softab

AMD Strix Halo Hardware Specifications

Hardware Target: AMD Ryzen AI Max+ 395 (Strix Halo) with Radeon 8060S iGPU

GPU Compute Specifications

Metric	Value
Architecture	RDNA 3.5 (gfx1151)
Compute Units	40
Max Clock	2.9 GHz
Peak FP16/BF16	59.4 TFLOPS
Peak INT8	59.4 TOPS
Peak INT4	118.8 TOPS
Memory	128GB LPDDR5X-8000 (unified)
Memory Bandwidth	256 GB/s theoretical, ~215 GB/s measured (84% efficiency)

Peak FP16 Calculation: 512 ops/clock/CU × 40 CU × 2.9e9 clock / 1e12 = 59.392 TFLOPS

Comparison with Competition

Spec	Strix Halo 395	Mac Studio M4 Max	DGX Spark	RTX PRO 6000
Price	$2,000	$3,499	$4,000	$8,200
Power (Max W)	120	300+	240	600
Memory	128GB	128GB	128GB	96GB GDDR7
Memory BW	256 GB/s	546 GB/s	273 GB/s	1792 GB/s
FP16 TFLOPS	59.4	34.1	62.5	251.9

Key Advantage: Strix Halo trades bandwidth for capacity—ideal for:

• Huge MoE models (70B+, 235B Q3_K)
• Long-context scenarios (32K+ tokens)
• Data that fits entirely in unified memory

Memory Architecture

Understanding GTT vs GART

• GART: Fixed reserved aperture set in BIOS (set to minimum: 512MB)
• GTT: Dynamically allocatable via TTM subsystem (configure in Linux)

Linux Memory Configuration

Create /etc/modprobe.d/amdgpu-ai.conf:

# GTT allocation via TTM (pages_limit × 4KB = total bytes)
# 31457280 pages × 4KB = 120 GiB
options ttm pages_limit=31457280

# Optional: Pre-allocate to reduce fragmentation
# Set equal to pages_limit for AI-dedicated systems
options ttm page_pool_size=31457280

# Legacy parameter (still set for compatibility)
options amdgpu gttsize=122800

Kernel Parameters

Add to GRUB_CMDLINE_LINUX_DEFAULT:

# Option 1: Maximum memory (kyuz0 toolboxes - 124GB GPU, 4GB OS)
amd_iommu=off              # May improve perf (unverified, see experiment-kernel-config.md)
amdgpu.gttsize=126976      # 124GB GTT size
ttm.pages_limit=32505856   # 124GB in pages (4KB pages)

# Option 2: Alternative full allocation (~128GB GPU)
amd_iommu=off              # May improve perf (unverified, see experiment-kernel-config.md)
amdgpu.gttsize=131072      # 128GB GTT size
ttm.pages_limit=33554432   # ~128GB in pages
amdgpu.cwsr_enable=0       # Fix MES hangs (add if needed)

Model Capacity Guide

Model	Size	Status
70B Q4_K	~37 GiB	✅ Comfortable
109B MoE (Llama 4 Scout)	~58 GiB	✅ Works well
235B MoE Q3_K	~97 GiB	⚠️ Near limit

VRAM Estimation (using kyuz0’s gguf-vram-estimator.py):

# Llama-4-Scout 17B Q4_K at different context lengths
# 4K context:  61.6 GB
# 32K context: 74.8 GB
# 1M context:  108.8 GB

# Qwen3-235B Q3_K (97 GiB model)
# 4K context:  ~97 GiB
# 32K context: ~110 GiB
# 130K context: ~130 GiB max on 128GB system

Performance Baseline

Measured Performance

Workload	Achieved	Theoretical	Utilization
GEMM (hipBLASLt, native gfx1151)	36.9 TFLOPS	59.4 TFLOPS	62%
GEMM (HSA_OVERRIDE gfx1100)	30.9 TFLOPS	59.4 TFLOPS	52%
GEMM (rocBLAS only)	5.1 TFLOPS	59.4 TFLOPS	9%
Flash Attention	~1 TFLOPS	-	Poor

llama.cpp Benchmarks (Llama-2-7B Q4_0, pp512/tg128)

Backend	Prompt (t/s)	Generation (t/s)	Notes
Vulkan AMDVLK	884	53	Best prompt processing
Vulkan RADV	729	55	Slightly better tg, more stable
ROCm HIP baseline	349	49	Without optimizations
ROCm + rocWMMA + FA	344	51	Better at long context
CPU only	294	29	Fallback

Key Finding: Vulkan outperforms ROCm 2-2.5x for prompt processing in most scenarios.

---

Back to: KNOWLEDGE_BASE.md