Today, we are sharing more details about VEGA and NAVI architectures. It should hopefully help you understand what is coming to professional market in the next 3 years.
Radeon Pro S9 Nano
While consumer Fiji is already EOL, there’s till place for it in the professional market. AMD decided to reuse R9 Nano for servers and call it Radeon Pro S9 Nano. We expect R9 Nano and S9 Nano to share the same specs, but the details are yet unknown. We have learned that this card should launch in few weeks.
VEGA 11
Like I said yesterday, full details of VEGA 11 were not yet disclosed. However, I did tell you that Polaris 10 will be replaced by Vega 11 next year. Of course, what I meant was the professional market. It does not mean there won’t be Polaris 10-based Radeons next year. It means that more powerful solutions will be offered in Radeon Pro series with VEGA GPU.
VEGA vs PASCAL
I decided to compile this chart for you to understand the differences between VEGA and Pascal. Like you can see Vega 10 has two stacks of HBM2 memory, hence 16GB configuration. However, it’s unclear if such configuration will be offered with Radeon RX series. Meanwhile dual Vega 10 will get 4 stacks of HBM2. This will also increase bandwidth from 512 GB/s to 1 TB/s.
Vega 10 has 750 GFLOPs of dual-precision computing power (that’s 1/16 of SP). Interestingly it will change with VEGA20, which should have a divider of 1/2 of SP. For such reason, some GPUs (like Hawaii) that offer good dual-precision performance will still be around next year (check the roadmap below). So as you can see Vega10 will compete with Pascal GP100 in half and single-precision computing.
AMD is also developing its NVLink alternative called xGMI for peer to peer GPU communication. It will be available with Naples architecture based on Zen and VEGA 20, which is currently expected in the second half of 2018.
| AMD VEGA Architecture | |||||
|---|---|---|---|---|---|
| VEGA11 | VEGA10 | VEGA10x2 | VEGA20 | GP100 TESLA P100 | |
| Available | 2017 | 1H 2017 | 2H 2017 | 2H 2018 | ‘Now’ |
| Architecture | 14nm GFX9 | 14 nm GFX9 | 14nm GFX9 | 7nm GFX9 | 16nm Pascal |
| Compute Clusters | ? | 64 | 2x 64 | 64 | 56 |
| Memory | ? | 16GB 2 stacks HBM2 | 4 stacks HBM2 | 16/32 GB 4 stacks HBM2 | 12/16GB 3/4 stacks HBM2 |
| Bandwidth | ? | 512 GB/s | 1 TB/s | 1 TB/s | 540/720 GB/s |
| TBP/TDP | ? | 225W | 300W | 150-300W | 250W |
| Interface | PCI-E 3.0 | PCI-E 3.0 | PCI-E 3.0 | PCI-E 4.0/xGMI | PCI-E 3.0/NVLink |
| Double-precision * | ? | 750 GF (1/16 SP) | ? | 1/2 SP | 4.7 TF |
| Single-precision * | ? | 12 TF | ? | ? | 9.3 TF |
| Half-precision * | ? | 24 TF | ? | ? | 18.7 TF |
| * floating point performance | |||||
The roadmap includes VEGA 10, VEGA 10 x2, VEGA 11, VEGA 20, NAVI 10, NAVI 10 x2 and NAVI 11. This is taken from official presentation. Please note it says SERVER GPU:
| AMD Server GPU Roadmap | |||||
|---|---|---|---|---|---|
| 2016 | 2017 | 2018 | 2019 | ||
| HPC | HAWAII | HAWAII | HAWAII | VEGA20 | VEGA20 |
| DEEPLEARNING | FIJI | POLARIS10 | VEGA10 VEGA10x2 | VEGA10 VEGA10x2 | NAVI10 NAVI10x2 |
| INFERENCE | Radeon Pro S9 Nano | POLARIS10/11 | VEGA11 | VEGA11 | NAVI11 |
