AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

I asked if they could run a benchmark that isolates geometry performance, and to try that one program that can show whether a gpu has some kind of tiled rasterization, but they didn't even notice me in the chat.

 

Try mentioning their names (Ryan, PCPer, etc) into your question/request. I asked about Quadro scores ("Do PCPer have quadro scores?") and they answered immediately. Or maybe I was lucky.

 

They are talking about tile based rendering now

 

Yeah, I missed it but found this pic.

 

So both last nights youtube guy and PCPer confirmed it's terrible at ethereum mining, about 30mh/sec. Good for consumers wanting to buy RX Vega at launch I guess

 

Since it seems to be tile based perhaps we will see large gains from drivers.

 

I watched the live stream. Just tried trianglebin on my 290 (GCN2) and it's completely different to what I saw Vega do. The whole upper triangle flashes weirdly all at once on the 290 whereas Vega flashes and changes in big tiles, then smaller ones as seen in the screenshot taken in the middle of the transition.

Vega is definitely doing things different than good old GCN.

 

For thermal issues, it shouldn't do much. Below 85C, there shouldn't be an appreciable error rate or the DRAM is faulty. The error rate ECC is meant to avoid would be spectacularly unlikely in the short demo period even once.
Above 85C and up to when HBM2's thermal trip sensor kicks in, the DRAM's refresh rate will increase. If that starts failing in that range, the DRAM is broken or it's melting down.

Where was that indicated, and what would that have to do with thermal issues?

Basic parity checks would not have the information necessary to fix anything.

Is there a specific citation of 4x256 dwords with regards to VGPRs?

 

GCN HD6670
Tiled Maxwell
Tiled Pascal


Links to the Kanter video on tiled raster. Thumbnail should be showing Maxwell. Looks like Vega has ~64 huge tiles. Different, but probably the equivalent of larger ROP caches from using L2. Tiled raster should have overlapping colors. May be differences in tools used so who knows.

 

But i watch this picture carefully，it‘s no different with my R9 Nano.

 

Oh,I see the different.Thank you.

 

Looks like it is rendering the triangles sequentially to me....

Although I don't expect much (performance gain) from drivers, I do think being a prosumer card the power and thermal management for the FE are probably tuned for quiet, reliable operation. I think Ryan was only averaging about 1450 on the clocks without touching the fan control. It is reasonable to assume that the RX version will be more aggressively profiled, and I can easily see something in the range of +10-15% performance. Of course it will be hotter and more power hungry, but as long as it isn't too loud most people probably won't care. I do think some people will probably be overly harsh in their judgement of the timing of the card, but entering a new market or re-entering an old one is always tough because you are always going to be viewed as late vs established competition. And I also suspect some did not really appreciate just how damn good Pascal was/is with regard to energy efficiency. Maxwell was already quite good and Pascal actually extended Nvidia's lead despite AMD making some decent gains of their own. Obviously, that is one area AMD will need to continue to improve, but it is not the end of the world. While I don't do any GPU dev work, I would probably be more interested in probing the capabilities of Vega than the rather stale (though amazingly efficient) Pascal. Even just for gaming purposes, I suspect Vega will probably be the more interesting card to own.... as long as the heat doesn't result in too much fan noise, I have grown to find it intolerable in my old age.

Even if Vega doesn't turn out to be the greatest thing since sliced bread (performance wise), I'm just happy to see AMD back competing in the high end. I do wish they would start executing the launches a little better. They have done reasonably well on the CPU side lately, and they've done well on the GPU side in the past. But this and the Polaris launch have both been rather, puzzling, I would say is the right word. Iron out a system that works and stick with that going forward.... It will be interesting to see how AMD prices the RX version, but I haven't seen any reason it can't be a successful (if not groundbreaking) re-entry to the high end.

 

1450 core clock would still put it right around 12 tflops, which is close to 1080 Ti levels of theoretical performance.

 

Let's be fair: unless it's based on some feedback from AMD, his comment about that doesn't have more weight than that of any other internet poster.

If current results are close to Vega RX, there must be some critical issue that could not easily be found with simulations. Or was only found very late in the process. But that doesn't make sense because the bad performance is pretty much across the board.

Were they counting on some software wizardry that didn't work out in the end? Or that are still work in progress?

 

Probably the shader use will make a huge difference.

 

Ryan stated that AMD told him there is no difference between Pro and Game mode, and that RX drivers will be newer, that's it. Also that Game mode has game optimizations and is working properly, his testing indicates that as well, the driver has Tiled Rendering, HBCC and the whole gang.

He also said he went over the phone with AMD BEFORE showing that on the stream, and they were OK with these results.

One thing is clear, AMD seems reluctant to announce any sort of major difference between the current driver and the upcoming RX driver, which leads many to believe there isn't in fact any major difference, otherwise AMD would have been quick to point out, especially in light of the current fan uproar.

In fact that idea in itself is ridiculous, considering you can play extremely well even on a Quadro which doesn't have a game mode or game drivers.




https://hothardware.com/reviews/nvidia-quadro-p6000-and-p5000-workstation-gpu-reviews?page=6

 

Has anyone yet been able to come up with a logical explanation as to why AMD chose not to provide cards to the press for one of their most important releases of last couple of years, outside of the shamefull PCWorld hands-off preview? I mean, the optics of this are just God-awful, akin to a studio refusing to screen a major movie to the critics which is never a good sign.

 

They want to delay the bad feedback ? Not overshadowing the good ryzen / threadreaper / Epyc momentum ?

 

Because they had to ship something to hit the deadline promised to their investors. There will be a proper launch for the gaming products and only a couple hardware geeks will care. I don't believe in magic patches though, so I reckon it's as good(bad) as it'll get.