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

there's no connection between one is HBM and one is not?

 

I think I'm gonna steal that for my sig

 

The question was why AMD has the tech in a high-end GPU.
One of the big reasons, I think, is that having a big impact on VRAM-limited games on cheap discrete GPUs is comparatively minor because cheap discrete GPUs are a minor consideration and wouldn't justify the expense in developing it. I'm not sure where games in general rank in importance, given the set of features Vega's memory subsystem caters to that have limited importance for anything consumer-facing.

 

I meant one company stating that you don't need more RAM just a smarter use of what you have and another putting as many modules as they can fit in the product. This is a gaming GPU so I have no idea in what scenario you will remotely need 11(!!) GB of ram...

 

oh, I guess this is certainly a way to look at it. Pros and cons to everything

 

Yes but at the same time this is a performance advantage over Nvidia who would need 2x the RAM to deliver same experience. I think it will be important for an APU also, having lets say 1GB of HBM and use system memory for the rest and still delivering a smooth experience in 1080p would be a killer for low-end VGAs.

 

It's likely more practical for their APU designs and they had the feature so ran with it on discrete parts. Memory management would get fun when rendering and a host program tried to share that pool. Gaming aside, using HBM to accelerate CPU performance on a low to mid range system could be interesting. We still haven't seen any real details on that, but it stands to reason there is a significant value add from the combination.

 

2x chips for the purposes of capacity, or bandwidth?
HBM is very expensive for its capacity, so it's not challenging to beat it on those grounds.
For bandwidth, the question is how much would be designed into a cheap discrete.

Just by being an APU, a significant chunk of the penalty "fixed" by the HBCC demo doesn't exist (limit system RAM capacity and it's unplayable). There are other implementations and software methods that could get quite a bit of the gain without a hardware controller. Because it appears to be overkill for this purpose and isn't in products that have need of this purpose, it may mean it mostly wasn't meant for this.

 

Simply take a look at the thread about 4gb VRAM not being enough and some of the usage of modern games at 1080p or 1440p, let alone 4k resolution. 11gb doesn't seem that far fetched for a 4k focused card and looking ahead.

 

Most games doesn't really use much more than 4GB, 8 is more than enough so 11 is just crazy...Unless! HDR games will use much more RAM(I dont know) then 11GB is logical.

 

Many modern games allocate more than 4GB of VRAM, but that doesn't mean the game accesses it all every frame (it can't, there wouldn't be enough bandwidth). Less than 50% of the loaded data is accessed during a peak single frame (at 64 KB page granularity, at cache line granularity even less). Vega is supposed to have hardware paging solution that uploads pages on demand from system memory to GPU memory. We don't yet know how well it works. It is entirely possible that a 8GB Vega is comparable to a traditional 16GB+ card.

Nvidia P100 has similar feature for CUDA compute use and it seems very promising (esp with NVLINK). Game data sets change much less per frame compared to long running professional compute workloads, meaning that similar technology should work even better for games.

 

Vega isnt fast enough, 3 cards from Nvidia in less than a year faster than anything AMD has and AMD Raja haven't been able to beat them yet.
They blame drivers isnt ready soon, then some failure to allocate dies as the factories are filled.
No, VR is going to save Rajas job.
no cigar and dancing and alcohol wont help when the hardware isnt fast enough.
are Raja sandbagging or are we in for a rough awakening?

 

In many ways performance at greater than 4 GB VRAM actual (not reservation) usage is less prevalent than gamers with 4K monitors. Hence, it's relevance is small, but important.

Regards,
SB

 

Always nice to declare the sky is falling before a product is even out.

The same was said before R300 launched, Rv770 launched, and Polaris launched. Will Vega be more like the former or more like the latter? We'll only know when the product is closer to launching.

Regards,
SB

 

Can't blame this on Raja, Vega's design was done well before Raja even came back, he is just making sure it gets out the door, or trying to. And if AMD ends up blaming him for it, they would be fools.

 

I'm not sure. Koduri was hired back as VP of Visual Computing in 2013, and Vega isn't breaking reaching market until 2017. New CPUs take roughly 5 years to get to market, and the most complex GPUs have taken somewhat less but are getting there.

I think 2013 is enough for him to take some ownership of it. I was cautiously optimistic about his words upon the creation of the RTG last year, but I don't think the results have yet earned the swagger, cigar, and swaying on the dance floor from the last time.

 

Actually Vega is rajas first design, it's been 5 years since he joined the company so he started Vega development.

Enviado desde mi HTC One mediante Tapatalk

 

has it been that long, seems like it was much shorter than that lol

 

I still do not see anything able to survive clock normalization.