AMD Instinct MI350X vs AMD Ryzen Z1 Extreme GPU Comparison

Where Each One Wins

The AMD Instinct MI350X and AMD Ryzen Z1 Extreme GPU occupy opposite ends of the AMD graphics spectrum, and the recorded data reflects that split clearly. The MI350X is a data center accelerator built for compute throughput, while the Z1 Extreme is a low-power mobile part designed for gaming handhelds. Neither device has benchmark scores in the database, and both sit at the 50th percentile among all GPUs, so direct performance comparisons rely entirely on architectural and specification differences rather than measured results.

The MI350X wins in every category that matters for large-scale compute workloads. Its FP32 throughput of 72.09 TFLOPS dwarfs the Z1 Extreme's 8.294 TFLOPS, an 8.7x advantage. Texture rate tells a similar story: 2,252.8 GTexel/s versus 129.6 GTexel/s, roughly 17.4x higher. Memory capacity and bandwidth are even more lopsided. The MI350X carries 288 GB of HBM3e across an 8192-bit bus, delivering 8.19 TB/s of bandwidth. The Z1 Extreme uses 16 GB of LPDDR5 on a 64-bit bus, yielding 51.20 GB/s. That is a 160x bandwidth gap and an 18x capacity gap. For training, inference, or any memory-bound workload, the MI350X simply has no competition in this pairing.

The Z1 Extreme wins where power and portability matter. Its TDP is 30 W, versus the MI350X's 1000 W, a 33.3x difference. The Z1 Extreme also has a working display output (1x USB Type-C), while the MI350X has no outputs at all. The Z1 Extreme supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4, making it a functional graphics solution for gaming and consumer applications. The MI350X lists N/A for all three APIs, confirming it is not designed for interactive rendering. The Z1 Extreme is also the only one with an active production status and a launch MSRP of 699 USD.

The use-case split is therefore stark. The MI350X is a compute accelerator with no display path, no graphics API support, and an OAM module form factor. The Z1 Extreme is a complete GPU solution with graphics APIs, a display output, and a compact 280 mm length. The MI350X wins on raw compute, memory, and bandwidth. The Z1 Extreme wins on efficiency, graphics capability, and usability as a consumer device.

FAQ

Q: Which GPU has higher FP32 compute performance?

A: The AMD Instinct MI350X delivers 72.09 TFLOPS FP32, while the AMD Ryzen Z1 Extreme GPU delivers 8.294 TFLOPS. The MI350X is approximately 8.7x higher based on the recorded figures.

Q: How do memory capacities compare?

A: The MI350X has 288 GB of HBM3e memory, whereas the Z1 Extreme has 16 GB of LPDDR5. The MI350X also uses an 8192-bit bus with 8.19 TB/s bandwidth, compared to the Z1 Extreme's 64-bit bus and 51.20 GB/s bandwidth.

Q: Which GPU supports DirectX?

A: Only the AMD Ryzen Z1 Extreme GPU supports DirectX, specifically DirectX 12 Ultimate (12_2). The AMD Instinct MI350X lists DirectX as N/A and also has no OpenGL or Vulkan support.

Q: What are the power requirements?

A: The MI350X has a TDP of 1000 W and a suggested PSU of 1400 W. The Z1 Extreme has a TDP of 30 W. The Z1 Extreme has no suggested PSU listed, while the MI350X requires none for power connectors since it uses an OAM module.

Q: Does either GPU have a display output?

A: The AMD Ryzen Z1 Extreme GPU has 1x USB Type-C display output. The AMD Instinct MI350X has no display outputs.

Q: Which GPU has more shading units?

A: The AMD Instinct MI350X has 16,384 shading units, while the AMD Ryzen Z1 Extreme GPU has 768 shading units. The MI350X also has 1,024 TMUs versus the Z1 Extreme's 48 TMUs.

Head-to-Head Benchmarks

The database contains no recorded benchmark scores for either GPU, so head-to-head comparisons must be drawn from the specification differences that directly determine performance. The largest single gap is memory bandwidth. The MI350X's 8.19 TB/s is 160x the Z1 Extreme's 51.20 GB/s. This is the defining characteristic of the MI350X as a compute accelerator; massive memory throughput is essential for AI training and inference workloads that stream data continuously to compute units.

FP32 compute shows the second-largest gap. The MI350X's 72.09 TFLOPS is 8.7x the Z1 Extreme's 8.294 TFLOPS. Notably, the MI350X achieves FP16 at the same rate as FP32 (72.09 TFLOPS, 1:1 ratio), while the Z1 Extreme doubles its FP32 rate for FP16, reaching 16.59 TFLOPS at a 2:1 ratio. Even with that doubling, the Z1 Extreme's FP16 peak of 16.59 TFLOPS remains well below the MI350X's FP32 peak.

Texture rate amplifies the compute difference. The MI350X reaches 2,252.8 GTexel/s, which is 17.4x the Z1 Extreme's 129.6 GTexel/s. Pixel rate, however, reverses the trend: the MI350X is listed at 0 MPixel/s, while the Z1 Extreme delivers 86.40 GPixel/s. The MI350X has 0 ROPs, so it cannot rasterize pixels at all. The Z1 Extreme has 32 ROPs, confirming it is the only one of the two that can output a rendered frame.

Clock speeds tell a different story. The Z1 Extreme boosts to 2700 MHz, 500 MHz higher than the MI350X's 2200 MHz boost. The Z1 Extreme also has a lower base clock at 800 MHz compared to the MI350X's 1000 MHz. Higher boost clocks on the Z1 Extreme do not compensate for the massive differences in shading units, TMUs, and memory bus width. The MI350X's 16,384 shading units and 1,024 TMUs simply outclass the Z1 Extreme's 768 shading units and 48 TMUs.

The release dates place these products two years apart. The Z1 Extreme launched on 2023-06-12, and the MI350X launched on 2025-06-11. Both share the TSMC foundry, but the MI350X uses a 3 nm process versus the Z1 Extreme's 4 nm process. Transistor counts differ by an order of magnitude: 185,000 million for the MI350X versus 25,390 million for the Z1 Extreme. Die size is similarly lopsided at 2380 mm² versus 178 mm². The transistor density, however, favors the Z1 Extreme at 142.6M / mm² versus the MI350X's 77.7M / mm², a reflection of the different design goals and memory packaging of the two chips.

Specification Differences

The two GPUs differ in nearly every measurable specification. The MI350X uses the MI350 256CU chip, while the Z1 Extreme uses the Phoenix chip. Architecture differs completely: CDNA 4.0 for the MI350X, RDNA 3.0 for the Z1 Extreme. Process nodes are 3 nm and 4 nm respectively, both from TSMC.

Memory is the most dramatic difference. The MI350X has 288 GB of HBM3e on an 8192-bit bus with 8.19 TB/s bandwidth. The Z1 Extreme has 16 GB of LPDDR5 on a 64-bit bus with 51.20 GB/s bandwidth. Memory clock also differs: the MI350X runs at 2000 MHz (8 Gbps effective), while the Z1 Extreme runs at 800 MHz (6.4 Gbps effective).

Compute resources differ by orders of magnitude. The MI350X has 16,384 shading units, 1,024 TMUs, and 0 ROPs. The Z1 Extreme has 768 shading units, 48 TMUs, 32 ROPs, and 12 ray tracing cores. The MI350X has no RT cores listed. Pixel rate is 0 MPixel/s for the MI350X and 86.40 GPixel/s for the Z1 Extreme. Texture rate is 2,252.8 GTexel/s versus 129.6 GTexel/s.

Power and physical specifications diverge completely. The MI350X has a 1000 W TDP and a 1400 W suggested PSU, uses an OAM module slot width, and has no power connectors. The Z1 Extreme has a 30 W TDP, no suggested PSU, no listed slot width, and no power connectors. Dimensions differ: the MI350X is 102 mm long and 165 mm wide, while the Z1 Extreme is 280 mm long, 111 mm high, and 21 mm wide.

Interface and output capabilities are mutually exclusive. The MI350X uses PCIe 5.0 x16 and has no display outputs. The Z1 Extreme has no bus interface listed and offers 1x USB Type-C display output. API support is absent for the MI350X (DirectX, OpenGL, and Vulkan all N/A) and fully present for the Z1 Extreme (DirectX 12 Ultimate, OpenGL 4.6, Vulkan 1.4). The Z1 Extreme is the only one with an active production status and carries a launch MSRP of 699 USD. The MI350X has no launch MSRP recorded.

Architecture Differences

The architectural divide between the two chips is fundamental. The MI350X is built on CDNA 4.0, AMD's compute-focused architecture for data center accelerators. The Z1 Extreme uses RDNA 3.0, AMD's graphics-focused architecture for consumer and console GPUs. This single difference explains most of the behavioral divergence: CDNA 4.0 prioritizes FP32 and FP16 throughput with massive memory bandwidth, while RDNA 3.0 prioritizes rasterization, ray tracing, and power efficiency.

The MI350X packs 185,000 million transistors into a 2380 mm² die, using a 3 nm TSMC process. The Z1 Extreme contains 25,390 million transistors on a 178 mm² die, using a 4 nm TSMC process. The MI350X has a transistor density of 77.7M / mm², while the Z1 Extreme reaches 142.6M / mm². The MI350X's lower density reflects the large amount of die area devoted to HBM3e memory controllers and the 8192-bit bus, which consume substantial space without adding logic density.

Compute feature sets diverge sharply. The MI350X has 16,384 shading units and 1,024 TMUs, with no ROPs and no RT cores. The Z1 Extreme has 768 shading units, 48 TMUs, 32 ROPs, and 12 RT cores. The MI350X is therefore incapable of pixel output, which is consistent with its 0 MPixel/s pixel rate. The Z1 Extreme's 86.40 GPixel/s pixel rate and 12 RT cores make it a complete graphics processor for games and consumer applications.

FP16 behavior highlights the architectural priorities. The MI350X runs FP16 at 72.09 TFLOPS with a 1:1 ratio to FP32, meaning it does not double throughput for reduced precision. The Z1 Extreme runs FP16 at 16.59 TFLOPS with a 2:1 ratio, doubling its FP32 rate. This suggests the MI350X is optimized for sustained FP32 and FP16 compute without shader reconfiguration, while the Z1 Extreme uses the RDNA 3.0 shader design to extract extra FP16 throughput.

Clock behavior also reflects the different design targets. The MI350X has a 1000 MHz base clock and 2200 MHz boost, prioritizing stability and sustained throughput in a 1000 W power envelope. The Z1 Extreme has an 800 MHz base clock and 2700 MHz boost, allowing aggressive single-core and low-load boosting within a 30 W power envelope. The higher boost clock on the Z1 Extreme is a classic mobile design choice, while the MI350X's lower boost clock keeps thermals manageable across 16,384 active shading units.

Memory architecture is the final differentiator. The MI350X uses HBM3e, a stacked memory technology, with an 8192-bit bus that enables 8.19 TB/s bandwidth. The Z1 Extreme uses LPDDR5, a low-power DRAM, with a 64-bit bus and 51.20 GB/s bandwidth. The MI350X's memory subsystem is designed for data center workloads that constantly stream large datasets, while the Z1 Extreme's memory is sized and powered for handheld gaming. The API support gap is the logical conclusion of these choices: CDNA 4.0 exposes no graphics APIs, while RDNA 3.0 supports DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.4.