AMD Instinct MI325X vs Lisuan Tech LX ULTRA Comparison

FAQ

Q: What are the core architectural identities of these two accelerators?

A: The AMD Instinct MI325X uses the CDNA 3.0 architecture with the Aqua Vanjaram chip, built on a 5 nm TSMC process. The Lisuan Tech LX ULTRA uses the TrueGPU architecture with the 7G105 chip, built on a 6 nm TSMC process.

Q: How do the memory subsystems compare?

A: The MI325X carries 256 GB of HBM3e on an 8192-bit bus, delivering 6.14 TB/s of bandwidth. The LX ULTRA carries 24 GB of GDDR6 on a 192-bit bus, delivering 432.0 GB/s of bandwidth.

Q: Which device has higher raw FP32 throughput?

A: The MI325X delivers 81.72 TFLOPS of FP32 compute, while the LX ULTRA delivers 24.58 TFLOPS. The MI325X is 3.3 times higher in this metric.

Q: What are the power requirements?

A: The MI325X has a TDP of 1000 W and a suggested PSU of 1400 W. The LX ULTRA has a TDP of 225 W and a suggested PSU of 550 W.

Q: Are there display outputs on either card?

A: The MI325X has no display outputs. The LX ULTRA provides 4x DisplayPort 1.4a outputs.

Q: What is the release timeline?

A: The MI325X was released on 2024-10-09. The LX ULTRA has a release date of 2026-03-16 and is marked as Active in production status.

Architecture Differences

The AMD Instinct MI325X and the Lisuan Tech LX ULTRA represent two fundamentally different design philosophies. The MI325X is a data-center compute accelerator built around the CDNA 3.0 architecture, implemented with the Aqua Vanjaram chip on a 5 nm TSMC process. It integrates 153,000 million transistors on a 1017 mm² die, yielding a transistor density of 150.4M per mm². The LX ULTRA uses the TrueGPU architecture with the 7G105 chip on a 6 nm TSMC process; its transistor count and die size are not recorded in the database.

Compute resource allocation diverges sharply. The MI325X has 19,456 shading units, 1,216 texture mapping units, and no ROPs, which aligns with its role as a compute-focused accelerator with a pixel rate of 0 MPixel/s. The LX ULTRA has 6,144 shading units, 192 TMUs, and 96 ROPs, producing a pixel rate of 192.0 GPixel/s. Neither device lists ray tracing cores or tensor cores in the recorded data. Texture throughput favors the MI325X at 2,553.6 GTexel/s versus 384.0 GTexel/s for the LX ULTRA.

Memory architecture is a defining difference. The MI325X uses 256 GB of HBM3e across an 8192-bit bus, achieving 6.14 TB/s bandwidth. The LX ULTRA uses 24 GB of GDDR6 across a 192-bit bus, achieving 432.0 GB/s. The memory clock differs as well: the MI325X runs at 1500 MHz with 6 Gbps effective, while the LX ULTRA runs at 2250 MHz with 18 Gbps effective. The MI325X has no listed base or boost clocks for the GPU core; the LX ULTRA also lacks listed core clocks.

The MI325X uses an OAM module slot width with no power connectors and no display outputs. The LX ULTRA is a dual-slot card with a 1x 16-pin power connector and 4x DisplayPort 1.4a outputs. The bus interface differs: PCIe 5.0 x16 for the MI325X versus PCIe 4.0 x16 for the LX ULTRA. API support also separates them. The MI325X lists DirectX, OpenGL, and Vulkan as N/A, while the LX ULTRA supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.3.

Where Each One Wins

The recorded data indicates a clean separation of use cases. The AMD Instinct MI325X wins decisively in memory capacity, memory bandwidth, FP32 throughput, and texture throughput. Its 256 GB HBM3e pool with 6.14 TB/s bandwidth positions it for large-scale compute workloads where memory residency and bandwidth dominate. The FP32 figure of 81.72 TFLOPS, combined with FP16 at 81.72 TFLOPS (1:1), shows a device optimized for sustained compute throughput without consumer-oriented graphics features.

The Lisuan Tech LX ULTRA wins in areas tied to graphics output and physical integration. It offers 4x DisplayPort 1.4a outputs, a full API stack with DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.3, and a 192.0 GPixel/s pixel rate. Its dual-slot form factor, 268 mm length, 112 mm height, and 40 mm width make it a physically installable card in standard systems. Its 225 W TDP and 550 W suggested PSU indicate a lower system-level power footprint compared to the MI325X.

The FP16 behavior differs meaningfully. The MI325X runs FP16 at 1:1 ratio with FP32, both at 81.72 TFLOPS. The LX ULTRA runs FP16 at 2:1 ratio, reaching 49.15 TFLOPS against 24.58 TFLOPS FP32. This means the MI325X offers symmetric FP32 and FP16 throughput, while the LX ULTRA provides higher relative FP16 performance within its own compute envelope.

Specification Differences

The two devices differ across nearly every recorded specification.

• Process node: 5 nm (TSMC) for the MI325X, 6 nm (TSMC) for the LX ULTRA.
• Transistors: 153,000 million for the MI325X, unknown for the LX ULTRA.
• Die size: 1017 mm² for the MI325X, unknown for the LX ULTRA.
• Transistor density: 150.4M / mm² for the MI325X, null for the LX ULTRA.
• Memory size: 256 GB for the MI325X, 24 GB for the LX ULTRA.
• Memory type: HBM3e for the MI325X, GDDR6 for the LX ULTRA.
• Memory bus width: 8192 bit for the MI325X, 192 bit for the LX ULTRA.
• Memory bandwidth: 6.14 TB/s for the MI325X, 432.0 GB/s for the LX ULTRA.
• Memory clock: 1500 MHz (6 Gbps effective) for the MI325X, 2250 MHz (18 Gbps effective) for the LX ULTRA.
• Shading units: 19,456 for the MI325X, 6,144 for the LX ULTRA.
• TMUs: 1,216 for the MI325X, 192 for the LX ULTRA.
• ROPs: 0 for the MI325X, 96 for the LX ULTRA.
• Pixel rate: 0 MPixel/s for the MI325X, 192.0 GPixel/s for the LX ULTRA.
• Texture rate: 2,553.6 GTexel/s for the MI325X, 384.0 GTexel/s for the LX ULTRA.
• FP32: 81.72 TFLOPS for the MI325X, 24.58 TFLOPS for the LX ULTRA.
• FP16: 81.72 TFLOPS (1:1) for the MI325X, 49.15 TFLOPS (2:1) for the LX ULTRA.
• TDP: 1000 W for the MI325X, 225 W for the LX ULTRA.
• Slot width: OAM Module for the MI325X, Dual-slot for the LX ULTRA.
• Power connectors: None for the MI325X, 1x 16-pin for the LX ULTRA.
• Suggested PSU: 1400 W for the MI325X, 550 W for the LX ULTRA.
• Bus interface: PCIe 5.0 x16 for the MI325X, PCIe 4.0 x16 for the LX ULTRA.
• Display outputs: No outputs for the MI325X, 4x DisplayPort 1.4a for the LX ULTRA.
• DirectX: N/A for the MI325X, 12 Ultimate (12_2) for the LX ULTRA.
• OpenGL: N/A for the MI325X, 4.6 for the LX ULTRA.
• Vulkan: N/A for the MI325X, 1.3 for the LX ULTRA.
• Dimensions: not recorded for the MI325X; 268 mm x 112 mm x 40 mm for the LX ULTRA.
• Production status: not recorded for the MI325X; Active for the LX ULTRA.
• Release date: 2024-10-09 for the MI325X, 2026-03-16 for the LX ULTRA.

Head-to-Head Benchmarks

The database contains no direct head-to-head benchmark entries for this pair, and no individual benchmark scores are recorded for either device. The comparison therefore rests on the recorded specification data and the derived compute ratios.

The largest single-metric advantage belongs to the MI325X in memory bandwidth. Its 6.14 TB/s is 14.2 times the LX ULTRA's 432.0 GB/s. This gap stems from the combination of HBM3e technology, the 8192-bit bus, and the larger memory pool. For workloads that stream large data sets, the MI325X holds a dominant positional advantage.

Memory capacity shows a 10.7 times difference: 256 GB versus 24 GB. The MI325X can hold over ten times more data on-device, which matters for models or datasets that exceed the LX ULTRA's capacity. The LX ULTRA must rely on host-side transfers or smaller working sets.

FP32 throughput gives the MI325X a 3.3 times advantage at 81.72 TFLOPS versus 24.58 TFLOPS. Texture rate shows a 6.65 times gap: 2,553.6 GTexel/s versus 384.0 GTexel/s. Shading unit count is 3.17 times higher on the MI325X (19,456 versus 6,144), and TMU count is 6.33 times higher (1,216 versus 192).

The LX ULTRA counters in graphics-specific metrics. Its 192.0 GPixel/s pixel rate contrasts with the MI325X's 0 MPixel/s, which reflects the MI325X having no ROPs. The LX ULTRA also holds the advantage in API compatibility, with DirectX 12 Ultimate, OpenGL 4.6, and Vulkan 1.3 support versus N/A entries for the MI325X. Display connectivity follows the same pattern: 4x DisplayPort 1.4a on the LX ULTRA, no outputs on the MI325X.

FP16 scaling differs between the two. The MI325X delivers 81.72 TFLOPS at FP16 with a 1:1 ratio to FP32. The LX ULTRA delivers 49.15 TFLOPS at FP16 with a 2:1 ratio, meaning its FP16 throughput is double its FP32 throughput. In absolute terms, the MI325X still leads FP16 by 1.66 times.

Power efficiency reverses the raw performance picture. The LX ULTRA operates at 225 W TDP, while the MI325X operates at 1000 W. The suggested PSU figures follow: 550 W for the LX ULTRA, 1400 W for the MI325X. The LX ULTRA delivers its 24.58 TFLOPS FP32 within a 225 W envelope, while the MI325X delivers 81.72 TFLOPS within a 1000 W envelope.

The physical form factors reflect their intended environments. The MI325X is an OAM module with no power connectors, designed for chassis-level integration. The LX ULTRA is a dual-slot card with a 1x 16-pin connector, 268 mm length, 112 mm height, and 40 mm width, suitable for standard expansion slots. The bus interfaces also differ: PCIe 5.0 x16 on the MI325X, PCIe 4.0 x16 on the LX ULTRA.

The production status field shows Active for the LX ULTRA and no recorded status for the MI325X. Release dates place the MI325X in October 2024 and the LX ULTRA in March 2026, a sequential timeline that does not affect the benchmark comparison but establishes their respective market appearances.