AMD recently launched a brand new page dedicated to the upcoming Polaris architecture and posted some pictures teasing the new GPUs. Our collegues over at Videocardz were able to make quick work of it and produce a die shot as well as an estimate of the approximate die size of the flagship Polaris GPU. The size corresponds to the number we heard about three few months ago: 232mm² large die for Polaris 10.
AMD teases die shots of Polaris 10 and Polaris 11
AMD has teased die shots of its upcoming GPUs in the past, Fiji being the most recent example and they were successfully analyzed to give an accurate representation of not only the die but its components as well (anyone remember the time HBM was still one of our “rumors”). In what appears to be history repeating itself, AMD has published teasers of both Polaris 10 ad Polaris 11 on its new page. Given below are the die reconstructions courtesy of Videocardz:
The partial but clear, reconstructed shot is that of Polaris 10 (aka Ellesmere XT). The die appears to be roughly 232mm² which is the exact same number we saw before. This makes it highly unlikely that we are looking at sheer coincidence and probably means that Polaris 10 GPU is, in fact, of that size. We are talking about the Samsung/GloFo 14nm LPP process here, so the size would be optimal in terms of yield and power efficiency. The second die (blurry and full shot) is Polaris 11 (aka Baffin XT). Polaris 11 is the smaller GPU which will be targeting the mobile GPU market and should offer the sweet spot of the Polaris efficiency claim (2.5x increase in perf/watt).
Not to scale.
According to the information we have about the 14nm LPP process, and based on transistor density increase, a 232mm² GPU would be roughly equivalent to a 464mm² 28nm processor – at the same TDP levels. Since we already know that AMD is going to be focusing not just on performance but power efficiency as well – this number could be be much higher, in fact we will discuss the number AMD is using below. We can however safely say that this die is more than capable of meeting the ‘minimum VR spec that AMD promises.
The more astute of our readers would notice one inconsistency with this information. A 232mm² does not fit with the positioning of the small Polaris die, which according to our estimates should be around 150mm² and doesn’t fit perfectly with our image of big Polaris which should have been near the 300mm² mark. If you were to take a look at the AMD slide given above and calculate the die shrinks associated with the 14nm jump – the number would be exactly 2.5. If you remember, Raja Koduri has previously tweeted about Polaris being “2.5 times brighter” in the past – so the number 2.5 times is one AMD is throwing around alot.
The performance levels of a 580mm² can be gained by a 232mm² die (at a very rough estimate) if we are talking bout a gain of 2.5 times in performance per watt. This is the theoretical peak, so the real number would probably be lower than this for Polaris 10 accounting for leakage when clocking at high rates. Also, we do not know what kind of power efficiency mix AMD is playing with here so it would be pointless to speculate further about the exact performance levels at this point.
AMD R9 400 Series 'Preliminary' Specifications
| WCCFTech | Radeon R9 480X | Radeon R9 480 | Radeon R9 470X | Radeon R9 470 |
|---|---|---|---|---|
| GPU | Polaris 10 / Ellesmere XT | Polaris 10 / Ellesmere Pro | Polaris 11 / Baffin XT | Polaris 11 / Baffin Pro |
| Fabrication Process | 14nm FinFET | 14nm FinFET | 14nm FinFET | 14nm FinFET |
| Compute Units | 40 | 36 | 20 | 16 |
| GCN Generation | GCN 4.0 | GCN 4.0 | GCN 4.0 | GCN 4.0 |
| Stream Processors | 2560 SPs | 2304 SPs | 1280 SPs | 1024 SPs |
| TMUs | TBA | TBA | TBA | TBA |
| ROPs | TBA | TBA | TBA | TBA |
| Memory Bus | 256-bit | 256-bit | 128-bit | 128-bit |
| Memory | 8GB GDDR5/X | 8GB GDDR5/X | 4GB GDDR5/X | 4GB GDDR5/X |
| TDP | ~110-150W | ~110-150W | <50W | <50W |
No comments:
Post a Comment