NVIDIA H20 vs NVIDIA H200 SXM 141 GB Comparison

Head-to-Head Benchmarks

The recorded database contains no direct head-to-head benchmark entries for the NVIDIA H20 and the NVIDIA H200 SXM 141 GB. Both cards have an average benchmark score of 0 and hold identical percentile rankings against all GPUs at the 50th percentile. This means the two accelerators occupy the same relative performance tier in the overall distribution, but without discrete benchmark scores, a direct performance comparison must be derived entirely from their recorded architectural and specification data.

The most substantial measurable difference between the two units appears in the shading unit count. The H200 SXM 141 GB carries 16,896 shading units, while the H20 has 9,984. That is a 69.2% higher count for the H200, which directly scales into its FP32 compute rating of 66.91 TFLOPS versus 39.54 TFLOPS for the H20. The H200 therefore delivers approximately 69.2% more single-precision floating-point throughput, a decisive margin for any workload that relies on general-purpose CUDA core execution.

Texture processing follows the same pattern. The H200 SXM 141 GB includes 528 texture mapping units, the H20 has 312. The resulting texture rate for the H200 is 1,045.4 GTexel/s, against 617.8 GTexel/s for the H20. Again, the gap is substantial, with the H200 achieving 69.2% higher texture fill performance. Both cards maintain the same 24 ROPs and identical pixel rates of 47.52 GPixel/s, meaning rasterization output is capped equally despite the compute disparity.

Tensor core counts mirror the shading unit distribution. The H200 SXM 141 GB has 528 tensor cores, the H20 has 312. FP16 throughput for the H200 is recorded at 133.8 TFLOPS with a 2:1 ratio, while the H20 achieves 79.07 TFLOPS under the same ratio. The H200 again holds a 69.2% advantage, confirming that the compute scaling is uniform across FP32, FP16, and tensor operations.

Memory capacity is the other major differentiator. The H200 SXM 141 GB carries 141 GB of HBM3e memory, while the H20 has 96 GB of HBM3. Bandwidth also favors the H200 at 4.89 TB/s versus 4.03 TB/s, a 21.3% increase. The memory clock differs as well: the H200 runs at 1593 MHz with 6.4 Gbps effective data rate, the H20 at 1313 MHz with 5.3 Gbps effective. Both use a 6144-bit bus, so the bandwidth gap comes entirely from the higher memory clock and the newer HBM3e standard.

Clock speeds show a partial trade-off. The H20 has a higher base clock at 1830 MHz, while the H200 SXM 141 GB starts at 1500 MHz. The boost clocks are identical at 1980 MHz. This means the H20 runs at a higher base frequency, but under sustained boost conditions, both cards reach the same peak. The H20's higher base clock does not compensate for its lower core count in aggregate throughput, as the recorded FP32 and FP16 figures demonstrate.

Power consumption also differs. The H200 SXM 141 GB is rated at 700 W TDP with a suggested PSU of 1100 W, while the H20 is rated at 500 W TDP with a suggested PSU of 900 W. The H200 uses more power to achieve its higher compute and memory performance. The H20's lower power draw may position it differently for power-constrained deployments, though no efficiency ratio is recorded in the database.

The Verdict

The data indicates two distinct positioning strategies within the same GH100 chip family. The H200 SXM 141 GB is the higher-throughput variant across every compute metric, with 69.2% more shading units, texture units, tensor cores, and FP32/FP16 throughput, plus 46.9% more memory capacity and 21.3% more bandwidth. The H20, conversely, offers a lower base clock advantage of 1830 MHz versus 1500 MHz, but its boost clock is identical at 1980 MHz.

For workloads that scale with raw compute throughput, FP16 tensor operations, or large memory-resident datasets, the H200 SXM 141 GB is the clear choice based on the recorded specifications. The 141 GB memory capacity and 4.89 TB/s bandwidth provide a substantial margin for model sizes and data movement. The H20's 96 GB and 4.03 TB/s remain capable but lower.

For deployments where power draw is a primary constraint, the H20's 500 W TDP versus 700 W TDP represents a 40% reduction in rated power consumption. The H20 also maintains a higher base clock, which may benefit workloads that operate below boost frequencies. No performance-per-watt figures are recorded, so a direct efficiency comparison cannot be made, but the absolute power difference is clear.

The identical 24 ROPs, 47.52 GPixel/s pixel rate, 6144-bit memory bus, and 1980 MHz boost clock indicate that both cards share the same memory interface width and rasterization ceiling. The H200 SXM 141 GB does not improve pixel fill or memory bus width, so workloads limited by those factors will see no difference between the two.

FAQ

Q: Which GPU has more FP32 compute throughput?

A: The NVIDIA H200 SXM 141 GB with 66.91 TFLOPS, compared to 39.54 TFLOPS for the H20.

Q: What is the memory capacity difference?

A: The H200 SXM 141 GB has 141 GB of HBM3e memory, while the H20 has 96 GB of HBM3, a difference of 45 GB.

Q: Do both cards use the same memory bus width?

A: Yes, both use a 6144-bit memory bus, but the H200 achieves higher bandwidth at 4.89 TB/s versus 4.03 TB/s due to faster memory clocks.

Q: Are the boost clocks the same?

A: Yes, both GPUs have a boost clock of 1980 MHz, though the H20 has a higher base clock of 1830 MHz versus 1500 MHz for the H200.

Q: Which GPU has more tensor cores?

A: The H200 SXM 141 GB has 528 tensor cores, while the H20 has 312 tensor cores.

Q: What is the TDP for each card?

A: The H200 SXM 141 GB is rated at 700 W, and the H20 is rated at 500 W.

Specification Differences

The two GPUs share the same GH100 chip, Hopper architecture, 5 nm process node, TSMC foundry, 80,000 million transistors, 814 mm² die size, and 98.3M / mm² transistor density. Both are SXM modules with PCIe 5.0 x16 interfaces, no display outputs, and N/A API support for DirectX, OpenGL, and Vulkan. Both have 24 ROPs, 47.52 GPixel/s pixel rate, and a 6144-bit memory bus. Both are Active in production status, have the same predecessor (Server Ada) and successor (Server Blackwell), and have no recorded launch MSRP.

The differences begin with shading units: 16,896 on the H200 versus 9,984 on the H20. Texture mapping units are 528 versus 312, and tensor cores are 528 versus 312. Texture rate is 1,045.4 GTexel/s versus 617.8 GTexel/s. FP32 is 66.91 TFLOPS versus 39.54 TFLOPS, and FP16 is 133.8 TFLOPS versus 79.07 TFLOPS. Base clocks differ at 1500 MHz for the H200 and 1830 MHz for the H20, with identical boost clocks of 1980 MHz. Memory clock is 1593 MHz with 6.4 Gbps effective for the H200, versus 1313 MHz with 5.3 Gbps effective for the H20. Memory size is 141 GB HBM3e versus 96 GB HBM3. Bandwidth is 4.89 TB/s versus 4.03 TB/s. TDP is 700 W versus 500 W. The H200 has an 8-pin EPS power connector, while the H20 has no recorded power connector. Suggested PSU is 1100 W for the H200 and 900 W for the H20. Release dates differ as well, with the H20 released earlier and the H200 released later.

Architecture Differences

Both GPUs are built on the GH100 chip using the Hopper architecture, fabricated by TSMC on a 5 nm process. The transistor count is identical at 80,000 million, and the die size is the same at 814 mm². The architectural foundation is therefore the same, but the configurations differ significantly.

The H200 SXM 141 GB activates a larger portion of the GH100 die, with 16,896 shading units and 528 tensor cores. The H20 uses a more limited configuration with 9,984 shading units and 312 tensor cores. This suggests different binning or enabled compute block counts within the same physical chip, though the database does not specify the exact mechanism.

Memory architecture differs in both capacity and type. The H200 uses HBM3e, a newer memory standard, while the H20 uses HBM3. Both share the same 6144-bit bus width, but the H200's memory runs at a higher clock, resulting in 4.89 TB/s bandwidth versus 4.03 TB/s. The H200's 141 GB capacity exceeds the H20's 96 GB, which is relevant for models that approach memory limits.

The H200 adds an 8-pin EPS power connector, which the H20 does not list. This aligns with the H200's higher TDP of 700 W versus 500 W. The H200 also requires a suggested PSU of 1100 W versus 900 W for the H20.

Both cards lack ray tracing cores in the recorded data, and both have N/A API support, indicating they are compute-focused accelerators without graphics output. The identical pixel rate of 47.52 GPixel/s confirms that rasterization hardware is unchanged between the two, while the compute and texture pipelines are expanded on the H200.

Where Each One Wins

The H200 SXM 141 GB wins in every compute throughput category recorded. Its 66.91 TFLOPS FP32 and 133.8 TFLOPS FP16 performance are directly tied to its higher shading unit and tensor core counts. Workloads that are compute-bound, such as large matrix operations, dense tensor calculations, or high-throughput FP16 inference, will see the largest benefit from the H200. The 528 tensor cores provide a 69.2% advantage over the H20's 312, which is decisive for AI training and inference tasks that leverage tensor operations.

The H200 also wins in memory capacity and bandwidth. With 141 GB versus 96 GB, the H200 can accommodate larger models or datasets entirely in GPU memory, reducing the need for memory swapping. The 4.89 TB/s bandwidth versus 4.03 TB/s further aids data movement, which is critical for memory-bound workloads that stream large tensors through the compute units.

The H20 wins in base clock speed, at 1830 MHz versus 1500 MHz for the H200. For workloads that run at base clock rather than boost, the H20 executes each instruction faster on a per-cycle basis. The H20 also consumes less power, at 500 W versus 700 W, which matters in power-constrained server environments where multiple accelerators are deployed per chassis. The lower power draw may also reduce cooling requirements, though the database does not record thermal specifications.

The H20 and H200 are evenly matched on pixel rate at 47.52 GPixel/s and ROP count at 24, so any workload limited by rasterization output will see no difference. Similarly, both use PCIe 5.0 x16, so host interface bandwidth is identical.

For memory-bound AI inference with very large models, the H200 SXM 141 GB is the stronger choice due to its 141 GB capacity and 4.89 TB/s bandwidth. For compute-bound training with dense FP16 or FP32 operations, the H200's 69.2% higher throughput wins outright. The H20's advantages are limited to a higher base clock and lower power draw, which may favor scenarios where sustained base-clock operation and power efficiency are prioritized over peak throughput.