AMD: Speculation, Rumors, and Discussion (Archive)
- Thread starter iMacmatician
- Start date
- Status
I don't see anything logical in that....
It doesn't matter which dies you're comparing: they're still pieces of silicon that use different processes and using different fabs. GF may started with their 14nm implementation 2 quarters before (or after) TSMC started with 16nm. And who knows which company is better in execution at implementing a new process from start to the same level of maturity.
Without that knowledge, it's pointless to claim that Vega will have a yield advantage just because it'll be a couple of quarters later than some chip from Nvidia.
I don't expect anything at all. All I know us that it took long time to get Polaris from early silicon to having a product on the market. (If we have to go by AMD's own claims of being ahead, they squandered their lead.)
With Polaris 10 having a clock speed that's in line with expectations, it doesn't look like AMD dramatically rearchitected GCN to allow higher clock speeds, so it would not be an outrageous bet to claim that process played a role in this, but that's about it.
well https://twitter.com/AMD/status/742455340839493632 confirms it
but I don't know how amd could possibly make a soc so rectangular
Because it was it is compatible with Xbox1 games so unless Nvidia discover a super emulation technology I dont see it possible unless you are using another iteration of GCN, so Polaris or Vega.
Never mind then.
CPU+GPU?
D
Deleted member 13524
Guest
Same site has the 1080 for over $900, so...
https://www.pcworldbusiness.co.uk/search?keywords=gtx+1080
Looks like the site is just price gouging rx 480's for now.
Looks like the site is just price gouging rx 480's for now.
Well that is a head scratcher because the live presentation showed 2x480 just beating a 1080FE, in a game ideal for AMD with AotS.
2x390x would be around 23-25% faster than a 1080FE using the internal benchmark IF using the 1.83 multiplier AMD inferred - this is not necessarily a real performance for the 390x but based upon the number provided by AMD for the dual 480.
So in that game a 390x is "faster" than 480 by a noteable amount when taking this back to single GPU, sure the 480 may perform better relatively in other games but so do Nvidia cards as AotS is one of their weaker ones.
Makes me think it is a speculation leak *shrug*.
Cheers
Last edited:
itsmydamnation
Veteran
Peak vs average, if we consider ATOS peak for AMD because its compute heavy has async and doesn't hurt AMD with tessaltion and thus can get high occupation and utilization on GCN 1.X. In that case the raw throughput rate of hawaii looks really good. But we know AMD have lots of changes coming to frontend and shaders. We also have a few snipits like this so if Polaris average and peak is much closer then GCN1.X then both situations can be true. Then there is the overclock possibilities if you dont care about TDP.
Multi GPU probably scales worse on newer cards before drivers are fully there.
firstminion
Newcomer
No
I believe I was misinterpreted, I meant that by focusing on Polaris and launching Vega later AMD have the marginal benefit of reducing Vega's risk (by expectation of increased yields).
If the rx 480 can hit 980ti stock performance things get interesting. If the rx 480 can over clock and be faster than over clocked 980tis then you have a huge market open up for amd.
$250 for an 8 gig card with 980ti performance for vr is a huge deal. You'd just greatly drop the entry price for vr
hoom
Veteran
Yeah I was like
But I guess most esports are older/less 3D intensive + tend to be played on min-quality for max responsiveness.
Yeah I was like
But I guess most esports are older/less 3D intensive + tend to be played on min-quality for max responsiveness.
The 460 could be close to a 280/290 in performance. That would be pretty good for most of the esport games
I didn't study the render but the first thing that popped into my head was GDDR5X.
256bit x 10ghz = 320GB/s
too many sixes with a 384bit bus
384bit x 6666mhz = 320GB/s
Every die I have ever seen was rectangular. Or are you talking about the above average aspect ratio? But I still don't see any connection between the shape of the die and the architecture of the chip. It's basically unrelated and simply the result of an optimization procedure for the layout of the chip.
I don't see how it would be especially difficult to fab chips with shapes of elongated rectangles. Sure, you loose slightly more area cutting them out of a wafer, but this is clearly not the dominating factor when talking about 300+mm² dies. So of course AMD, nV or intel can possibly design (and have done so!) dies with such proportions.
As mentioned, this is simply the result of optimizing the layout and even the manufacturability of the die. Traditionally, one wants to put all external interfaces on the outer perimeter. And if possible, one wants to avoid driving some signals over the whole die all the time, so certain components should be better placed in proximity to each other. And with more and more lithography steps necessary during production, it may be also advantageous, to shape the die so more copies fit to a single mask within the reticle limit to reduce the number of individual exposures for a wafer. A square die of let's say 375mm² would be pretty bad for that (only a single copy fits to the photomask). But make it 15x25mm² and all of the sudden two copies fit to a single mask halving the number of individual exposures a stepper unit has to perform for a wafer. And this increases the throughput and reduces the cost (there is a reason why ASML specifies the throughput of their machines as 275 wafers per hour with 96 shots for instance, it gets less with more shots per wafer
).- Status
