AMD Instinct MI350X vs Intel Arc B390 Comparison

The AMD Instinct MI350X and Intel Arc B390 occupy opposite ends of the GPU spectrum. The MI350X is a compute accelerator with a 1000 W power envelope, while the B390 is an integrated graphics processor with an 80 W power envelope. The database contains no direct head-to-head benchmark results between the two, but the recorded specifications and the one available benchmark score for the B390 allow for a structural comparison.

Head-to-Head Benchmarks

No direct head-to-head benchmark entries exist in the database for this pairing. The MI350X has no recorded benchmark scores, and its average benchmark score is zero. The B390 has a single recorded result: a score of 1482 in the 3DMark Steel Nomad DX12 test. This score places the B390 at the 9th percentile of all GPUs in the database.

The B390's nearest rivals in the database provide context for its performance. The NVIDIA GeForce GT 520MX scores 1463, which is 1.3% lower than the B390. The NVIDIA GeForce 800M scores 1460, which is 1.5% lower. The NVIDIA GeForce GT 625 OEM scores 1446, which is 2.5% lower. The NVIDIA GeForce GT 710 scores 1443, which is 2.7% lower. These are narrow margins, all within 3 percentage points. The B390 leads this small cluster, but the performance gap between it and these older discrete parts is minimal.

The MI350X, by contrast, has no benchmark entries. Its percentile rank of 50 reflects the absence of measured data, not a measured performance level. The database assigns it a neutral percentile because no scores were recorded. The FP32 compute rate of 72.09 TFLOPS and the texture rate of 2,252.8 GTexel/s are specification-derived figures, not benchmark results. The MI350X cannot be placed on the same 3DMark leaderboard as the B390 because no benchmark run was logged for it.

The B390's 7.680 TFLOPS FP32 rate and 120.0 GTexel/s texture rate are the only comparable measured specification points. The MI350X delivers roughly 9.4 times the FP32 throughput of the B390 (72.09 divided by 7.680 equals 9.39). The texture rate difference is larger: 2,252.8 divided by 120.0 equals 18.77 times. The pixel rate difference is extreme in the opposite direction, because the MI350X has a pixel rate of 0 MPixel/s while the B390 has 60.00 GPixel/s. The MI350X has no ROP units, so it cannot rasterize pixels. The B390 has 24 ROPs, which explains its 60.00 GPixel/s pixel rate.

In memory bandwidth, the MI350X lists 8.19 TB/s over an 8192-bit bus using HBM3e. The B390 uses system shared memory with system dependent bandwidth, so no numeric comparison can be made. The MI350X has 288 GB of dedicated HBM3e memory. The B390 has no dedicated memory; it shares system memory and the bus width is listed as system shared.

The Verdict

The data indicates two entirely different products for two entirely different workloads. The MI350X is a 1000 W accelerator with 16,384 shading units, 1,024 TMUs, and zero ROPs. It has no display outputs and exposes no DirectX, OpenGL, or Vulkan APIs. It is not a graphics card in the conventional sense; it is a compute module with a 2380 mm² die, 185,000 million transistors, and a 3 nm TSMC process node. Its FP16 rate matches its FP32 rate at 72.09 TFLOPS with a 1:1 ratio, indicating compute-oriented design with no specialized half-precision boost.

The B390 is an integrated processor with 1,536 shading units, 48 TMUs, 24 ROPs, and 12 ray tracing cores. It supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. It has an 80 W TDP and is marked as production Active. Its FP16 rate is 15.36 TFLOPS with a 2:1 ratio, meaning it doubles FP16 throughput relative to FP32. Its 3DMark Steel Nomad score of 1482 places it near the bottom of the database, around the 9th percentile.

For compute workloads that rely on FP32 or FP16 throughput, the MI350X is the clear choice. Its FP32 rate is 72.09 TFLOPS versus 7.680 TFLOPS for the B390. For graphics rendering, the B390 is the only viable option of the two, because the MI350X has no ROPs, no pixel rate, and no graphics API support. The MI350X cannot produce a rasterized frame. The B390 can, at a modest 60.00 GPixel/s.

The MI350X's 1:1 FP16 to FP32 ratio means the accelerator does not gain extra throughput from half precision. The B390's 2:1 ratio means it processes FP16 at twice its FP32 rate, which could benefit workloads that use half-precision shaders or compute kernels. However, even at 15.36 TFLOPS FP16, the B390 remains far below the MI350X's 72.09 TFLOPS FP32 rate.

The MI350X has no launch MSRP in the database, so no pricing statement can be made. The B390 also has no launch MSRP.

Architecture Differences

The MI350X uses the CDNA 4.0 architecture on a 3 nm TSMC process. Its chip is labeled MI350 256CU, indicating 256 compute units. The die measures 2380 mm² and contains 185,000 million transistors, yielding a transistor density of 77.7 million transistors per square millimeter. The B390 uses the Xe3-LPG architecture on a 3 nm Intel process, with a Panther Lake chip. Its transistor count and die size are unknown in the database.

The MI350X has 16,384 shading units and 1,024 TMUs. It has no ROPs, no ray tracing cores, and no tensor cores listed. The B390 has 1,536 shading units, 48 TMUs, 24 ROPs, and 12 ray tracing cores. The MI350X's memory subsystem uses 288 GB of HBM3e on an 8192-bit bus with 8.19 TB/s bandwidth. The B390 uses system shared memory with no dedicated bus width or bandwidth figures. The MI350X memory clock is 2000 MHz with 8 Gbps effective speed. The B390 memory clock is marked as system shared.

Clocks differ in base and boost behavior. The MI350X has a 1000 MHz base clock and a 2200 MHz boost clock. The B390 has a 300 MHz base clock and a 2500 MHz boost clock. The B390 boosts higher, but the MI350X maintains a much higher base frequency. The MI350X has a pixel rate of 0 MPixel/s and a texture rate of 2,252.8 GTexel/s. The B390 has a pixel rate of 60.00 GPixel/s and a texture rate of 120.0 GTexel/s.

The MI350X is an OAM module with no power connectors and a suggested PSU of 1400 W. Its TDP is 1000 W. It connects via PCIe 5.0 x16. The B390 is an IGP with no power connectors and no suggested PSU, with an 80 W TDP. Its bus interface is listed as IGP, meaning it is integrated into a processor package.

API support separates the two clearly. The MI350X lists N/A for DirectX, OpenGL, and Vulkan. The B390 supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. The MI350X has no display outputs. The B390's display outputs are portable device dependent.

Release dates also differ. The MI350X was released on June 11, 2025. The B390 was released on January 26, 2026. The MI350X's predecessor is listed as Radeon Instinct. The B390 has no predecessor listed.

FAQ

Q: Which GPU has higher FP32 compute throughput?

A: The AMD Instinct MI350X delivers 72.09 TFLOPS FP32. The Intel Arc B390 delivers 7.680 TFLOPS FP32. The MI350X is approximately 9.4 times higher.

Q: Can the AMD Instinct MI350X run DirectX 12 games?

A: No. The MI350X lists DirectX as N/A, along with OpenGL and Vulkan. It also has no display outputs and a pixel rate of 0 MPixel/s. The Intel Arc B390 supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4.

Q: What is the Intel Arc B390's 3DMark Steel Nomad score?

A: The B390 scored 1482 in the 3DMark Steel Nomad DX12 test. This places it at the 9th percentile of all GPUs in the database. Its nearest rivals are the NVIDIA GeForce GT 520MX at 1463 (1.3% lower), the NVIDIA GeForce 800M at 1460 (1.5% lower), the NVIDIA GeForce GT 625 OEM at 1446 (2.5% lower), and the NVIDIA GeForce GT 710 at 1443 (2.7% lower).

Q: How much memory does each GPU have?

A: The MI350X has 288 GB of HBM3e memory on an 8192-bit bus with 8.19 TB/s bandwidth. The B390 has no dedicated memory; its memory size, type, and bus width are all listed as system shared, with system dependent bandwidth.

Q: What are the power requirements?

A: The MI350X has a 1000 W TDP and a suggested PSU of 1400 W. The B390 has an 80 W TDP and no suggested PSU listed.

Q: Which GPU has ray tracing support?

A: The B390 has 12 ray tracing cores. The MI350X has no ray tracing cores listed.

Where Each One Wins

The MI350X wins in compute throughput, memory capacity, and memory bandwidth. Its FP32 rate of 72.09 TFLOPS dwarfs the B390's 7.680 TFLOPS. Its FP16 rate matches at 72.09 TFLOPS with a 1:1 ratio, which means half-precision workloads gain no extra speed, but the absolute FP16 throughput still exceeds the B390's 15.36 TFLOPS. The texture rate of 2,252.8 GTexel/s is 18.77 times the B390's 120.0 GTexel/s. The 288 GB HBM3e pool with 8.19 TB/s bandwidth is designed for large datasets that the B390's system shared memory cannot match in capacity or dedicated bandwidth. The 8192-bit bus width is a server-class memory interface. The 1000 W TDP and 1400 W suggested PSU indicate a device intended for sustained, power-hungry compute loads, not casual use.

The B390 wins in graphics features, rasterization, and API compatibility. It has 24 ROPs and a pixel rate of 60.00 GPixel/s, while the MI350X has zero ROPs and a pixel rate of 0 MPixel/s. Only the B390 can rasterize frames. It supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4, while the MI350X lists N/A for all three. The 12 ray tracing cores give the B390 hardware acceleration for ray traced effects, which the MI350X lacks entirely. The B390's boost clock of 2500 MHz is higher than the MI350X's 2200 MHz boost, and its 80 W TDP makes it suitable for integrated, low-power environments. The B390 also has a 2:1 FP16 ratio, doubling its FP16 throughput to 15.36 TFLOPS, which is useful for graphics workloads that use half-precision data.

The database shows no workload where the MI350X can output video. It has no display outputs. The B390's display outputs are portable device dependent, but they exist. For any task involving a screen, the B390 is the only functional option. For any task involving raw compute on large memory buffers, the MI350X is the only serious option. The 3DMark Steel Nomad score of 1482 for the B390, with nearest rivals all within 2.7%, confirms that the B390 sits at the low end of the performance spectrum. The MI350X has no benchmark score, so its percentile of 50 is a placeholder, not a measured result.

The transistor and die data reinforce the divide. The MI350X uses 185,000 million transistors on a 2380 mm² die at 77.7 million transistors per square millimeter. The B390's transistor count and die size are unknown. The MI350X is a 3 nm TSMC product; the B390 is a 3 nm Intel product. Both use 3 nm processes, but the scale of the MI350X silicon is far larger. The MI350X is an OAM module, the B390 is an IGP. The MI350X connects through PCIe 5.0 x16, the B390 through its integrated bus interface. These are not competing products; they are complementary parts of a broader hardware ecosystem.