NVIDIA L4

The NVIDIA L4 is a single-slot server accelerator built on the Ada Lovelace architecture, featuring 24 GB of GDDR6 memory and a 72 W TDP. Its average benchmark score of 128,665 places it in the 97th percentile of all GPUs, a strong showing for a card that requires no auxiliary power connectors. Against its nearest rivals, the L4 lands in a tight cluster: it trails the GeForce RTX 3090 Ti by 2.5%, the Radeon PRO W6800 by 3.7%, and the Radeon RX 9070 GRE by 4.3%, while leading the Radeon PRO W7700 by 4.2%. These margins are narrow, making the L4 a competitive option in its performance tier.

How It Compares

The L4 sits 2.5% behind the NVIDIA GeForce RTX 3090 Ti, which posts an average score of 131,911. That is a remarkably small gap given the L4’s 72 W TDP and single-slot design, and it suggests the Ada Lovelace architecture extracts a high level of performance from a much lower power envelope. In practice, the L4 will deliver near-parity with the 3090 Ti in many compute workloads, though the 3090 Ti retains a slight edge in aggregate benchmarks.

Against the AMD Radeon PRO W6800, the L4 is 3.7% slower, with the PRO W6800 averaging 133,588. This is another close result, and the L4’s 24 GB memory capacity matches the PRO W6800’s professional positioning. The data shows a meaningful but not decisive advantage for the AMD card, and the L4’s power efficiency may offset that gap in dense server deployments.

The L4 is 4.2% faster than the AMD Radeon PRO W7700, which averages 123,434. This is the L4’s clearest win among its nearest rivals. The PRO W7700 is a capable workstation card, but the L4’s higher average score and larger memory capacity give it an edge in memory-intensive tasks. The 4.2% delta is consistent across the benchmark suite, indicating a stable performance advantage.

The AMD Radeon RX 9070 GRE leads the L4 by 4.3%, with an average score of 134,417. This is the largest deficit in the rival group, but it is still a narrow margin. The RX 9070 GRE is a consumer-oriented card, while the L4 is designed for server use, so the comparison highlights how close the L4 comes to a high-end consumer GPU despite its headless design and low power draw.

Ray Tracing and Feature Set

The L4 includes 60 RT cores and 240 tensor cores, both built on the Ada Lovelace architecture. This hardware enables hardware-accelerated ray tracing and tensor-based compute, making the card suitable for rendering, AI inference, and similar workloads. The API support is comprehensive: DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4 are all present. DirectX 12 Ultimate support indicates feature level 12_2, which includes advanced ray tracing and mesh shader capabilities.

The card has no display outputs, which confirms its role as a server accelerator rather than a desktop graphics card. It belongs to the Server Ada (Lxx) generation, with the predecessor listed as Server Ampere and the successor as Server Hopper. The tensor cores are a key differentiator for AI workloads, while the RT cores provide dedicated ray tracing throughput. The FP32 and FP16 performance are both rated at 30.29 TFLOPS, with a 1:1 ratio, meaning the card does not lose half-rate FP16 throughput, which is advantageous for mixed-precision compute.

Power and Cooling

The L4 has a TDP of 72 W, an exceptionally low figure for a card with this level of performance. It requires no power connectors, and the suggested PSU is 250 W. This makes the L4 easy to integrate into existing server chassis without additional power cabling. The single-slot form factor is 169 mm in length and 56 mm in height, allowing for dense multi-card configurations.

The low power draw also means cooling requirements are modest. A single-slot cooler is sufficient to handle the 72 W TDP, and the compact dimensions fit standard server racks. The lack of power connectors simplifies installation, and the 250 W PSU recommendation leaves ample headroom for the rest of the system. This combination of low power and compact size is a major advantage for data center environments where space and thermal budgets are tight.

FAQ

Q: What is the NVIDIA L4’s average benchmark score?

A: The average benchmark score is 128,665, placing it in the 97th percentile of all GPUs.

Q: How does the L4 compare to the GeForce RTX 3090 Ti?

A: The L4 trails the RTX 3090 Ti by 2.5%, with the 3090 Ti averaging 131,911.

Q: Does the L4 require power connectors?

A: No. The L4 has a 72 W TDP and no power connectors, and the suggested PSU is 250 W.

Q: What memory does the L4 have?

A: It has 24 GB of GDDR6 memory on a 192-bit bus, with 300.1 GB/s of bandwidth.

Q: What APIs does the L4 support?

A: It supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4.

Q: What is the L4’s form factor?

A: It is a single-slot card measuring 169 mm in length and 56 mm in height, with a PCIe 4.0 x16 interface.

Benchmark Performance

The L4’s benchmark results show a clear performance profile. In Geekbench OpenCL, it scores 140,838, while in Geekbench Vulkan it scores 116,491. The OpenCL result is substantially higher, indicating that the card’s compute performance is better expressed through OpenCL workloads. The average of these two tests is 128,665, which is the figure used for comparisons against rivals.

Relative to the RTX 3090 Ti, the L4 is 2.5% slower. That is a narrow margin, and the L4’s 24 GB memory capacity may close the gap in memory-bound tasks. Against the Radeon PRO W6800, the L4 is 3.7% slower, with the PRO W6800 averaging 133,588. The Radeon RX 9070 GRE is the strongest rival, leading by 4.3% with an average of 134,417. The L4’s best showing is against the Radeon PRO W7700, where it is 4.2% faster.

The 97th percentile ranking puts the L4 above the vast majority of GPUs, even though it is not the fastest in its immediate rival group. The deltas are all within a 4.3% range, meaning the L4 is highly competitive with these cards. The Vulkan score is notably lower than the OpenCL score, which may be relevant for workloads that rely on Vulkan compute. Overall, the benchmark data indicates a well-rounded accelerator that trades a small amount of peak performance for drastically lower power consumption.

Who Should Consider It

The L4 is best suited for server and data center deployments where power efficiency and density are priorities. Its 72 W TDP and single-slot design make it ideal for multi-GPU systems, and the absence of power connectors simplifies cabling. The 24 GB memory capacity is substantial for large models and datasets, and the 300.1 GB/s bandwidth provides enough throughput for many compute tasks.

For high-resolution compute workloads, the L4’s 24 GB GDDR6 buffer is a key asset. The card’s 97th percentile ranking ensures it can handle demanding tasks, while the 60 RT cores and 240 tensor cores add specialized acceleration for ray tracing and AI. However, because the L4 has no display outputs, it is not suitable for direct desktop use. It is a headless accelerator designed for servers, and its performance is competitive with high-end workstation cards despite its low power draw.

Users who prioritize raw benchmark scores above all else may prefer the RX 9070 GRE or RTX 3090 Ti, which lead by 4.3% and 2.5%, respectively. But for environments where power, space, and thermal limits are critical, the L4’s combination of performance and efficiency is compelling. The single-slot form factor and 169 mm length allow for high-density configurations, and the 250 W PSU recommendation means even modest server power supplies can support it.

Memory Subsystem

The L4 is equipped with 24 GB of GDDR6 memory, a substantial amount for a server accelerator. The memory bus is 192-bit, and the total bandwidth is 300.1 GB/s. The memory clock is 1563 MHz, with an effective data rate of 12.5 Gbps. This configuration provides a balance between capacity and bandwidth, though the 192-bit bus limits raw throughput compared to wider-memory cards.

For high-resolution workloads, the 24 GB capacity is the standout feature. Large textures, big batch sizes, and complex models can reside entirely in memory, reducing the need for frequent data transfers. The 300.1 GB/s bandwidth is sufficient for many compute tasks, but memory-bound workloads may see a bottleneck when accessing large datasets. Still, the capacity far exceeds what most consumer GPUs offer, making the L4 a strong choice for inference, rendering, and other memory-intensive applications.

The 1:1 FP32/FP16 ratio (30.29 TFLOPS each) further enhances compute flexibility, and the 240 tensor cores can accelerate AI workloads. The memory subsystem, combined with the Ada Lovelace architecture, positions the L4 as a versatile accelerator for server environments. Its 24 GB buffer and 300.1 GB/s bandwidth are well matched to the card’s compute capabilities, and the low 72 W TDP means that memory capacity does not come at the cost of power efficiency.