AMD Radeon Instinct MI300 vs Intel Arc B390 Comparison

Where Each One Wins

The AMD Radeon Instinct MI300 and the Intel Arc B390 occupy entirely separate performance tiers, with the data showing no overlap in their intended roles. The MI300 is a data center compute accelerator built around the CDNA 3.0 architecture, designed for throughput-oriented workloads where massive memory pools and extreme bandwidth matter more than display output. The Arc B390, by contrast, is an integrated graphics processor within Intel's Panther Lake mobile platform, aimed at portable devices where power draw and physical footprint are primary constraints.

Benchmark results confirm the separation. The Arc B390 has a recorded 3DMark Steel Nomad DX12 score of 1482, placing it in the 9th percentile of all GPUs in the database. Its nearest recorded rivals are low-end discrete parts: the NVIDIA GeForce GT 520MX (1463, 1.3% behind), the GeForce 800M (1460, 1.5% behind), the GT 625 OEM (1446, 2.5% behind), and the GT 710 (1443, 2.7% behind). The MI300 has no recorded benchmark scores and no nearest rivals in the database, which is consistent with its role as a compute accelerator rather than a rasterization-focused graphics card.

The wins are therefore categorical rather than marginal. The MI300 wins in raw compute throughput, memory capacity, and memory bandwidth. The Arc B390 wins in power efficiency per watt, physical integration, and API support for consumer graphics workloads. The Arc B390 also wins in clock speed range, boosting to 2500 MHz compared to the MI300's 1700 MHz boost, though this advantage is irrelevant given the massive difference in shading unit count.

Architecture Differences

The two processors share almost nothing architecturally. The MI300 uses the CDNA 3.0 architecture on TSMC's 5 nm process, with a die size of 1017 mm² and 153,000 million transistors, yielding a transistor density of 150.4 million per square millimeter. The chip, codenamed Aqua Vanjaram, is a dedicated accelerator with no display outputs and no consumer graphics API support listed. Its memory subsystem consists of 128 GB of HBM3 on an 8192-bit bus, delivering 6.55 TB/s of bandwidth.

The Arc B390 uses Intel's Xe3-LPG architecture on Intel's 3 nm process, with transistor count and die size listed as unknown. It is an integrated graphics processor (IGP) within the Panther Lake generation, drawing its memory from the system pool: capacity, type, bus width, and bandwidth are all system dependent. The B390 includes 12 ray tracing cores, a feature entirely absent from the MI300's specification sheet. It also supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4, while the MI300 lists no consumer APIs at all.

The compute pipelines diverge sharply. The MI300 packs 14,080 shading units, 880 texture mapping units, and zero ROPs, with a pixel rate of 0 MPixel/s. Its texture rate reaches 1,496.0 GTexel/s. The Arc B390 has 1,536 shading units, 48 TMUs, and 24 ROPs, with a pixel rate of 60.00 GPixel/s and a texture rate of 120.0 GTexel/s. The MI300's FP32 throughput is 47.87 TFLOPS versus 7.680 TFLOPS for the B390; its FP16 throughput is 383.0 TFLOPS (8:1 ratio) versus 15.36 TFLOPS (2:1 ratio) for the B390.

FAQ

Q: Which processor has higher FP32 compute throughput?

A: The AMD Radeon Instinct MI300 delivers 47.87 TFLOPS of FP32 performance, which is approximately 6.2 times the 7.680 TFLOPS of the Intel Arc B390.

Q: Does the Intel Arc B390 support ray tracing?

A: Yes, the Arc B390 includes 12 ray tracing cores. The MI300 specification sheet does not list ray tracing cores.

Q: What memory configuration does each processor use?

A: The MI300 uses 128 GB of HBM3 on an 8192-bit bus with 6.55 TB/s bandwidth. The Arc B390 uses system shared memory, with capacity, type, bus width, and bandwidth all dependent on the host system.

Q: How do the recorded benchmark scores compare?

A: The Arc B390 has a 3DMark Steel Nomad DX12 score of 1482. The MI300 has no recorded benchmark scores in the database.

Q: What is the power consumption difference?

A: The MI300 has a TDP of 600 W and requires two 8-pin power connectors with a suggested PSU of 1000 W. The Arc B390 has a TDP of 80 W, uses no power connectors, and has no suggested PSU listed.

Q: Which processor has a smaller manufacturing process?

A: The Intel Arc B390 uses a 3 nm process from Intel, while the AMD MI300 uses a 5 nm process from TSMC.

Specification Differences

The two parts differ across nearly every measurable specification. The MI300 uses a 5 nm TSMC process, while the B390 uses a 3 nm Intel process. Transistor count is 153,000 million for the MI300 versus unknown for the B390; die size is 1017 mm² versus unknown. The MI300's base clock is 1000 MHz with a boost of 1700 MHz, while the B390 runs at 300 MHz base and 2500 MHz boost. Memory clocks differ: the MI300 operates at 1600 MHz with 6.4 Gbps effective, while the B390's memory clock is system shared.

Shading units: 14,080 for the MI300 versus 1,536 for the B390. TMUs: 880 versus 48. ROPs: 0 versus 24. The B390 has 12 ray tracing cores; the MI300 has none. Pixel rate: 0 MPixel/s versus 60.00 GPixel/s. Texture rate: 1,496.0 GTexel/s versus 120.0 GTexel/s. FP32: 47.87 TFLOPS versus 7.680 TFLOPS. FP16: 383.0 TFLOPS (8:1) versus 15.36 TFLOPS (2:1). TDP: 600 W versus 80 W. Power connectors: 2x 8-pin versus none. Bus interface: PCIe 5.0 x16 versus IGP. Display outputs: none versus portable device dependent.

API support differs completely. The MI300 lists no DirectX, OpenGL, or Vulkan versions. The B390 supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. The MI300 has dimensions of 267 mm (10.5 inches) in length and 111 mm (4.4 inches) in height. The B390 has no listed physical dimensions. Release dates are also far apart: the MI300 launched on January 3, 2023, while the B390 is dated January 26, 2026. The MI300's predecessor is the FirePro Data Center series; the B390 has no predecessor listed. The B390's production status is active; the MI300's is not listed.

Head-to-Head Benchmarks

The database contains no direct head-to-head benchmark comparisons between the MI300 and the B390, and the MI300 has no benchmark entries at all. The only recorded score belongs to the B390: 1482 in 3DMark Steel Nomad DX12. This absence of comparative data is itself informative. The MI300 is not built for the rasterization workloads that 3DMark measures; its zero ROP count and lack of display outputs indicate it cannot drive a monitor or render traditional graphics frames.

The B390's nearest rival data provides the only quantitative performance context available. Its 1482 score places it 1.3% ahead of the GeForce GT 520MX (1463), 1.5% ahead of the GeForce 800M (1460), 2.5% ahead of the GT 625 OEM (1446), and 2.7% ahead of the GT 710 (1443). These margins are small, placing the B390 at the extreme low end of the database's performance distribution with a 9th percentile rank.

For the MI300, the meaningful numbers are compute specifications rather than benchmark scores. Its FP32 throughput of 47.87 TFLOPS is 6.2 times the B390's 7.680 TFLOPS. Its FP16 throughput of 383.0 TFLOPS is 24.9 times the B390's 15.36 TFLOPS. Its memory bandwidth of 6.55 TB/s is several orders of magnitude beyond the B390's system-dependent shared memory. Its texture rate of 1,496.0 GTexel/s is 12.5 times the B390's 120.0 GTexel/s.

The pixel rate comparison runs the other direction: the B390 renders 60.00 GPixel/s, while the MI300 produces 0 MPixel/s. The B390's 24 ROPs deliver actual rasterization output; the MI300 has none. The B390 also carries 12 ray tracing cores, enabling hardware-accelerated ray tracing workloads that the MI300 cannot perform. The MI300's 600 W TDP and 1000 W suggested PSU reflect its data center positioning, while the B390's 80 W TDP and connector-free design suit its integrated mobile role.

The data supports a straightforward interpretation. The MI300 is a compute accelerator with no graphics output capability, trading rasterization hardware for extreme FP32 and FP16 throughput, massive HBM3 capacity, and industry-leading bandwidth. The B390 is a low-power integrated GPU with modern API support, ray tracing hardware, and modest rasterization performance that sits near the bottom of the database's benchmark distribution. Their specification sheets share almost no common ground, and the recorded data confirms they serve different markets entirely.