Performance Rating
AMD Radeon Instinct MI355X
Contents:
Memory ML Performance Compute Power Architecture & Compatibility ML Software Support Clocks & Performance Power Consumption Rendering Benchmarks AdditionalMemory
Memory Size
288
ГБ
Memory Type
Memory Bandwidth
8.19 TB/s
Memory Bus Width
ML Performance
FP16 (Half Precision)
629.1
TFLOPS
BF16 (Brain Float)
TF32 (TensorFloat)
Compute Power
FP32 (Single Precision)
+40,311%
78.64
TFLOPS
FP64 (Double Precision)
78.64
TFLOPS
CUDA Cores
+12,700%
16,384
RT Cores
Architecture & Compatibility
GPU Architecture
SM (Streaming Multiprocessor)
PCIe Version
ML Software Support
CUDA Version
Clocks & Performance
Base Clock
1,000
Boost Clock
2,400
Memory Clock
2,000
Power Consumption
TDP/TGP
-95%
65 W
Recommended PSU
Power Connector
Rendering
Texture Units (TMU)
+12,700%
1,024
ROP
L2 Cache
16 MB
Benchmarks
MLPerf, llama2-70b-99.9 (mxfp4)
MLPerf, llama3.1-405b (mxfp4)
Llama.cpp: Backend: AMD ROCm HIP - Model: GLM-4.7-Flash-IQ4_XS - Test: Prompt Processing 1024
Llama.cpp: Backend: AMD ROCm HIP - Model: GLM-4.7-Flash-IQ4_XS - Test: Prompt Processing 2048
Llama.cpp: Backend: AMD ROCm HIP - Model: GLM-4.7-Flash-IQ4_XS - Test: Prompt Processing 512
Llama.cpp: Backend: AMD ROCm HIP - Model: GLM-4.7-Flash-IQ4_XS - Test: Text Generation 128
Llama.cpp: Backend: AMD ROCm HIP - Model: Llama-3.1-Tulu-3-8B-Q8_0 - Test: Text Generation 128
Llama.cpp: Backend: AMD ROCm HIP - Model: MiniMax-M2.5-UD-TQ1_0 - Test: Prompt Processing 1024
Llama.cpp: Backend: AMD ROCm HIP - Model: MiniMax-M2.5-UD-TQ1_0 - Test: Prompt Processing 2048
Llama.cpp: Backend: AMD ROCm HIP - Model: MiniMax-M2.5-UD-TQ1_0 - Test: Prompt Processing 512
Llama.cpp: Backend: AMD ROCm HIP - Model: MiniMax-M2.5-UD-TQ1_0 - Test: Text Generation 128
Llama.cpp: Backend: AMD ROCm HIP - Model: Qwen3-8B-Q8_0 - Test: Prompt Processing 1024
Llama.cpp: Backend: AMD ROCm HIP - Model: Qwen3-8B-Q8_0 - Test: Prompt Processing 2048
Llama.cpp: Backend: AMD ROCm HIP - Model: Qwen3-8B-Q8_0 - Test: Prompt Processing 512
Llama.cpp: Backend: AMD ROCm HIP - Model: Qwen3-8B-Q8_0 - Test: Text Generation 128
Llama.cpp: Backend: AMD ROCm HIP - Model: gpt-oss-20b-Q8_0 - Test: Prompt Processing 1024
Llama.cpp: Backend: AMD ROCm HIP - Model: gpt-oss-20b-Q8_0 - Test: Prompt Processing 2048
Llama.cpp: Backend: AMD ROCm HIP - Model: gpt-oss-20b-Q8_0 - Test: Prompt Processing 512
Llama.cpp: Backend: AMD ROCm HIP - Model: gpt-oss-20b-Q8_0 - Test: Text Generation 128
Llama.cpp: Backend: Vulkan - Model: GLM-4.7-Flash-IQ4_XS - Test: Prompt Processing 1024
Llama.cpp: Backend: Vulkan - Model: GLM-4.7-Flash-IQ4_XS - Test: Prompt Processing 2048
Llama.cpp: Backend: Vulkan - Model: GLM-4.7-Flash-IQ4_XS - Test: Prompt Processing 512
Llama.cpp: Backend: Vulkan - Model: GLM-4.7-Flash-IQ4_XS - Test: Text Generation 128
Llama.cpp: Backend: Vulkan - Model: Llama-3.1-Tulu-3-8B-Q8_0 - Test: Prompt Processing 1024
Llama.cpp: Backend: Vulkan - Model: Llama-3.1-Tulu-3-8B-Q8_0 - Test: Prompt Processing 2048
Llama.cpp: Backend: Vulkan - Model: Llama-3.1-Tulu-3-8B-Q8_0 - Test: Prompt Processing 512
Llama.cpp: Backend: Vulkan - Model: Llama-3.1-Tulu-3-8B-Q8_0 - Test: Text Generation 128
Llama.cpp: Backend: Vulkan - Model: MiniMax-M2.5-UD-TQ1_0 - Test: Prompt Processing 1024
Llama.cpp: Backend: Vulkan - Model: MiniMax-M2.5-UD-TQ1_0 - Test: Prompt Processing 2048
Llama.cpp: Backend: Vulkan - Model: MiniMax-M2.5-UD-TQ1_0 - Test: Prompt Processing 512
Llama.cpp: Backend: Vulkan - Model: MiniMax-M2.5-UD-TQ1_0 - Test: Text Generation 128
Llama.cpp: Backend: Vulkan - Model: Mistral-7B-Instruct-v0.3-Q8_0 - Test: Text Generation 128
Llama.cpp: Backend: Vulkan - Model: Qwen3-8B-Q8_0 - Test: Prompt Processing 1024
Llama.cpp: Backend: Vulkan - Model: Qwen3-8B-Q8_0 - Test: Prompt Processing 2048
Llama.cpp: Backend: Vulkan - Model: Qwen3-8B-Q8_0 - Test: Prompt Processing 512
Llama.cpp: Backend: Vulkan - Model: Qwen3-8B-Q8_0 - Test: Text Generation 128
Llama.cpp: Backend: Vulkan - Model: gpt-oss-20b-Q8_0 - Test: Prompt Processing 1024
Llama.cpp: Backend: Vulkan - Model: gpt-oss-20b-Q8_0 - Test: Prompt Processing 2048
Llama.cpp: Backend: Vulkan - Model: gpt-oss-20b-Q8_0 - Test: Prompt Processing 512
Llama.cpp: Backend: Vulkan - Model: gpt-oss-20b-Q8_0 - Test: Text Generation 128
Additional
Slots
Release Date
Display Outputs
Renting is cheaper than buying
H100 80Gb SXM (80GB)
NVIDIA
Флагманская GPU на архитектуре Hopper с 80 ГБ памяти HBM3 в форм-факторе SXM. Создана для обучения и инференса больших языковых моделей, HPC задач. Поддерживает FP64, FP32, TF32, FP16, BF16, FP8 и INT8 вычисления с Tensor Cores 4-го поколения.
from 299.11 RUB/hour
