GPU Wiki Architecture
TeraScale 3
TeraScale 3 (internally designated VLIW4) is the final evolution of AMD's VLIW shader family, used exclusively in the Radeon HD 6900 series (Cayman) launched in late 2010. By reducing the VLIW cluster width from 5 to 4 shader processors, AMD improved instruction packing efficiency — a direct response to compiler difficulty squeezing full utilization out of VLIW5 workloads.
- Launch era
- 2010 to 2012
- Process nodes
- 40nm (TSMC)
- Predecessor
- TeraScale 2 (VLIW5)
- Successor
- GCN 1.0
- Source review
- Verified
Related GPUs
| GPU | Architecture | Launch Date | Shader Processors | Memory | Bus Width | TDP/TBP | Segment |
|---|---|---|---|---|---|---|---|
| Radeon HD 6930 | TeraScale 3 | 2011-12-01 | 1,280 | 1 GB GDDR5 | 256-bit | 186 W | Gaming |
| Radeon HD 6970 | TeraScale 3 | 2010-12-14 | 1,536 | 2 GB GDDR5 | 256-bit | 250 W | Gaming |
| Radeon HD 6990 | TeraScale 3 | 2011-03-08 | 1,536 | 2 GB GDDR5 | 256-bit | 375 W | Gaming |
| Radeon HD 6950 | TeraScale 3 | 2010-12-14 | 1,408 | 2 GB GDDR5 | 256-bit | 200 W | Gaming |
| FirePro V5900 | TeraScale 3 | 2011-05-24 | 512 | 2 GB GDDR5 | 256-bit | 75 W | Workstation |
| FirePro V7900 SDI | TeraScale 3 | 2011-05-24 | 1,280 | 2 GB GDDR5 | 256-bit | 150 W | Workstation |
| FirePro V7900 | TeraScale 3 | 2011-05-24 | 1,280 | 2 GB GDDR5 | 256-bit | 150 W | Workstation |
Architecture History
The HD 6970 and HD 6950 were competitive with NVIDIA's GeForce GTX 580 and represented AMD's last VLIW-based high-end desktop GPU before GCN.
More architecture facts
- Architecture name
- TeraScale 3 (VLIW4)
- Important chips
- Cayman XT (HD 6970), Cayman Pro (HD 6950)
- Memory technologies
- GDDR5 (256-bit)
- Architecture feature generation
- VLIW4 Stream Processors; DirectX 11 (FL 11_0); OpenCL 1.1
- Consumer series
- Radeon HD 6970, HD 6950
- Workstation / professional series
- ATI FirePro V9800 (Cayman)
- Data center series
- (None; compute focus shifted to GCN development)
What this architecture changed
VLIW4 cluster design: each shader processor group handles 4-wide VLIW instructions rather than VLIW5, improving compiler utilization and die efficiency.
Increase in shader processor count (1536 on Cayman XT) relative to Cypress (1600) with a smaller die by switching to VLIW4.
DirectX 11 feature level 11_0.
Side port memory removed compared to some earlier designs; unified GDDR5 bus.
Why it mattered
TeraScale 3 / VLIW4 represents AMD acknowledging the compiler limitations of wide VLIW designs. The move to 4-wide clusters was a stepping stone: it improved utilization without the full redesign that GCN would require. Cayman is historically significant as the last VLIW-architecture high-end AMD GPU, marking the end of a shader design philosophy that traced back to the ATI R300 generation. It is relatively uncommon in collector markets compared to the larger Evergreen lineup.