inference/research
verifiedmedium · 0.75updated 2026-10-17

AMD Radeon VII

amd vega 20 (GCN 5.1, gfx906) 2019

Fields

id
accelerators/amd-radeon-vii
Vendor
amd
Architecture
vega 20 (GCN 5.1, gfx906)
Release year
2019
Process (nm)
7
Form factors
pcie
VRAM (GB)
16
Memory type
hbm2
Memory bus (bit)
4,096
Memory bandwidth (GB/s)
1,024
Bandwidth basis
AMD product-page spec table (Wayback capture): Memory Bandwidth up to 1024 GB/s with Memory Interface 4096-bit. The same table prints Memory Speed 4 Gbps, and 4 Gbps x 4096-bit / 8 = 2048 GB/s, which disagrees with the printed bandwidth; the bandwidth row is stored and the disagreement is left unresolved.
Peak FLOPS
precisionfp16tflops27.7denseyesvendor_claimyessource_idamd-radeon-vii-product-page-archive-2022basisvendor-spec-sheetbasis_detailAMD product-page row Peak Half Precision Compute Performance 27.7 TFLOPs. No sparsity footnote is printed on the page, so dense:true.precisionfp32tflops13.8denseyesvendor_claimyessource_idamd-radeon-vii-product-page-archive-2022basisvendor-spec-sheetbasis_detailAMD product-page row Peak Single Precision Compute Performance 13.8 TFLOPs. No sparsity footnote, so dense:true.precisionfp64tflops3.46denseyesvendor_claimyessource_idamd-radeon-vii-product-page-archive-2022basisvendor-spec-sheetbasis_detailAMD product-page row Peak Double Precision Compute Performance 3.46 TFLOPs. 3.46 / 13.8 = 1:4 FP64, the Vega 20 rate, distinct from the 1:2 rate printed for Radeon Pro VII.
TDP (W)
300
Unified memory
no
Consumer part
yes
notes
Consumer Vega 20 card: 60 compute units, 3840 stream processors, 13.2 billion transistors, base 1400 MHz, boost up to 1750 MHz, peak up to 1800 MHz, typical board power 300 W, launched February 2019. All figures come from the Wayback capture of the AMD product page because the live URL 404s. BANDWIDTH DISAGREEMENT: the page prints both Memory Speed 4 Gbps and Memory Bandwidth up to 1024 GB/s on a 4096-bit bus, and those two cannot both be true (4 Gbps x 4096 / 8 = 2048 GB/s). The bandwidth row is what is stored; the speed row is recorded here and not silently corrected. gfx906 is shared with Radeon Pro VII and Instinct MI50/MI50X, so a card reporting gfx906 is not necessarily this SKU. ROCm support for gfx906 has been dropped from current releases, which limits which engines can actually reach the 16 GB of HBM2. 16 GB holds a 7B-13B Q4 and not a 34B Q4.

Notes

Consumer Vega 20 card: 60 compute units, 3840 stream processors, 13.2 billion transistors, base 1400 MHz, boost up to 1750 MHz, peak up to 1800 MHz, typical board power 300 W, launched February 2019. All figures come from the Wayback capture of the AMD product page because the live URL 404s. BANDWIDTH DISAGREEMENT: the page prints both Memory Speed 4 Gbps and Memory Bandwidth up to 1024 GB/s on a 4096-bit bus, and those two cannot both be true (4 Gbps x 4096 / 8 = 2048 GB/s). The bandwidth row is what is stored; the speed row is recorded here and not silently corrected. gfx906 is shared with Radeon Pro VII and Instinct MI50/MI50X, so a card reporting gfx906 is not necessarily this SKU. ROCm support for gfx906 has been dropped from current releases, which limits which engines can actually reach the 16 GB of HBM2. 16 GB holds a 7B-13B Q4 and not a 34B Q4.

Cross-references

Interconnect (1)