Radeon Instinct MI350P vs H200 NVL

Primary details

GPU architecture, market segment, value for money and other general parameters compared.

Place in the ranking	not rated	not rated
Place by popularity	not in top-100	not in top-100
Architecture	Hopper (2022−2024)	CDNA 4.0 (2025−2026)
GPU code name	GH100	MI350 128CU
Market segment	Workstation	Workstation
Release date	18 November 2024 (1 year ago)	7 May 2026 (less than a year ago)

Detailed specifications

General parameters such as number of shaders, GPU core base clock and boost clock speeds, manufacturing process, texturing and calculation speed. Real power consumption of some graphics cards can well exceed their nominal TDP, especially if overclocked.

Pipelines / CUDA cores	16896	8192
Core clock speed	1365 MHz	1000 MHz
Boost clock speed	1785 MHz	2200 MHz
Number of transistors	80,000 million	73,000 million
Manufacturing process technology	5 nm	3 nm
Power consumption (TDP)	600 Watt	600 Watt
Texture fill rate	942.5	1,126.4
Floating-point processing power	60.32 TFLOPS	36.04 TFLOPS
ROPs	24	no data
TMUs	528	512
Tensor Cores	528	no data
L1 Cache	33 MB	2 MB
L2 Cache	50 MB	16 MB
L3 Cache	no data	128 MB

Form factor & compatibility

Information on compatibility with other computer components. Useful when choosing a future computer configuration or upgrading an existing one. For desktop graphics cards it's interface and bus (motherboard compatibility), additional power connectors (power supply compatibility).

Interface	PCIe 5.0 x16	PCIe 5.0 x16
Length	267 mm	267 mm
Width	2-slot	2-slot
Supplementary power connectors	8-pin EPS	1x 16-pin

VRAM capacity and type

Parameters of VRAM installed: its type, size, bus, clock and resulting bandwidth. Integrated GPUs have no dedicated video RAM and use a shared part of system RAM.

Memory type	HBM3e	HBM3e
Maximum RAM amount	141 GB	144 GB
Memory bus width	6144 Bit	8192 Bit
Memory clock speed	1593 MHz	2000 MHz
Memory bandwidth	4.89 TB/s	8.19 TB/s
Resizable BAR	+	-

Connectivity and outputs

This section shows the types and number of video connectors on each GPU. The data applies specifically to desktop reference models (for example, NVIDIA’s Founders Edition). OEM partners often modify both the number and types of ports. On notebook GPUs, video‐output options are determined by the laptop’s design rather than the graphics chip itself.

Display Connectors

No outputs

No outputs

API and SDK support

List of supported 3D and general-purpose computing APIs, including their specific versions.

DirectX	N/A	N/A
Shader Model	N/A	N/A
OpenGL	N/A	N/A
OpenCL	3.0	3.0
Vulkan	N/A	N/A
CUDA	9.0	-
DLSS	+	-

Pros & cons summary

---

Recency	18 November 2024	7 May 2026
Maximum RAM amount	141 GB	144 GB
Chip lithography	5 nm	3 nm