Intel Arc Pro B390 vs NVIDIA H20 Comparison

Where Each One Wins

The recorded data separates these two accelerators into entirely different operational categories. The Intel Arc Pro B390 is an integrated graphics processor (IGP) designed for portable devices, while the NVIDIA H20 is a server-class SXM module with no display outputs. Their benchmark profiles do not overlap, as evidenced by the empty head-to-head benchmark array and zero wins recorded for each side. The functional split is defined by form factor and deployment environment rather than direct performance comparison.

The Intel Arc Pro B390 wins in the domain of client-side graphics integration. It uses system shared memory, draws from the host platform's power budget at 80 W TDP, and requires no power connectors. Its display outputs are listed as portable device dependent, meaning it exists to drive screens in mobile systems. The H20 cannot win in this category because it has no display outputs at all. Any workload that requires rendering to a local panel falls exclusively to the Intel part.

The NVIDIA H20 wins in the domain of dedicated compute acceleration. It carries 96 GB of HBM3 memory across a 6144-bit bus, delivering 4.03 TB/s of bandwidth. The Intel part depends on system shared memory with system dependent bandwidth, a fundamental disadvantage for memory-bound server workloads. The H20 also provides 312 tensor cores, a feature class entirely absent from the Arc Pro B390's specification sheet. Machine learning inference and training tasks that leverage tensor operations have no equivalent path on the Intel IGP.

The physical deployment also dictates distinct wins. The H20 is an SXM module with a 500 W TDP and a suggested 900 W power supply, placing it in rack-mounted server infrastructure with dedicated cooling. The Arc Pro B390 is soldered into a portable device motherboard, consuming a fraction of the power and requiring no auxiliary cooling beyond the host chassis. Neither part can substitute for the other in these environments.

FAQ

Q: Which GPU has higher FP32 compute throughput?

A: The NVIDIA H20 delivers 39.54 TFLOPS FP32, which is approximately 5.1 times the 7.680 TFLOPS of the Intel Arc Pro B390.

Q: Do both GPUs support DirectX 12 Ultimate?

A: No. The Intel Arc Pro B390 supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. The NVIDIA H20 lists N/A for DirectX, OpenGL, and Vulkan, reflecting its server-oriented design with no graphics API support.

Q: What memory configurations do the two parts use?

A: The Intel Arc Pro B390 uses system shared memory with system dependent bandwidth. The NVIDIA H20 uses 96 GB of HBM3 memory on a 6144-bit bus with 4.03 TB/s bandwidth.

Q: Which product has a higher pixel fill rate?

A: The Intel Arc Pro B390 records 60.00 GPixel/s, while the NVIDIA H20 records 47.52 GPixel/s, giving the Intel part a higher pixel throughput despite its lower overall compute.

Q: What are the process nodes and foundries for each chip?

A: The Intel Arc Pro B390 uses a 3 nm process at Intel foundry. The NVIDIA H20 uses a 5 nm process at TSMC.

Q: Are both products currently in production?

A: Yes, both the Intel Arc Pro B390 and the NVIDIA H20 are listed with Active production status. The Intel part released on 2026-01-26, while the NVIDIA part released on 2024-01-31.

Head-to-Head Benchmarks

The head-to-head benchmark array is empty, and both items record zero wins and zero average benchmark scores. Direct numerical competition cannot be established from the recorded data. However, the specification tables provide measurable deltas that define the performance envelope of each part.

The most significant compute gap appears in FP32 throughput. The H20 produces 39.54 TFLOPS against 7.680 TFLOPS for the Arc Pro B390, a difference of 31.86 TFLOPS. In FP16, the H20 produces 79.07 TFLOPS versus 15.36 TFLOPS for the Intel part, a 63.71 TFLOPS gap. Both figures use a 2:1 ratio, so the relative scaling is consistent. The H20's advantage stems from 9984 shading units versus 1536, a 6.5 times unit count advantage, combined with higher clocks: 1980 MHz boost versus 2500 MHz boost. The Intel part actually clocks higher, but the sheer unit count difference overwhelms the clock advantage.

Texture throughput shows a similar pattern. The H20 records 617.8 GTexel/s from 312 TMUs, while the Arc Pro B390 records 120.0 GTexel/s from 48 TMUs. The pixel rate flips direction: 60.00 GPixel/s for Intel versus 47.52 GPixel/s for NVIDIA, despite both parts having 24 ROPs. The Intel part achieves its higher pixel rate through the 2500 MHz boost clock.

Memory bandwidth presents the largest absolute delta. The H20's 4.03 TB/s over a 6144-bit HBM3 interface compares to system dependent bandwidth for the Arc Pro B390. The H20's memory clock is listed at 1313 MHz with 5.3 Gbps effective, while the Intel part's memory clock is system shared. For memory-intensive workloads, the H20's dedicated HBM3 stack removes the contention and latency variability inherent to shared system memory.

Ray tracing hardware also splits the two. The Arc Pro B390 includes 12 RT cores, while the H20 lists no RT core count. The Intel part supports DirectX 12 Ultimate, which mandates hardware ray tracing support. The H20's absence of graphics APIs and RT cores indicates its compute-focused role.

The pixel rate result is the only measured metric where the Intel part leads. This aligns with its role as a display-driving IGP, where fixed-function output logic and rasterization matter more than raw compute density. The H20's lower pixel rate, despite far more shading units, reflects its design priority toward tensor and FP32 workloads rather than rasterization.

Specification Differences

The two products differ across nearly every recorded specification field. Process node: Intel uses 3 nm at Intel foundry, NVIDIA uses 5 nm at TSMC. Transistor count: the H20 has 80,000 million transistors on an 814 mm² die with a density of 98.3M per mm², while the Arc Pro B390's transistor count, die size, and density are unknown.

Clock behavior diverges sharply. The Arc Pro B390 runs at 300 MHz base and 2500 MHz boost. The H20 runs at 1830 MHz base and 1980 MHz boost. The Intel part has a much wider clock range and a higher peak clock, while the H20 operates in a narrow, sustained band suited to continuous server load.

Memory configuration is a total divergence. The Arc Pro B390 has system shared memory, system shared type, system shared bus width, and system dependent bandwidth. The H20 has 96 GB of HBM3, 6144-bit bus width, and 4.03 TB/s bandwidth. The memory clock also differs: system shared versus 1313 MHz (5.3 Gbps effective).

Compute unit counts: 1536 shading units, 48 TMUs, 24 ROPs, 12 RT cores, and no tensor cores for Intel. The H20 has 9984 shading units, 312 TMUs, 24 ROPs, no RT core count listed, and 312 tensor cores. The H20's tensor core count equals its TMU count, a common pattern in Hopper architecture.

Power and physical specifications: the Arc Pro B390 is an IGP at 80 W TDP with no power connectors and no suggested PSU. The H20 is an SXM module at 500 W TDP with a suggested 900 W PSU and no listed power connectors. Bus interface: IGP for Intel, PCIe 5.0 x16 for NVIDIA. Display outputs: portable device dependent for Intel, none for NVIDIA.

API support: Intel lists DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. NVIDIA lists N/A for all three. Release dates: 2026-01-26 for Intel, 2024-01-31 for NVIDIA. Predecessors: HD Graphics-WM for Intel, Server Ada for NVIDIA. Successors: none for Intel, Server Blackwell for NVIDIA. Both parts share Active production status and a 50th percentile ranking against all GPUs.

Architecture Differences

The Intel Arc Pro B390 uses the Xe3-LPG architecture on the Panther Lake chip, belonging to the Arc Graphics-WM (Panther Lake) generation. The NVIDIA H20 uses the Hopper architecture on the GH100 chip, belonging to the Server Hopper (Hxx) generation. These are fundamentally different design lineages: one extends Intel's integrated graphics roadmap, the other extends NVIDIA's data center compute line.

The Xe3-LPG architecture integrates graphics and compute into a mobile processor package. Its 3 nm process at Intel foundry represents a leading-edge node for power efficiency, consistent with an 80 W IGP. The architecture includes 12 RT cores and supports DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4, indicating a full client graphics feature set. The lack of tensor cores means AI acceleration relies on the host system or is absent entirely. The 2:1 FP16 ratio suggests shared execution resources rather than dedicated tensor hardware.

Hopper is a server compute architecture with no graphics API support. The 5 nm process at TSMC produces an 814 mm² die containing 80,000 million transistors, yielding a density of 98.3M per mm². The 312 tensor cores are the architectural centerpiece, designed for matrix operations in AI training and inference. The 96 GB HBM3 stack with 4.03 TB/s bandwidth feeds these tensor cores at a scale impossible for shared memory systems. The 500 W TDP and SXM form factor place it in actively cooled server chassis with dedicated power delivery.

The architectural priorities are visible in the unit allocations. Intel dedicates transistors to RT cores, ROPs, and API compatibility for display output. NVIDIA dedicates transistors to tensor cores, TMUs, and memory controllers for data throughput. The Intel part's 24 ROPs match the H20's 24 ROPs, but Intel achieves higher pixel rate through clock speed. The H20's 312 TMUs versus 48 for Intel shows where texture throughput matters: the H20 targets workloads that sample large data arrays, while the Intel part targets conventional 3D rendering.

The absence of DirectX support on the H20 is not a deficiency but a design boundary. Hopper GPUs do not expose graphics APIs because they are not installed in systems with displays. The Arc Pro B390's support for DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4 defines its role as a client rendering device. The two architectures share no common workload domain where a direct comparison would be meaningful.