Intel Arc Pro B390 vs NVIDIA L20 Comparison

Head-to-Head Benchmarks

The Intel Arc Pro B390 and the NVIDIA L20 occupy opposing ends of the performance spectrum, and the recorded data confirms this without ambiguity. The NVIDIA L20 delivers a Geekbench OpenCL score of 274,276 and a Geekbench Vulkan score of 228,018, producing an average benchmark score of 251,147 across those tests. The Intel Arc Pro B390 holds a 50th percentile ranking among all GPUs in the database, with an average benchmark score of 0 and no recorded entries in the benchmark suite. The L20, by contrast, sits at the 99th percentile, a position that places it among the top 1% of all GPUs tracked by the database.

The performance gap is decisive in every measured category. In raw FP32 throughput, the L20 reaches 59.35 TFLOPS, while the Arc Pro B390 computes 7.680 TFLOPS, meaning the NVIDIA part delivers roughly 7.7 times the single-precision compute throughput. The FP16 comparison is similarly lopsided: the L20 produces 59.35 TFLOPS at a 1:1 ratio, whereas the Arc Pro B390 manages 15.36 TFLOPS at a 2:1 ratio. Even accounting for the Intel part's advantage of packing two FP16 operations into a single FP32 cycle, the L20 still exceeds it by a factor of approximately 3.9 in raw half-precision output.

Pixel throughput tells the same story. The L20 sustains 322.6 GPixel/s against the Arc Pro B390's 60.00 GPixel/s, a 5.4x advantage in fill rate. Texture rate widens the margin further: 927.4 GTexel/s versus 120.0 GTexel/s, or 7.7x. The L20 also carries 92 ray tracing cores and 368 tensor cores, while the Arc Pro B390 lists 12 ray tracing cores and no tensor core count in the database. The L20's nearest rivals in the database, the NVIDIA PG506-232 and AMD Radeon PRO W7900D, trail it by 11.6% and 14.2% respectively in average benchmark score. Above it, the NVIDIA L40 and RTX 6000 Ada Generation lead by 11.6% and 12.6%, positioning the L20 as a strong mid-to-upper tier server compute card rather than a flagship.

Neither part has head-to-head benchmark entries in the database, so direct same-scenario comparisons are unavailable. The closest available reference points are the L20's nearest rivals, which bracket its average score: the PG506-232 at 225,124 and the L40 at 284,111. The Arc Pro B390, lacking any benchmark records, cannot be placed on that scale numerically. Its 50th percentile ranking, however, indicates it lands in the middle of the database distribution, which is consistent with its integrated design and compute specifications.

Clock behavior differs as well. The Arc Pro B390 runs at a 300 MHz base and 2500 MHz boost, while the L20 operates at 1440 MHz base and 2520 MHz boost. The L20 therefore sustains a much higher floor frequency, and its 5 nm TSMC process node allows 76,300 million transistors on a 609 mm² die with a density of 125.3M transistors per mm². The Intel part uses a 3 nm Intel node with unknown transistor count and die size. The L20's higher clock rates and far larger execution resources combine to produce the massive throughput differential observed in every rate metric.

The Verdict

The data indicates two distinct use cases with minimal overlap. The NVIDIA L20 is a compute-oriented accelerator with 48 GB of GDDR6 memory on a 384-bit bus delivering 864.0 GB/s of bandwidth. Its 11,776 shading units, 368 texture mapping units, and 128 ROPs provide the execution width needed for heavy rendering, AI inference, and large dataset workloads. Its 99th percentile ranking, along with benchmark scores that place it 11.6% ahead of the PG506-232 and 14.2% ahead of the Radeon PRO W7900D, confirms it as a high-throughput server part. The only cards above it in its nearest rival set, the L40 and RTX 6000 Ada Generation, outscore it by 11.6% and 12.6%, respectively, making the L20 a strong but not top-tier choice within the database's server GPU hierarchy.

The Intel Arc Pro B390, by contrast, is an integrated graphics processor (IGP) with system-shared memory and no dedicated VRAM. Its 80 W TDP, lack of power connectors, and IGP bus interface position it for portable devices where space and power are constrained. Its 1536 shading units, 48 TMUs, and 24 ROPs are modest by comparison, and its pixel rate of 60.00 GPixel/s and texture rate of 120.0 GTexel/s reflect that scale. The 50th percentile ranking suggests it performs around the median of all GPUs in the database, which is notable for an integrated solution but far below the L20's 99th percentile.

From the recorded data, the L20 is the correct selection for workloads requiring sustained compute throughput, large memory capacity, or high-bandwidth access to that memory. The Arc Pro B390 is appropriate for systems where a discrete GPU cannot be installed, where power draw must stay low, or where the workload fits within shared system memory. The two products do not compete for the same socket or chassis: the L20 is a dual-slot PCIe 4.0 x16 card measuring 267 mm in length and 111 mm in height, while the Arc Pro B390 is an IGP with no slot width, dimensions, or power connector requirements listed. There is no benchmark score for the Arc Pro B390 to compare directly, so any decision between them must rest on the architectural and specification differences documented below.

FAQ

Q: Which GPU has the higher benchmark score?

A: The NVIDIA L20 has an average benchmark score of 251,147, based on a Geekbench OpenCL score of 274,276 and a Geekbench Vulkan score of 228,018. The Intel Arc Pro B390 has no recorded benchmark scores and an average benchmark score of 0.

Q: How does the L20 compare to its nearest rivals?

A: The L20's average score of 251,147 is 11.6% higher than the NVIDIA PG506-232 (225,124) and 14.2% higher than the AMD Radeon PRO W7900D (219,827). It trails the NVIDIA L40 (284,111) by 11.6% and the RTX 6000 Ada Generation (287,237) by 12.6%.

Q: What memory configurations do the two GPUs use?

A: The NVIDIA L20 has 48 GB of GDDR6 memory on a 384-bit bus with 864.0 GB/s bandwidth and 18 Gbps effective memory speed. The Intel Arc Pro B390 uses system shared memory, with system shared type, bus width, and system dependent bandwidth.

Q: Which GPU has higher FP32 compute throughput?

A: The NVIDIA L20 delivers 59.35 TFLOPS FP32, while the Intel Arc Pro B390 delivers 7.680 TFLOPS FP32. The L20 also matches its FP32 rate in FP16 at 59.35 TFLOPS, whereas the Arc Pro B390 reaches 15.36 TFLOPS FP16 via a 2:1 ratio.

Q: What are the power requirements for each GPU?

A: The NVIDIA L20 has a 275 W TDP, uses a single 16-pin power connector, and lists a suggested PSU of 600 W. The Intel Arc Pro B390 has an 80 W TDP, no power connectors, and no suggested PSU listed.

Q: Do both GPUs support the same graphics APIs?

A: Yes, both support DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. The L20 also provides 4x DisplayPort 1.4a outputs, while the Arc Pro B390's display outputs are listed as portable device dependent.

Specification Differences

The two GPUs differ in nearly every measurable specification. The Intel Arc Pro B390 uses a 3 nm process node from Intel, while the NVIDIA L20 uses a 5 nm node from TSMC. The L20 contains 76,300 million transistors on a 609 mm² die with a density of 125.3M transistors per mm²; the Arc Pro B390's transistor count and die size are listed as unknown. The L20's base clock is 1440 MHz and its boost clock is 2520 MHz, compared to 300 MHz base and 2500 MHz boost for the Arc Pro B390.

Memory separates the two completely. The L20 has 48 GB of GDDR6 on a 384-bit bus with 864.0 GB/s bandwidth and 2250 MHz memory clock (18 Gbps effective). The Arc Pro B390 has system shared memory of system shared type and bus width, with system dependent bandwidth. Shading units number 11,776 on the L20 versus 1,536 on the Arc Pro B390. TMUs total 368 versus 48, and ROPs total 128 versus 24. Ray tracing cores are 92 on the L20 and 12 on the Arc Pro B390; tensor cores are 368 on the L20 and not listed for the Arc Pro B390.

The L20's TDP is 275 W with a dual-slot form factor, a 16-pin power connector, and a suggested PSU of 600 W. The Arc Pro B390 has an 80 W TDP, is an IGP with no slot width, and uses no power connectors. The L20 connects via PCIe 4.0 x16 and measures 267 mm by 111 mm; the Arc Pro B390 uses an IGP bus interface with no dimensions listed. Display outputs on the L20 are 4x DisplayPort 1.4a, while the Arc Pro B390's outputs are portable device dependent. The L20 was released on 2023-11-15, and the Arc Pro B390 on 2026-01-26.

Architecture Differences

The NVIDIA L20 is built on the Ada Lovelace architecture using the AD102 chip, while the Intel Arc Pro B390 uses the Xe3-LPG architecture with the Panther Lake chip. The L20 belongs to the Server Ada (Lxx) generation, and the Arc Pro B390 belongs to the Arc Graphics-WM (Panther Lake) generation. The L20's predecessor is listed as Server Ampere and its successor as Server Hopper; the Arc Pro B390's predecessor is HD Graphics-WM and its successor is not listed.

The L20 is a discrete server accelerator with dedicated GDDR6 memory, 368 tensor cores for AI workloads, and 92 ray tracing cores for accelerated ray tracing. Its 1:1 FP16 to FP32 ratio indicates full-rate half-precision execution, which is a common design choice for compute-heavy server parts. The Arc Pro B390 has no tensor core count listed and a 2:1 FP16 ratio, meaning it processes FP16 at half the rate of FP32. Its 12 ray tracing cores and 24 ROPs align with an integrated graphics design intended for lighter workloads.

Process technology differs by foundry and node: Intel fabricates the Arc Pro B390 on its own 3 nm process, while TSMC fabricates the L20 on 5 nm. The L20's transistor density of 125.3M per mm² reflects a mature, high-density server process. The Arc Pro B390's transistor density is not listed. The L20's memory clock of 2250 MHz with 18 Gbps effective transfer rate and 864.0 GB/s bandwidth supports large data movement, whereas the Arc Pro B390 relies entirely on system memory bandwidth, which is system dependent.

The L20's 275 W power envelope and dual-slot cooling solution indicate a card designed for sustained server workloads. The Arc Pro B390's 80 W TDP and IGP form factor indicate integration into a host processor package. Both parts support DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4, so API compatibility is not a differentiator. The architectural separation is clear: Ada Lovelace is a high-throughput discrete compute architecture with dedicated memory and AI acceleration, while Xe3-LPG is a power-efficient integrated architecture sharing system resources.