Intel Arc B390 vs NVIDIA H100 CNX Comparison

The Verdict

The Intel Arc B390 is an integrated graphics processor aimed at portable devices, while the NVIDIA H100 CNX is a dual-slot server accelerator with no display outputs. The recorded data shows the Arc B390 has a 3DMark Steel Nomad DX12 score of 1482, placing it in the 9th percentile of all GPUs. Its nearest rivals, all NVIDIA parts from the GeForce GT 500M/600M/700M era, score between 1443 and 1463, meaning the Arc B390 leads them by 1.3% to 2.7%. The H100 CNX has no benchmark entries in the database, so direct comparative performance cannot be quantified; its 50th percentile rank is based on its class and specifications, not measured results.

For builders, the Arc B390 suits a thin-and-light device that needs basic 3D acceleration and modern API support. The H100 CNX targets server workloads requiring massive memory bandwidth and compute throughput, but it lacks any display outputs, so it is not a client graphics card. Neither part serves the same user, and the data does not support recommending one over the other for a shared use case.

Architecture Differences

The Arc B390 uses the Panther Lake chip with Xe3-LPG architecture, belonging to the Arc Graphics-M (Panther Lake) generation. It is built on a 3 nm process at Intel, with a base clock of 300 MHz and a boost clock of 2500 MHz. Its memory subsystem is entirely system-shared, meaning size, type, bus width, and bandwidth all depend on the host platform. The chip integrates 1536 shading units, 48 texture mapping units, 24 ROPs, and 12 ray tracing cores. It offers FP32 performance of 7.680 TFLOPS and FP16 performance of 15.36 TFLOPS with a 2:1 ratio. The TDP is 80 W, and it uses no power connectors because it is an integrated graphics processor (IGP) with an IGP bus interface. Display outputs are portable device dependent.

The H100 CNX uses the GH100 chip with Hopper architecture, part of the Server Hopper (Hxx) generation. It is built on a 5 nm process at TSMC, with 80,000 million transistors on a 814 mm² die, giving a transistor density of 98.3M per mm². Base clock is 690 MHz, boost clock is 1845 MHz, and memory runs at 1593 MHz with 3.2 Gbps effective. It has 80 GB of HBM2e memory on a 5120-bit bus, delivering 2.04 TB/s of bandwidth. The chip contains 14592 shading units, 456 TMUs, 24 ROPs, and 456 tensor cores; it has no listed RT cores. FP32 output is 53.84 TFLOPS, and FP16 output is 215.4 TFLOPS with a 4:1 ratio. TDP is 350 W, requiring an 8-pin EPS connector and a 750 W suggested PSU. It uses a PCIe 5.0 x16 interface and measures 267 mm in length and 111 mm in height. The H100 CNX has no display outputs.

The process node difference is significant: Intel uses 3 nm versus TSMC's 5 nm, but the H100 CNX compensates with a much larger die and far more transistors. The Arc B390 relies on shared system memory, while the H100 CNX has dedicated HBM2e with a 5120-bit bus. The H100 CNX also adds tensor cores, which the Arc B390 does not list.

Head-to-Head Benchmarks

The head-to-head benchmark table is empty, and neither part records a win count in direct comparisons. The only benchmark score in the entire dataset belongs to the Arc B390: 1482 in 3DMark Steel Nomad DX12. The H100 CNX has no benchmark entries, so any attempt to compare their measured performance would be speculative. What the data does show is how the Arc B390 stacks against its nearest rivals, all of which are low-end NVIDIA mobile or OEM parts. The Arc B390 scores 1.3% higher than the GeForce GT 520MX (1463), 1.5% higher than the GeForce 800M (1460), 2.5% higher than the GeForce GT 625 OEM (1446), and 2.7% higher than the GeForce GT 710 (1443). These are narrow margins, indicating the Arc B390 sits at the very bottom of the performance scale among all GPUs, just above some of the weakest discrete parts from a decade ago.

For the H100 CNX, the absence of benchmark data means the database cannot confirm any advantage in real-world tests. Its theoretical compute figures are far higher, but no measured score exists to verify them. The only quantitative comparison available is the percentile rank: the H100 CNX sits at the 50th percentile versus the Arc B390's 9th percentile, but that rank likely reflects the H100 CNX's server positioning rather than an actual performance test.

Specification Differences

The two parts differ in nearly every measurable specification. The Arc B390 uses a 3 nm process from Intel; the H100 CNX uses a 5 nm process from TSMC. Transistor count is unknown for the Arc B390, while the H100 CNX has 80,000 million transistors. Die size is unknown for the Arc B390, while the H100 CNX measures 814 mm². Transistor density is not listed for the Arc B390, but the H100 CNX has 98.3M per mm².

Clock speeds differ: the Arc B390 runs at 300 MHz base and 2500 MHz boost; the H100 CNX runs at 690 MHz base and 1845 MHz boost. Memory is system shared for the Arc B390, while the H100 CNX has 80 GB of HBM2e at 1593 MHz with 3.2 Gbps effective. Bus width is system shared for the Arc B390 versus 5120 bit for the H100 CNX. Bandwidth is system dependent for the Arc B390 versus 2.04 TB/s for the H100 CNX.

Compute resources differ substantially: the Arc B390 has 1536 shading units, 48 TMUs, 24 ROPs, and 12 RT cores; the H100 CNX has 14592 shading units, 456 TMUs, 24 ROPs, and 456 tensor cores, with no RT cores listed. Pixel rate is 60.00 GPixel/s for the Arc B390 versus 44.28 GPixel/s for the H100 CNX. Texture rate is 120.0 GTexel/s for the Arc B390 versus 841.3 GTexel/s for the H100 CNX. FP32 is 7.680 TFLOPS for the Arc B390 versus 53.84 TFLOPS for the H100 CNX. FP16 is 15.36 TFLOPS (2:1) for the Arc B390 versus 215.4 TFLOPS (4:1) for the H100 CNX.

Power and physical specs differ: the Arc B390 has an 80 W TDP, is an IGP with no power connectors, and uses an IGP bus interface. The H100 CNX has a 350 W TDP, is dual-slot, uses an 8-pin EPS connector, requires a 750 W suggested PSU, and uses PCIe 5.0 x16. Display outputs are portable device dependent for the Arc B390, while the H100 CNX has no outputs. API support also differs: the Arc B390 supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4; the H100 CNX lists no API support. Release dates differ: the Arc B390 launched on 2026-01-26, while the H100 CNX launched on 2023-03-20. The H100 CNX has a predecessor (Server Ada) and successor (Server Blackwell), while the Arc B390 lists none.

FAQ

Q: Which GPU has higher FP32 compute?

A: The NVIDIA H100 CNX delivers 53.84 TFLOPS of FP32, which is 7.0 times the Intel Arc B390's 7.680 TFLOPS.

Q: Does the Intel Arc B390 support ray tracing?

A: Yes, the Arc B390 has 12 dedicated RT cores, while the H100 CNX lists no RT cores in the database.

Q: What memory does the NVIDIA H100 CNX use?

A: The H100 CNX uses 80 GB of HBM2e with a 5120-bit bus and 2.04 TB/s of bandwidth. The Arc B390 uses system shared memory with system dependent bandwidth.

Q: Can the NVIDIA H100 CNX output to a display?

A: No, the H100 CNX has no display outputs. The Arc B390's display outputs are portable device dependent, meaning it relies on the host device's panel.

Q: How does the Arc B390 compare to its nearest rivals?

A: The Arc B390 scores 1482 in 3DMark Steel Nomad DX12, beating the GeForce GT 520MX by 1.3%, the GeForce 800M by 1.5%, the GeForce GT 625 OEM by 2.5%, and the GeForce GT 710 by 2.7%.

Q: Which GPU has more texture mapping units?

A: The H100 CNX has 456 TMUs versus 48 on the Arc B390, resulting in a texture rate of 841.3 GTexel/s versus 120.0 GTexel/s.

Where Each One Wins

The Intel Arc B390 wins in integrated graphics deployment. It is an IGP with no power connectors, an 80 W TDP, and an IGP bus interface, meaning it drops into a portable device without additional wiring. Its display outputs are portable device dependent, so it can drive the host panel directly. It also supports modern client APIs: DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4, which the H100 CNX lacks entirely. Its pixel rate of 60.00 GPixel/s exceeds the H100 CNX's 44.28 GPixel/s, which could matter for rasterization-bound tasks at lower resolutions. Its boost clock of 2500 MHz is also higher than the H100 CNX's 1845 MHz.

The NVIDIA H100 CNX wins in raw compute throughput. Its FP32 output of 53.84 TFLOPS is roughly seven times the Arc B390's, and its FP16 output of 215.4 TFLOPS is fourteen times higher. Texture rate is 841.3 GTexel/s versus 120.0 GTexel/s, a sevenfold advantage. Memory bandwidth is 2.04 TB/s from HBM2e, while the Arc B390 depends on system memory with no fixed bandwidth. The H100 CNX also has 14592 shading units, 456 TMUs, and 456 tensor cores, versus 1536 shading units and 48 TMUs on the Arc B390. Its 80 GB memory capacity dwarfs whatever the host system allocates to the Arc B390. The H100 CNX uses a PCIe 5.0 x16 interface, giving it a dedicated high-bandwidth connection to the host, whereas the Arc B390 shares system resources entirely.

The H100 CNX also wins on physical footprint in a server context: it is a dual-slot card with an 8-pin EPS connector and a 750 W suggested PSU, which is standard for rack deployment. The Arc B390 cannot be installed in a desktop or server at all, since it is an IGP. The H100 CNX's 350 W TDP is much higher than the Arc B390's 80 W, but that power budget buys a 53.84 TFLOPS FP32 result. Its transistor count of 80,000 million on an 814 mm² die indicates a far more complex chip, which aligns with its server positioning.

For a user building a portable device, the Arc B390 is the only viable option because the H100 CNX has no display outputs and cannot function as a client GPU. For a user populating a server rack, the H100 CNX is the only viable option because the Arc B390 is physically an IGP with no server interface. The data does not support any crossover use case.