NVIDIA L40S vs NVIDIA Tesla V100 SXM2 16 GB Comparison

FAQ

Q: How does the NVIDIA L40S compare to the Tesla V100 SXM2 16 GB in overall benchmark performance?

A: The L40S has a significantly higher average benchmark score of 295,763 compared to the V100's 114,395, a difference of roughly 158%. The L40S also sits in the 99th percentile of all GPUs, while the V100 sits in the 95th percentile.

Q: Which GPU wins in Geekbench OpenCL and by how much?

A: The L40S wins decisively, scoring 330,727 versus 87,454 for the V100. This represents a delta of 278.2%, meaning the L40S is nearly four times faster in this compute workload.

Q: What about Vulkan performance? Is the gap as large?

A: The L40S also wins in Vulkan, scoring 260,799 against the V100's 141,336. The delta is 84.5%, a substantial but smaller margin than OpenCL.

Q: What are the memory specifications of each card?

A: The L40S has 48 GB of GDDR6 memory on a 384-bit bus with 864.0 GB/s bandwidth. The V100 has 16 GB of HBM2 memory on a 4096-bit bus with 897.0 GB/s bandwidth. Despite the V100's higher bandwidth, its capacity is a third of the L40S.

Q: Do both cards support ray tracing and Tensor cores?

A: The L40S has 142 RT cores and 568 Tensor cores. The V100 has 640 Tensor cores but no RT cores. DirectX support differs as well: the L40S supports DirectX 12 Ultimate (12_2), while the V100 supports DirectX 12 (12_1).

Q: What is the production status of these GPUs?

A: Both are listed as end-of-life. The L40S was released on 2022-10-12, and the V100 was released earlier on 2017-06-20.

Architecture Differences

The NVIDIA L40S is built on the AD102 chip using the Ada Lovelace architecture, fabricated on a 5 nm process at TSMC. It packs 76,300 million transistors on a 609 mm² die, giving a transistor density of 125.3 million per mm². The Tesla V100 SXM2 16 GB uses the GV100 chip with the older Volta architecture, on a 12 nm process, also at TSMC. It contains 21,100 million transistors on a larger 815 mm² die, but with a much lower density of 25.9 million per mm².

Clock behavior diverges sharply. The L40S runs at a base of 1110 MHz and boosts to 2520 MHz, while the V100 starts higher at 1312 MHz base but only boosts to 1530 MHz. Memory clocks also differ: the L40S uses 2250 MHz (18 Gbps effective) GDDR6, whereas the V100 uses 876 MHz (1752 Mbps effective) HBM2.

The L40S has 18,176 shading units, 568 TMUs, and 192 ROPs. It also includes 142 RT cores and 568 Tensor cores. The V100 has 5,120 shading units, 320 TMUs, and 128 ROPs, with 640 Tensor cores but no RT cores. This makes the L40S the only one of the two capable of hardware ray tracing.

Compute throughput shows a generation leap. The L40S delivers 91.61 TFLOPS FP32 and 91.61 TFLOPS FP16 (1:1 ratio). The V100 delivers 15.67 TFLOPS FP32 and 31.33 TFLOPS FP16 (2:1 ratio). Pixel and texture rates follow: the L40S reaches 483.8 GPixel/s and 1,431.4 GTexel/s, versus 195.8 GPixel/s and 489.6 GTexel/s for the V100.

The L40S is a dual-slot PCIe 4.0 x16 card with a 1x 16-pin power connector and a 300 W TDP. The V100 is an SXM module with no power connectors and a 250 W TDP. The L40S has display outputs (1x HDMI 2.1, 3x DisplayPort 1.4a), while the V100 has none. The L40S measures 267 mm in length and 111 mm in height; the V100 has no listed dimensions.

API support differs. The L40S supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.3. The V100 supports DirectX 12 (12_1), OpenGL 4.6, and Vulkan 1.4. The L40S generation is listed as Server Ada (Lxx), and the V100 as Tesla Volta (Vxx).

Head-to-Head Benchmarks

The Geekbench OpenCL test shows a massive win for the L40S. Its score of 330,727 dwarfs the V100's 87,454, a delta of 278.2%. In practical terms, the L40S excels in compute-heavy workloads like FP32 throughput, where it holds a 91.61 TFLOPS versus 15.67 TFLOPS.

In Geekbench Vulkan, the L40S again takes the lead, with 260,799 versus 141,336, a delta of 84.5%. While the V100 has a higher memory bandwidth per byte (897.0 GB/s versus 864.0 GB/s), the L40S compensates with more than triple the memory capacity (48 GB versus 16 GB) and more than triple the shading units.

One key area where the V100 holds an edge is FP16 compute. The V100's 31.33 TFLOPS FP16, while lower than the L40S's 91.61 TFLOPS, is achieved at a 2:1 ratio, meaning it is optimized for certain mixed-precision tasks. However, the L40S's 1:1 ratio makes it far more flexible for FP16 workloads that need full throughput.

The L40S also wins in raw pixel and texture rates: 483.8 GPixel/s versus 195.8 GPixel/s, and 1,431.4 GTexel/s versus 489.6 GTexel/s. The V100 has no RT cores, so any ray-traced workload is impossible on it, while the L40S has 142 RT cores.

The benchmark data shows two wins for the L40S and zero for the V100. The closest rival to the V100 in the database is the NVIDIA RTX A5500 Mobile, which scores 113,944, just 0.4% behind. The V100's nearest rivals include the RTX 4000 SFF Ada Generation (117,088, 2.3% ahead) and the NVIDIA GB10 (117,393, 2.6% ahead). The AMD Radeon PRO W7900 trails by 3.3%.

For the L40S, its nearest rivals are all higher-end cards: the NVIDIA H200 NVL leads by 11.7%, the AMD Instinct MI300X leads by 7%, and the NVIDIA RTX 6000 Ada Generation trails by 3%. The NVIDIA L40 is 4.1% behind. This places the L40S in the upper echelon of the database, far above the V100.

The Verdict

The data is clear: the NVIDIA L40S is the superior GPU by every measurable benchmark metric. Its average score of 295,763 is 158% higher than the V100's 114,395. In OpenCL, it is 278.2% faster; in Vulkan, it is 84.5% faster. The L40S also offers a 48 GB memory capacity versus 16 GB, and it supports ray tracing, which the V100 cannot.

The V100, however, retains niche advantages. Its HBM2 memory provides 897.0 GB/s bandwidth, slightly higher than the L40S's 864.0 GB/s. Its FP16 throughput at 31.33 TFLOPS, while lower, is achieved with a 2:1 ratio, which may be relevant for specific legacy workloads. It also has a lower TDP of 250 W versus 300 W, and it is an SXM module, which may fit certain server chassis.

For any modern workload, the L40S is the recommended choice. It dominates in compute, memory capacity, and feature support. The V100 is end-of-life and outclassed by even the RTX A5500 Mobile in the database, which scores 113,944 versus 114,395.

The verdict: choose the L40S for AI, rendering, or any compute-heavy task. Choose the V100 only if you have a specific need for HBM2 bandwidth or a 2:1 FP16 ratio, and you cannot use a newer card.

Specification Differences

• Process Node: L40S: 5 nm; V100: 12 nm
• Transistors: L40S: 76,300 million; V100: 21,100 million
• Die Size: L40S: 609 mm²; V100: 815 mm²
• Transistor Density: L40S: 125.3M / mm²; V100: 25.9M / mm²
• Base Clock: L40S: 1110 MHz; V100: 1312 MHz
• Boost Clock: L40S: 2520 MHz; V100: 1530 MHz
• Memory Clock: L40S: 2250 MHz (18 Gbps effective); V100: 876 MHz (1752 Mbps effective)
• Memory Size: L40S: 48 GB; V100: 16 GB
• Memory Type: L40S: GDDR6; V100: HBM2
• Memory Bus Width: L40S: 384 bit; V100: 4096 bit
• Memory Bandwidth: L40S: 864.0 GB/s; V100: 897.0 GB/s
• Shading Units: L40S: 18,176; V100: 5,120
• TMUs: L40S: 568; V100: 320
• ROPs: L40S: 192; V100: 128
• RT Cores: L40S: 142; V100: None
• Tensor Cores: L40S: 568; V100: 640
• Pixel Rate: L40S: 483.8 GPixel/s; V100: 195.8 GPixel/s
• Texture Rate: L40S: 1,431.4 GTexel/s; V100: 489.6 GTexel/s
• FP32: L40S: 91.61 TFLOPS; V100: 15.67 TFLOPS
• FP16: L40S: 91.61 TFLOPS (1:1); V100: 31.33 TFLOPS (2:1)
• TDP: L40S: 300 W; V100: 250 W
• Slot Width: L40S: Dual-slot; V100: SXM Module
• Power Connectors: L40S: 1x 16-pin; V100: None
• Suggested PSU: L40S: 700 W; V100: 600 W
• Bus Interface: L40S: PCIe 4.0 x16; V100: PCIe 3.0 x16
• Display Outputs: L40S: 1x HDMI 2.1, 3x DisplayPort 1.4a; V100: No outputs
• DirectX: L40S: 12 Ultimate (12_2); V100: 12 (12_1)
• Release Date: L40S: 2022-10-12; V100: 2017-06-20

Where Each One Wins

The L40S wins in nearly every category that matters for modern workloads. It dominates raw compute: 91.61 TFLOPS FP32 versus 15.67 TFLOPS. It offers triple the memory capacity (48 GB versus 16 GB), which is critical for large AI models or complex scenes. It is the only card with RT cores, making it suitable for ray-traced rendering. Its Vulkan score of 260,799 versus 141,336 shows a clear advantage in graphics APIs.

The V100's only wins are narrow and specific. It has a higher memory bandwidth at 897.0 GB/s versus 864.0 GB/s, which could benefit workloads that are bandwidth-bound rather than capacity-bound. Its FP16 2:1 ratio means it can process half-precision data at higher throughput relative to FP32, though the absolute FP16 number is lower than the L40S. It also has a lower TDP at 250 W, and its SXM module format may be a requirement for certain existing server deployments.

The V100's Tensor core count is higher at 640 versus 568, but without RT cores, its utility in ray tracing is absent. For any user considering the V100 today, the data suggests it is outclassed by the RTX A5500 Mobile in the database, so its only justification is legacy compatibility.

In summary, choose the L40S for AI training, simulation, or any compute-heavy task. Choose the V100 for its HBM2 bandwidth and 2:1 FP16 ratio, but expect no other meaningful performance advantages.