1.2% and the NVIDIA Quadro P6000 by 1.4%. These are close, competitive matchups, but against a peer group that the A10M's score nearly doubles.

The headline number: the A10M's Geekbench OpenCL result is 96% higher than the CMP 90HX's. The head-to-head record shows one win for the A10M and zero for the CMP 90HX.

Where Each One Wins

Based strictly on the recorded data, the A10M wins everywhere a score exists. The only measured comparison is Geekbench OpenCL, a general compute benchmark, and the A10M takes it by 96%. There is no test in the database where the CMP 90HX comes out ahead.

That said, the raw specifications suggest where each card's strengths would theoretically lie, even without benchmark confirmation. The CMP 90HX holds a substantial memory bandwidth advantage: 760.3 GB/s versus 500.2 GB/s for the A10M, thanks to GDDR6X running at 19 Gbps effective versus the A10M's GDDR6 at 12.5 Gbps effective. Workloads that stream large amounts of data and saturate bandwidth rather than raw shader throughput are the only place the CMP 90HX's spec sheet argues in its favor.

The A10M counters with more of everything computational: 7168 shading units to 6400, 224 TMUs to 200, 224 tensor cores to 200, and 56 RT cores to 50. Its FP32 output of 23.44 TFLOPS beats the CMP 90HX's 21.89 TFLOPS, and FP16 is identical to FP32 on both cards (1:1 ratio), so the A10M leads there too at 23.44 TFLOPS against 21.89 TFLOPS. Texture fill rate favors the A10M, 366.2 GTexel/s to 342.0 GTexel/s.

Curiously, pixel fill rate goes the other way on paper: the CMP 90HX's higher boost clock of 1710 MHz versus 1635 MHz, combined with the same 80 ROPs on both cards, gives it 136.8 GPixel/s against the A10M's 130.8 GPixel/s. That is a narrow spec-sheet edge with no benchmark in the database to confirm it matters, and it is dwarfed by the measured 96% compute gap.

The A10M also carries 20 GB of memory to the CMP 90HX's 10 GB, doubling capacity for large datasets, models, or scenes. For professional and compute workloads, the recorded data and the spec sheet point the same direction: the A10M is the stronger card in every measured and nearly every theoretical dimension except memory bandwidth and pixel rate.

Architecture Differences

There is very little architectural daylight here, which makes the performance gap more instructive. Both cards use:

• Chip: GA102
• Architecture: Ampere
• Process node: 8 nm
• Foundry: Samsung
• Transistors: 28,300 million
• Die size: 628 mm²
• Transistor density: 45.1M / mm²
• ROPs: 80
• API support: DirectX 12 Ultimate (12_2), OpenGL 4.6, Vulkan 1.4

The divergence is in how each card configures that shared silicon and what generation it belongs to. The A10M is part of the Server Ampere (Axx) generation, with the Tesla Turing line as its predecessor and Server Ada as its successor. The CMP 90HX belongs to NVIDIA's Mining GPUs generation, with no listed predecessor or successor. Its purpose is baked into its lineage: it is a purpose-built mining card, and the database reflects that with a 2021-07-27 release date and an End-of-life status. The A10M is also End-of-life, but it sits in a continuous professional GPU lineage.

Feature counts differ modestly: the A10M enables 768 more shading units, 24 more TMUs, 24 more tensor cores, and 6 more RT cores than the CMP 90HX from the same physical die. Both support the same API feature set on paper, but the CMP 90HX's mining pedigree and severely limited host interface (covered below) make those API listings largely academic.

Specification Differences

The fields where the two cards actually diverge:

• Memory size: A10M 20 GB vs CMP 90HX 10 GB
• Memory type: A10M GDDR6 vs CMP 90HX GDDR6X
• Memory bandwidth: A10M 500.2 GB/s vs CMP 90HX 760.3 GB/s
• Effective memory speed: A10M 12.5 Gbps vs CMP 90HX 19 Gbps
• Base clock: A10M 975 MHz vs CMP 90HX 1500 MHz
• Boost clock: A10M 1635 MHz vs CMP 90HX 1710 MHz
• Shading units: 7168 vs 6400
• TMUs: 224 vs 200
• RT cores: 56 vs 50
• Tensor cores: 224 vs 200
• FP32/FP16: 23.44 TFLOPS vs 21.89 TFLOPS (both 1:1 FP16)
• Pixel rate: 130.8 GPixel/s vs 136.8 GPixel/s
• Texture rate: 366.2 GTexel/s vs 342.0 GTexel/s
• TDP: 150 W vs 320 W
• Slot width: Single-slot vs Dual-slot
• Power connectors: 8-pin EPS vs 2x 8-pin
• Suggested PSU: 450 W vs 700 W
• Bus interface: PCIe 4.0 x16 vs PCIe 1.0 x4
• Length: 267 mm (10.5 inches) vs 285 mm (11.2 inches)
• Release date: A10M not recorded vs CMP 90HX 2021-07-27
• Generation: Server Ampere (Axx) vs Mining GPUs

The bus interface is arguably the most consequential difference after the benchmark scores. The A10M offers PCIe 4.0 x16, a full modern host link. The CMP 90HX is limited to PCIe 1.0 x4, a narrow legacy interface that constrains data exchange with the host system. For any workload that moves data over the bus, not just within the GPU, this is a hard limitation.

Power and physical characteristics also diverge sharply. The A10M runs at 150 W in a single-slot form factor with an 8-pin EPS connector, suitable for dense server deployments with a 450 W suggested PSU. The CMP 90HX draws 320 W, needs two 8-pin connectors and a 700 W suggested PSU, and occupies a dual-slot footprint. It is also longer at 285 mm versus 267 mm. Both cards share the same 112 mm height and neither has any display outputs, reinforcing their headless compute roles.

FAQ

Q: Which card is faster overall?

A: The A10M. It scored 135230 in Geekbench OpenCL versus 69000 for the CMP 90HX, a 96% advantage, and it holds the higher percentile ranking, 96th versus 90th against all GPUs in the database.

Q: Do they use the same chip?

A: Yes. Both are built on the GA102 die, Ampere architecture, Samsung 8 nm process, with 28,300 million transistors and a 628 mm² die. The A10M simply enables more of the die's functional units.

Q: Does the CMP 90HX have any hardware advantage?

A: On paper, yes: memory bandwidth (760.3 GB/s vs 500.2 GB/s from faster GDDR6X), a higher boost clock (1710 MHz vs 1635 MHz), and slightly higher pixel fill rate (136.8 vs 130.8 GPixel/s). None of this translates into a benchmark win in the recorded data.

Q: Which has more memory?

A: The A10M, with 20 GB of GDDR6 versus 10 GB of GDDR6X on the CMP 90HX.

Q: How do their power requirements differ?

A: The A10M has a 150 W TDP, single-slot design, an 8-pin EPS connector, and a 450 W suggested PSU. The CMP 90HX has a 320 W TDP, dual-slot design, two 8-pin connectors, and a 700 W suggested PSU.

Q: Are either of these cards still in production?

A: No. Both are listed as End-of-life. The CMP 90HX was released on 2021-07-27; no release date is recorded for the A10M.

The Verdict

For compute, professional visualization, and any general GPU workload, the data points unambiguously to the A10M. It delivers 96% higher Geekbench OpenCL performance, twice the memory capacity, more shading, tensor, and RT resources, a full PCIe 4.0 x16 interface, and dramatically lower power draw in a single-slot package. It competes with current-generation workstation cards like the RTX 4000 Ada Generation, finishing dead even at a 0% delta.

The CMP 90HX has no benchmark win in the database and only two spec-sheet edges, memory bandwidth and pixel rate, alongside a PCIe 1.0 x4 interface that severely limits host communication. Its measured peer group, the Arc A770, Radeon Instinct MI25, and Quadro P6000, sits an entire tier below the A10M's rivals. Unless a workload is dominated purely by memory bandwidth and nothing else, the recorded data makes the choice clear: the A10M is the stronger performer in every measured dimension.