Seems entirely pointless to bring Fiji to 14nm if Vega exists. The architecture is already made, the R&D has been spent. It wouldn't be any cheaper to bring Fiji to 14nm than to create a smaller Vega.Seeing all these low clock Vega 687F:C3/687F:C1 benchmarks with varying performance brackets, I got the sneaking suspicion there's a 14nm Fiji dieshrink in the lower end of the Vega stack.
I'm probably completely wrong but if Pascal could do so well being mostly a Maxwell dieshrink at high clocks, Fiji at high clocks could do pretty well too, it'll probably alleviate some of the frontend bottleneck running at those clocks as well.
Unless the frontend runs its own clock independent of the core clock.
http://cdn.wccftech.com/wp-content/...R9-Fury_HBM_OC_1000-MHz_Unlocked-1140x883.png
Here's a non-X Fury on steroids, wonder how much the 1TB/s bandwidth OC helps the graphics score.
Obviously for something close to reticle limit, they don't have a choice and have to use the best performing foundry but (at least for the GPU portion of the console chips) there is no reason why if you have a process which is, say, 20% worse (in terms of max clock or power) that for a GPU you couldn't just design it to be wider but slower. For something as parallel as a GPU wide and slow should work quite well at the cost of die area but it allows the design to stick the process' sweet spot. Of course, this is something you have to decide at right at the beginning of your design. And from The Stilt's clock/watt charts for Ryzen, there can be more than one such spot.GF's wafers are by default cheaper for AMD at least due to WSA. More wafers bought means lower penalty. That's why I find it weird that console chips are made at TSMC...
Seems entirely pointless to bring Fiji to 14nm if Vega exists. The architecture is already made, the R&D has been spent. It wouldn't be any cheaper to bring Fiji to 14nm than to create a smaller Vega.
As for Pascal, it didn't automatically benefit from the new process in terms of clockspeeds. NVIDIA tweaked it so it can do it. If you read the Polaris whitepaper, RX 480's clockspeed bump was mostly from some clever circuit design rather than the process.
It has the exact same speficiations of the Nano, so it's probably just 28nm Fiji.I forget the chart, but I recall that the lower end MI pro cards are Fiji-based (maybe the lowest tier). Was that on 14nm or no? If AMD is still making new Fiji cards for any purpose, why wouldn't they bring them to 14nm? I'm assuming there are massive costs here, and lack of demand/need to justify those costs?
Where are you getting these figures? Are you sure there isn't a decimal point missing? 7.0-8.0 mil?It has the exact same speficiations of the Nano, so it's probably just 28nm Fiji.
There are indeed fairly large costs that need to be amortized. Mask costs range from 70-80mil on 14nm, and setting up a new production chain, tooling, and the actual design costs (because you can't just send a 28nm chip and copy paste it into 14nm) are costs as well.
It has the exact same speficiations of the Nano, so it's probably just 28nm Fiji.
There are indeed fairly large costs that need to be amortized. Mask costs range from 70-80mil on 14nm, and setting up a new production chain, tooling, and the actual design costs (because you can't just send a 28nm chip and copy paste it into 14nm) are costs as well.
If you're doing all of this, you might as well make a chip from your latest architecture. Costs shouldn't be any higher.
SteamIs that what it officially looks like in Steam? Or is that a shop?
I'm fairly certain I heard Lisa Su mention it in the investor call. I may be wrong.Where are you getting these figures? Are you sure there isn't a decimal point missing? 7.0-8.0 mil?
I got this data for 20nm. The enormous jump in costs seems unlikely.
193i multi-patterning, a direct consequence of 193i resolution limit, is a major driver of escalating cost. At 20nm node, for example, a set of masks costs $5 million, and the cost continues to mount unabated. Here are the detrimental effects of multi-patterning:
At 14nm node, there are at least 14 design layers for which single patterning would not work, thus requiring 27 critical mask layers. At 10nm node, the number of critical mask layers jumps to 42 or more. As critical layers soar, so do masks per set. While 44 masks are needed to process 45nm node ICs, 68 masks are required at 14/16nm. While patterning steps (i.e., mask steps) grow proportionately, the number and cost of 193i tools rise as well.
http://www.multibeamcorp.com/GSAIoTSecurity.pdf
Thanks. Picked up both now. Enjoying playing Vega on my new prey machine....Ladies and gentlemen, get out your Radeon RX VEGA to play Prey!
Steam
Thanks. Picked up both now. Enjoying playing Vega on my new prey machine....
/s
This looks like a single game taken from the downloads menu and then put over that ad. I have never ever seen an ad displayed in the downloads menu, and I doubt AMD has the clout to make itself the first.Steam
Are you sure? It looks like a photoshop with the vega logo on the left and the steam page on the right
I wonder if the density (transistors/area) has something to do with that? That is, does a design able to hit higher speeds have to have at least some parts designed to be less dense so that a design able to clock faster would have to 'waste' some die space to reach those clocks.
And might power be equally affected by such decisions.
Designing a less dense design might cost a bit more in terms of dies per wafer, but if it means better clocks or perf/watts and hence is able to command a higher price it would make sense, especially if GF's wafers are cheaper than TSMC (which they should be if their process is worse).
Half the time my card is getting 80% of the gtx 1070 fps. Other half it's actually outperforming the 1080ti by 30%.How many FPS are you getting?
He's just posting it out of sarcasm. The fact that one part of that duo isn't available
I deliberately asked if it were Steam or a photoshop though. The outright lie is unappreciated, if it is indeed a fake. This would not be the first time guesses were posted as facts...
He took the photoshopped image from reddit. Look at the URL:
https://i.redd.it/v7hywugg7ivy.png
He never said that is what he sees.Sure did. User in Australia said that that is what he sees.