Workload and traffic model
Fix precision and workload shape before comparing compute rates. Track external-memory bytes, shared-memory reuse and occupancy separately; a peak arithmetic rate is only one bound on useful throughput.
AMD / SPECIMEN
Polysilicon / lower-layer view of AMD Radeon VII · Vega20, photographed by FritzchensFritz. Identification follows the linked photographer's caption. This is a documented photographic specimen, not a claim that every product with this name uses identical silicon. The original page contains the preparation, microscopy and capture details.

Original source · Retrieved not recorded
FritzchensFritz · photographer · Included in this image
Attribution and license link retained for every photograph; separate per-file source metadata and acquisition hashes preserved.
Photographed specimen identified by its creator as AMD Radeon VII · Vega20. Polysilicon / lower-layer view. This image is not assigned to other SKUs or used to infer die dimensions.
| Identity & fabrication | |
|---|---|
| Process label | Published analysisWikimedia Commons · FritzchensFritz Scope: physical-specimen Photographer-supplied label, not an independently verified foundry disclosure. |
| Additional specifications | |
| Photographed specimen | Published analysisWikimedia Commons · FritzchensFritz Scope: physical-specimen |
| Photographic view | Published analysisWikimedia Commons · FritzchensFritz Scope: physical-specimen |
Start with documented specifications, then use the study prompts below. The prompts are engineering guidance, not a claim about this product’s private RTL or physical implementation.
Fix precision and workload shape before comparing compute rates. Track external-memory bytes, shared-memory reuse and occupancy separately; a peak arithmetic rate is only one bound on useful throughput.
Choose a representative producer/consumer path and specify acceptance, ordering, backpressure, outstanding capacity and reset recovery. Turn these contracts into assertions and test scenarios before optimizing the datapath.
Obtain macro abstracts, pin maps, timing libraries and activity-based power estimates. Explore macro proximity, routing channels and power-grid reachability together. An annotated die photograph can suggest study questions, but cannot supply a netlist, PDK or signoff constraints.
Polysilicon / lower-layer view of AMD Radeon VII · Vega20, photographed by FritzchensFritz. Identification follows the linked photographer's caption. This is a documented photographic specimen, not a claim that every product with this name uses identical silicon. The original page contains the preparation, microscopy and capture details.
This record has identification and source evidence. A deeper architectural article has not yet been written. Visible patterns alone do not establish block functions, capacity or an internal protocol.
Wikimedia Commons · FritzchensFritz
Photographed specimen identified by its creator as AMD Radeon VII · Vega20. Polysilicon / lower-layer view. This image is not assigned to other SKUs or used to infer die dimensions.
Link not yet checked
View original die photoAccess to a source does not grant permission to reproduce its images.
AMD@7nm@GCN 5th gen@Vega20@Radeon VII@ - @ DSCx2 polysilicon layer@IR Makro.jpg ↗ · Wikimedia Commons · FritzchensFritz