AMD Instinct MI455X

AMD Instinct MI455X
AMD Instinct MI455X graphics card review

AMD Instinct MI455X: 432 GB of HBM4 and the First DeepSeek-R1 Results

On July 23, 2026, AMD introduced the Instinct MI455X as part of Helios, a server rack for training and running neural networks. The MI455X is a module with direct liquid cooling. Each accelerator has 432 GB of HBM4 memory, while the full rack with 72 GPUs provides a total capacity of approximately 31 TB.

In November, AMD reported that the Helios system achieved 21,333 tokens/s on DeepSeek-R1, calculated per MI455X. On November 12, the company compared this figure with NVIDIA Vera Rubin's result and claimed an advantage of up to 30%. Four days later, Nebius published a new Vera Rubin result on the same model-more than 22,000 tokens/s per GPU on average.

How the MI455X Differs from the MI450X

Both accelerators for Helios are based on the CDNA 5 architecture. AMD specifies the same peak memory bandwidth for both: 23.3 TB/s. Their peak compute figures also match, including 40.3 PFLOPS in the MXFP4 format. The main difference between the models is how much data fits in each GPU's local memory.

Accelerator HBM4 memory per GPU Number of HBM4 stacks
AMD Instinct MI455X 432 GB 12
AMD Instinct MI450X 288 GB 8

The additional 144 GB is especially useful when the model's weights are just slightly too large for the memory of a single accelerator. If the entire copy can fit on an MI455X, the server can process requests without exchanging intermediate results between GPUs. Another way to use the extra memory is to allocate more space for the cache used by long conversations.

Does DeepSeek-R1 Fit in 432 GB?

The main DeepSeek-R1 model contains approximately 671 billion parameters without the additional token-prediction module. When generating a single token, the model uses about 37 billion parameters, but it may select different experts for the next token. Loading the entire model requires the weights of all experts.

  • With ideal packing at four bits per parameter, the weights occupy approximately 335.5 GB (312.5 GiB). This is the theoretical minimum without metadata.
  • In MXFP4, each group of 32 values has a shared scale byte. Storing all parameters in this format increases the calculated size to 356.5 GB (332 GiB).
  • For a single request with a context of 131,072 tokens, DeepSeek-R1's compact MLA cache in BF16 requires an additional approximately 9.2 GB. This estimate is based on the model configuration and does not account for working buffers or program overhead.

According to these calculations, the four-bit weights fit in the nominal memory of a single MI455X and exceed the capacity of a single MI450X. How many long requests an MI455X can serve simultaneously depends on the actual format of the weights and cache. AMD has not disclosed enough details about its configuration to connect the result of 21,333 tokens/s with placing the model on a single GPU.

DeepSeek-R1 on the MI455X, MI355X, and NVIDIA Vera Rubin

In the MLPerf Inference Offline scenario, the system receives all requests at once. The test measures the aggregate generation rate across all requests. In the last column, public MLPerf results show the system rate divided by the number of accelerators. For the MI455X, the per-GPU figure was reported by AMD itself.

System Offline result source Number of GPUs Tokens/s per system Tokens/s per GPU
AMD Helios, Instinct MI455X AMD lab, November Not specified - 21,333
NVIDIA Vera Rubin NVL72 MLPerf v6.1, September, Preview 72 1,183,327 16,435
Nebius, NVIDIA Vera Rubin NVL72 MLPerf v6.1, November, Available 72 1,586,667 22,037
Crusoe, AMD Instinct MI355X MLPerf v6.1, September 512 2,901,950 5,668

For its comparison, AMD used the September Vera Rubin result. Relative to that result, the MI455X figure is indeed approximately 30% higher. In Nebius's November publication, the average per Rubin was already 3.3% higher than AMD's figure. Both differences apply to specific systems. The MI455X result came from an internal test that was not validated by MLCommons. AMD did not specify the number of accelerators or the computational precision.

For an interactive service, another question remains: how many requests the system can handle under a response-time limit. In the Server scenario, Nebius's November rack achieved 1,752,403 tokens/s. For the MI455X, AMD's November announcement included only Offline results, so the response latency of the two systems cannot be compared based on these publications.

What Tasks Is the MI455X For?

The clearest use case for the MI455X is a model that is too large to fit comfortably in 288 GB per GPU. The additional memory may reduce the number of accelerators needed for one model copy or allow the server to handle more long conversations simultaneously. If the same workload fits on the MI450X, the larger model has no stated advantage in HBM4 bandwidth. In that case, the choice depends on the measured speed and the cost of the complete configuration for the required context length and acceptable time to first token.

GPU Comparison

Compare Instinct MI455X with other GPUs

Please select GPU from the dropdown list.

Basic

Label Name
AMD
Platform
Professional
Launch Date
July 2026
Accelerated Compute Dies (XCD)
8
Cooling
Direct Liquid Cooling
Form Factor
EAM
GPU Lithography
TSMC 2nm / 3nm FinFET
I/O Dies (IOD)
2
Work Group Processors (WGP)
256
Model Name
AMD Instinct MI455X
Generation
Instinct MI400 Series
Boost Clock
2400 MHz
Transistors
320 billion
Compute Units
256
Foundry
TSMC
Process Size
2nm / 3nm FinFET
Architecture
CDNA 5

Memory Specifications

HBM Memory Stacks
12
HBM4 Interface Channels
192
Memory ECC
Yes
Memory Partitions
4
Memory Size
432 GB
Memory Type
HBM4
Bandwidth
Memory bandwidth refers to the data transfer rate between the graphics chip and the video memory. It is measured in bytes per second, and the formula to calculate it is: memory bandwidth = working frequency × memory bus width / 8 bits.
23.3 TB/s

Display and Media

AV1 Encode/Decode
AV1 decode (4 shared engines)
H.264 Hardware Encode/Decode
H.264 decode (4 shared engines)
H.265 HEVC Hardware Encode/Decode
H.265 decode (4 shared engines)
JPEG/MJPEG Codec Cores
40
Video Decoder Engines
4 shared engines (H.264, H.265, VP9, AV1)

Theoretical Performance

FP32 Matrix
315 TFLOPS
FP32 Vector
315 TFLOPS
FP16 (half)
An important metric for measuring GPU performance is floating-point computing capability. Half-precision floating-point numbers (16-bit) are used for applications like machine learning, where lower precision is acceptable. Single-precision floating-point numbers (32-bit) are used for common multimedia and graphics processing tasks, while double-precision floating-point numbers (64-bit) are required for scientific computing that demands a wide numeric range and high accuracy.
315 TFLOPS
FP64 (double)
An important metric for measuring GPU performance is floating-point computing capability. Double-precision floating-point numbers (64-bit) are required for scientific computing that demands a wide numeric range and high accuracy, while single-precision floating-point numbers (32-bit) are used for common multimedia and graphics processing tasks. Half-precision floating-point numbers (16-bit) are used for applications like machine learning, where lower precision is acceptable.
5 TFLOPS
FP32 (float)
An important metric for measuring GPU performance is floating-point computing capability. Single-precision floating-point numbers (32-bit) are used for common multimedia and graphics processing tasks, while double-precision floating-point numbers (64-bit) are required for scientific computing that demands a wide numeric range and high accuracy. Half-precision floating-point numbers (16-bit) are used for applications like machine learning, where lower precision is acceptable.
315 TFLOPS

AI Features

AI Software Frameworks Supported by GPU
PyTorch, TensorFlow, JAX, ONNX Runtime, vLLM, SGLang, Triton, Kokkos, RAJA
BF16 Matrix
5033 TFLOPS
BF16 Matrix Sparse
10066 TFLOPS
FP16 Matrix
5033 TFLOPS
FP16 Matrix Sparse
10066 TFLOPS
FP8 Matrix
20133 TFLOPS
INT8 Matrix
5033 TOPS
INT8 Matrix Sparse
10066 TOPS
MXFP4
40265 TFLOPS
MXFP6
20133 TFLOPS
MXFP8
20133 TFLOPS

Interconnect and Scalability

CPU-to-GPU Interconnect
256 GB/s
GPU Interconnect
UALink / UALoE
Scale-out Bandwidth
600 GB/s
Scale-up Bandwidth
3.6 TB/s

Miscellaneous

Virtualization
SR-IOV
L2 Cache
192 MB

Benchmarks

FP32 (float)
Score
315 TFLOPS

Compared to Other GPU

FP32 (float) / TFLOPS
315 +0%
91.375 -71%
80.086 -74.6%
66.228 -79%