GCN 5.0 | PC Tech Vault

Vega offered strong FP16 and FP64 compute throughput and substantial HBM2 frame buffers, making it popular for GPU compute workloads and cryptocurrency mining, though it consumed more power than competing NVIDIA Pascal products at similar performance levels.

More architecture facts

Architecture name: GCN 5.0 (Vega)
Important chips: Vega 10 (RX Vega 64 / Vega 56), Vega 12 (Radeon Pro Vega — mobile), Vega 20 (Radeon Pro VII — 7nm variant, covered under GCN 5.1)
Memory technologies: HBM2 (Vega 10: up to 8 GB at 484 GB/s)
Architecture feature generation: GCN 5.0 Stream Processors; Next-Generation Compute Units (NCU); DirectX 12 (FL 12_0); Vulkan 1.1; OpenCL 2.0; Rapid Packed Math (FP16 at 2x rate)
Consumer series: Radeon RX Vega 64, RX Vega 56, RX Vega 64 Liquid Cooled
Workstation / professional series: Radeon Pro WX 8200 (Vega 56 derivative), Radeon Pro Vega 48/56 (Mac Pro 2019)
Data center series: Radeon Instinct MI25 (Vega 10)

What this architecture changed

HBM2 memory: up to 8 GB at 484 GB/s on Vega 10, providing substantially more bandwidth than GDDR5X competitors.
Rapid Packed Math: FP16 operations at 2x the rate of FP32, enabling faster compute on half-precision workloads.
Next Compute Units (NCUs): redesigned Compute Units with improved instruction cache and prefetch.
High Compute Unit count (64 CUs on Vega 10) targeting professional and compute-oriented workloads.
Vulkan 1.1 compliance.

Why it mattered

Vega occupied AMD's high end at a time when GCN was approaching architectural maturity on 14nm. Its HBM2 integration and compute focus made it a preferred platform for early machine learning researchers who needed AMD hardware, and its large frame buffers made it useful for professional visualization. The generation's higher power draw relative to Pascal was a recognized weakness, and it accelerated AMD's development of the cleaner RDNA architecture.