NVIDIA B200 vs NVIDIA Tesla V100 SXM2 32 GB Comparison

NVIDIA B200 and NVIDIA Tesla V100 SXM2 32 GB represent two distinct generations of NVIDIA’s server-grade compute efforts. The data shows a decisive performance gulf, but the comparison is more nuanced than a simple score difference. The B200, built on the Blackwell architecture, is a current-generation part with a 100th-percentile ranking in the database, while the V100 SXM2 32 GB, a Volta-era product, remains a capable legacy option with a 96th-percentile score.

Head-to-Head Benchmarks

The sole head-to-head benchmark available is Geekbench OpenCL, where the B200 delivers a score of 345,482 against the V100 SXM2 32 GB’s 131,696. This represents a delta of 162.3% in favor of the B200. That is not a marginal improvement; it is a 2.6x jump in raw compute throughput as measured by this workload. The B200’s score places it at 100% of all GPUs in the database, meaning it sits at the absolute top of the distribution. The V100 SXM2 32 GB, with its 131,696 points, holds a 96th-percentile position, which is a strong result for a part of its vintage but a full 4 percentage points below the B200’s ceiling.

Looking at the B200’s nearest rivals provides context for its lead over the older card. The B200 outperforms the NVIDIA H200 NVL by 3.2%, a relatively slim margin that suggests the B200 is not a generational outlier but rather a refinement of the Hopper-class performance envelope. Against the AMD Instinct MI300X, the B200 is 8.6% faster, and it beats the NVIDIA L40S by a more substantial 16.8%. However, the B200 trails the NVIDIA B300 SXM6 AC by 6.6%, indicating that within the Blackwell family itself, the B200 is not the absolute peak. For the V100 SXM2 32 GB, its nearest rivals show a much tighter cluster: it is 0.2% behind the NVIDIA Tesla V100 PCIe 16 GB, 0.9% ahead of the AMD Radeon PRO V620, 0.9% behind the AMD Instinct MI100, and 1.4% ahead of the AMD Radeon Pro W6800X Duo. These sub-2% deltas confirm that the V100 SXM2 32 GB is operating at the performance level of its immediate contemporaries, and the 162.3% gap to the B200 is a reflection of architectural progress rather than a defect in the older card.

Architecture Differences

The B200 is built on TSMC’s 5 nm process, packing 104,000 million transistors into a GB100 chip. The V100 SXM2 32 GB uses a 12 nm process with a GV100 die containing 21,100 million transistors. The transistor count difference is stark—nearly a 5x increase—and the die size for the V100 SXM2 32 GB is 815 mm², with a transistor density of 25.9 million per mm². No die size or density figure is available for the B200, but the shift to 5 nm is a primary driver of its performance advantage. The B200’s base clock is 700 MHz with a boost of 1965 MHz, while the V100 SXM2 32 GB runs at 1290 MHz base and 1530 MHz boost. The B200 has a lower base clock but a significantly higher boost ceiling, allowing it to scale up under load.

Memory configurations diverge sharply. The B200 features 90 GB of HBM3e on a 4096-bit bus, delivering 4.10 TB/s of bandwidth. The V100 SXM2 32 GB offers 32 GB of HBM2 on the same 4096-bit bus width, but bandwidth is limited to 898.0 GB/s. The B200’s memory bandwidth is over 4.5x higher, which is critical for data-intensive workloads. The B200’s memory clock is listed as 2000 MHz with 8 Gbps effective, whereas the V100 SXM2 32 GB runs at 877 MHz with 1754 Mbps effective. The B200’s shading unit count is 18,944 versus 5,120 for the V100 SXM2 32 GB, and texture mapping units are 592 versus 320. Raster output units differ with the B200 at 24 and the V100 SXM2 32 GB at 128, which is an unusual inversion—the older card has more ROPs, but the B200’s pixel rate is 47.16 GPixel/s versus 195.8 GPixel/s for the V100 SXM2 32 GB, indicating the V100’s ROP advantage does not translate to higher pixel throughput in this comparison.

Compute capabilities show a clear hierarchy. The B200’s FP32 throughput is 74.45 TFLOPS, compared to 15.67 TFLOPS for the V100 SXM2 32 GB. FP16 performance is where the B200 truly separates itself: 1,191.2 TFLOPS with a 16:1 ratio, versus 31.33 TFLOPS with a 2:1 ratio for the V100 SXM2 32 GB. Tensor core counts are 592 for the B200 and 640 for the V100 SXM2 32 GB, but the B200’s newer tensor core design and higher clock speeds yield far greater FP16 output. The B200 has a TDP of 1000 W with a suggested PSU of 1400 W, while the V100 SXM2 32 GB draws 250 W with a 600 W suggested PSU. The B200 uses PCIe 5.0 x16, whereas the V100 SXM2 32 GB is limited to PCIe 3.0 x16. Neither card has display outputs, and both use SXM modules. The B200’s production status is Active, while the V100 SXM2 32 GB is End-of-life, having been released on 2018-03-26.

The Verdict

The data indicates that the NVIDIA B200 is the superior compute device in every measured category. Its Geekbench OpenCL score of 345,482 is more than double the V100 SXM2 32 GB’s 131,696, and its 100th-percentile ranking against all GPUs means no other part in the database matches its average benchmark score. The B200’s nearest rival, the B300 SXM6 AC, is 6.6% faster, so the B200 is not the absolute fastest Blackwell part, but it is the one with the highest percentile in this dataset. The V100 SXM2 32 GB, while holding a strong 96th-percentile position, is 162.3% behind the B200 in the head-to-head test. For users choosing strictly on raw performance, the B200 is the clear selection.

The V100 SXM2 32 GB should not be dismissed. Its 250 W TDP is a fraction of the B200’s 1000 W, and its 32 GB of HBM2 memory is sufficient for many legacy workloads. The V100 SXM2 32 GB’s performance cluster, within 1.4% of its nearest rivals, suggests it remains competitive within its own generation. The B200’s advantages—more memory, higher bandwidth, larger shading unit count, and broader FP16 capability—are all measurable in the benchmark data. The production status difference is also telling: the B200 is Active, while the V100 SXM2 32 GB is End-of-life, which affects long-term availability and driver support. For new deployments, the B200 is the data-backed choice. For maintaining existing Volta-based infrastructure, the V100 SXM2 32 GB remains a functional, if dated, option.

FAQ

Q: How much faster is the NVIDIA B200 than the NVIDIA Tesla V100 SXM2 32 GB in Geekbench OpenCL?

A: The B200 scores 345,482, while the V100 SXM2 32 GB scores 131,696, yielding a delta of 162.3% in favor of the B200.

Q: What is the memory capacity and type for each card?

A: The B200 has 90 GB of HBM3e with 4.10 TB/s bandwidth, while the V100 SXM2 32 GB has 32 GB of HBM2 with 898.0 GB/s bandwidth.

Q: Which card has a higher FP32 throughput?

A: The B200 delivers 74.45 TFLOPS of FP32, compared to 15.67 TFLOPS for the V100 SXM2 32 GB.

Q: What are the TDP requirements for each card?

A: The B200 has a TDP of 1000 W with a suggested PSU of 1400 W, while the V100 SXM2 32 GB has a TDP of 250 W with a suggested PSU of 600 W.

Q: How does the B200 compare to its nearest rival, the NVIDIA H200 NVL?

A: The B200 is 3.2% ahead of the H200 NVL in average benchmark score, with the B200 at 345,482 and the H200 NVL at 334,891.

Q: Is the V100 SXM2 32 GB still in production?

A: No, its production status is End-of-life, and it was released on 2018-03-26. The B200’s production status is Active.

Where Each One Wins

The B200 wins decisively in raw compute benchmarks. Its Geekbench OpenCL score is 162.3% higher, and its FP16 output of 1,191.2 TFLOPS dwarfs the V100 SXM2 32 GB’s 31.33 TFLOPS. The B200’s 90 GB memory capacity and 4.10 TB/s bandwidth make it suitable for large model training and inference workloads where memory footprint and bandwidth are the limiting factors. Its 100th-percentile ranking confirms it is the top performer in the database, and its nearest rivals—the H200 NVL at 3.2% slower and the MI300X at 8.6% slower—do not close the gap meaningfully. The B200 also benefits from being Active production status and PCIe 5.0 connectivity, which future-proofs system integration.

The V100 SXM2 32 GB wins in efficiency and legacy compatibility. Its 250 W TDP is 75% lower than the B200’s 1000 W, making it far easier to cool and power in existing infrastructure. Its 32 GB of HBM2 memory, while smaller, is still substantial for many inference tasks. The V100 SXM2 32 GB’s 96th-percentile score shows it outperforms 96% of all GPUs in the database, and its nearest rivals—the V100 PCIe 16 GB at 0.2% faster and the MI100 at 0.9% faster—are essentially matched, meaning the SXM2 32 GB variant is not obsolete in its peer group. The V100 SXM2 32 GB’s higher ROP count (128 versus 24) and pixel rate (195.8 GPixel/s versus 47.16 GPixel/s) suggest it may handle certain rasterization tasks better, although its lack of display outputs limits that utility. For users with Volta-specific software stacks or power constraints, the V100 SXM2 32 GB remains a viable choice, but the data shows the B200 is the performance leader by a wide margin.