It appears I’ve just discovered first Polaris 11 device in CompuBench database.
Polaris 11 “67FF” with 16 Compute Units
So far Polaris 11 was quite a mystery. The smaller Polaris chip that was shown competing with GTX 950 back in January, was not in the spotlight two months later at Capsaicin event. This silicon is expected to replace Pitcairn-based products with more power efficient solution.
AMD Polaris 11, Device ID 67FF:C8 codenamed “Goose”
The word ‘SKU’ in the post title is intentional, as we know that Polaris 11 will launch in many different configurations.
It’s important, because CompuBench shows active Compute Units, so for instance GTX 970 is shown with 13 active units. Therefore, it does not mean that full Polaris 11 silicon has 16 CUs, but this particular SKU does.
Polaris 11 OpenCL device information (CompuBench) / information according to Khronos
If 16 Compute Units are correct, and Polaris GCN 4.0 architecture still has 64 Stream Processors per CU, then Polaris 11 67FF SKU should feature 1024 unified cores.
You might wonder why it’s only 1024 SP, but let me remind you that new FinFET architectures have much higher transistor density per mm2. For example, Pascal GP100 has twice as many transistors as Maxwell GM200. Plus, AMD’s Polaris is using even smaller fabrication node than Pascal (14nm vs 16nm). In other words, those unified cores (CUDAs/Stream Processors) will be much more efficient than its predecessors.
To illustrate what is known to this day about Polaris and Vega, I made this table.
Please note, specs of Polaris 10 67DF and Polaris 67FF are not necessarily showing full silicons.
| AMD Polaris / Vega GPU Specifications | |||
|---|---|---|---|
| April 8th, 2016 | AMD Vega 10 | AMD Polaris 10 (67DF) | AMD Polaris 11 (67FF) |
| GPU | Vega 10 / Greenland | Polaris 10 / Ellesmere | Polaris 11 / Baffin |
| Positioning | Enthusiast | High-end | Mid-range |
| Fabrication Process | 14nm FinFET | 14nm FinFET | 14nm FinFET |
| Compute Units | 64 | 36+ | 16+ |
| Stream Processors | 4096 | 2304+ | 1024+ |
| Computing Power | ~8.2 TFLOPs | ~3.7 TFLOPs | ~ 2.0 TFLOPs |
| Core clock | ~1000 MHz | ~800 MHz | ~1000 MHz |
| Effective Memory Clock | ~2000 MHz | ~6000 MHz | ~7000 MHz |
| Memory Bus | 4096-bit | 256-bit | 128-bit |
| Memory | 16GB HBM2 | 8GB GDDR5 | 4GB GDDR5 |
| Bandwidth | 1024 GB/s | 192 GB/s | 112 GB/s |
| Launch Date | Q4’16 – Q1’17 | Q2’16 | Q2’16 |
AMD Polaris 11 vs GeForce GTX 950 OpenCL performance
This is the first benchmark of Polaris 11 at CompuBench, so I find it unreasonable to compare those results to anything. Still, just to give you some idea on how it performs at this stage, here’s a comparison between Polaris11 and GTX 950 in OpenCL benchmark:
CES2016 Polaris 11 demonstration
Back in January, when AMD announced Polaris architecture, were were told that Radeon card based on Polaris 11 can offer stable 60 frames per second in Star Wars Battlefront, while the whole system drew just 86W from power socket. Meanwhile system with GeForce GTX 950 delivering the same framerate required 140W.
Here’s a video from PCPerspective showing this demo:
