AMD was kind enough to send a document on the technical advances of Tonga/R9 285 our way Sunday night in the middle of Labor Day weekend here in the US. We hope you enjoy a list of bullet points that may or may not be worth a damn, given how badly this was messaged, it is the best we can do.
Yes a list of bullet points has replaced an actual technical brief, it seems that AMD doesn’t actually want anyone asking questions about Tonga. Normally this happens when the product is essentially a BS update with nothing real to offer the customer. This time it may or may not be, and the ‘advances’ in question may or may not be real, we can’t actually tell you for sure. With that in mind, take this with a big grain of salt, it could very well be complete fluff.
Now that Tonga/R9 285 is here, it brings a new updated version of GCN with it. You might recall GCN numbering went 1, 1.1, now this, it might be 1.2, it might be 3 if you count 1.1 as 2, or something like that. It is as clear as mud and since there was no opportunity to ask a question, who the hell knows? Not us but that is OK because we really don’t care at this point, whee!
Tonga brings new ‘technical’ five bullet points, improved tessellation performance, lossless delta color compression, updated “ISA instruction set” (Author’s note: Yes I know… it is that way in the slides), high quality scaler, and a new multimedia engine. How do these work, and what technically do they bring to the table? No clue really, not even sure if they are actually real but there is a good chance that there may be something under the heatspreader.
For tessellation AMD claims 4 primitives per clock and a 2-4x throughput improvement. They have a graph that shows that too so it has to be legit, right? That is all we have on the topic, a potential technical advance that, well, they didn’t think was worth talking about on a technical level.
Same with the updated ISA sans echo because it isn’t really even a good joke in the first place. It is said to have 16-bit FP and Int instructions that AMD claims helps low power GPU compute and media processing. We are going to call complete BS on this, it might have been believable if they hadn’t tried to spin it as ‘low power’ but that makes us think there is nothing real behind the curtain. Same with instructions that allow “Sharing between SIMD lanes”, whatever that means. Haven’t got a clue about what this actually does much less why we should care. Please note this may be a technical advance but we can’t give you the first clue about what it actually does.
This is all wrapped up with “Improved compute task scheduling” which once again may actually have some merit or may be about as realistic as the inevitable chromed bikini for the woman with the large sword on the box. Who knows? Not us. AMD didn’t actually list the instructions much less go into any detail about them at all. Call us a tad skeptical here, then again how can we argue with a nebulous ‘better’ claim?
We now come to frame buffer compression, complete with a graph showing 40% better memory performance on 3DMark Firestrike. Frame buffer compression is something as old as the hills and we are actually pretty stunned that it wasn’t done up until this point, we thought it was. Welcome to the 1990’s AMD, assuming that this too is real. And works on anything but Firestrike. Even if it does it may not actually be worth a damn because of overhead, power consumption, or latency induced by the compression, something an actual brief would have answered.
Then AMD claims upgraded display controllers, specifically the upscaler. We are going to be flat out skeptical here because if your don’t use Windows, AMD’s GPUs kinda blow because of horrific driver quality. Video related functions are especially worthless in such situations so the new “Polyphase filter with up to 10 horizontal taps and six vertical taps, low pass kernel” is worth its weight in electrons used to describe it. Same with the new pre-scaler and “hardware block implementation within the display pipe”. What do they do? Beats me once again.
AMD takes a slide to describe the new multimedia engine that they call a powerhouse but only compare it to the GTX760 in later slides. We won’t speculate as to why they picked this particular one, maybe because it is the one thing that they can actually beat. Now. Maybe, still not sure if they can actually do it though.
This new update possibly brings MJPEG to the list of codecs as well as H.264 High Profile Level 5.2 support. They claim best in class H.264 encode performance but in our tests, without Windows mind you, it doesn’t actually work on previous models so we won’t hold our breath here. AMD claims some best in class H.264 features which is awesome. Unfortunately the entire world has moved on to H.265 because no one is dumb enough to use H.264 for anything but trivial 4K work. A golf clap at best here, again only if real.
Then there is a list of best in class HD playback and transcoding features as long as H.265 isn’t on your list of needed items. Once again if these ‘features’ ever come to non-Windows OSes we may actually care. There may be some real tech under the hood, but if the last 4-5 generation are any guide, don’t count on it.
Powertune is the next bit up, something SemiAccurate covered in exhausting detail here. We have absolutely no clue if Tonga improves on the last version of the tech, or even if it is a new iteration thereof. Nothing was said in the slides worth mentioning other than a dumbed down version of the above article. Complete with graphs. Simple graphs. Shoot me now please.
So now Tonga is out and it is the first PI/Pirate Islands card. This new line brings advances so great that AMD doesn’t actually want to tell you about them. The tech under the hood, if it is actually real or new, may actually be worth a damn but don’t take our word for it because we have no way to tell. Sadly that seems to be the way semiconductor companies want to talk about things now, the tech isn’t worth briefing press on anymore.S|A
Latest posts by Charlie Demerjian (see all)
- Intel shows off 10nm 112Gbps SerDes - Mar 12, 2019
- Intel releases Compute Express Link spec - Mar 11, 2019
- Qualcomm rolls out a second gen 5G modem called X55 - Feb 19, 2019
- What is Intel’s Foveros tech and what isn’t it? - Feb 11, 2019
- Why SemiAccurate called 10nm wrong - Jan 25, 2019