Looks like AMDs upcoming 14nm GPU is a low end like previously speculated.
Only 232mm2.
Quote:
Project F is yet another codename for the chip designed at AMD HQ. What we've learned from previous LinkedIn leaks is that just because something is designed and reaches engineering state, it does not mean it will end up as a product for end-users. Project F could be one of such designs.
According to the profile of former AMD employee Project F is a discrete GPU designed in 14nm Low-Power Plus manufacturing process. Such chip would be manufactured by either Samsung or Global Foundaries. 14LPP is a next iteration of 14nm process bringing improvements over 14LPE (Lower-Power Early) process in performance and power efficiency roughly by 15%.
I need to make a chart of die size vs transistors vs node over time. But I don't think that's unreasonable. Pitcairn and Cayman both had similar transistor counts, at 2.8 and 2.64 billion respectively, and (very) roughly the same performance. Cayman was 389mm2 and Pitcairn was 212mm2. That was from just a single node shrink (and to be fair a very large architectural shift). Here, we're going to see two since 20nm was skipped. Assuming the changes to GCN are big enough, this might end up being around 290X/980/780Ti tier while drawing much less power on a much smaller and hopefully cheaper die. Die size raises costs exponentially, or so I've heard, so smaller and denser is better, especially early on.
I need to make a chart of die size vs transistors vs node over time. But I don't think that's unreasonable. Pitcairn and Cayman both had similar transistor counts, at 2.8 and 2.64 billion respectively, and (very) roughly the same performance. Cayman was 389mm2 and Pitcairn was 212mm2. That was from just a single node shrink (and to be fair a very large architectural shift). Here, we're going to see two since 20nm was skipped. Assuming the changes to GCN are big enough, this might end up being around 290X/980/780Ti tier while drawing much less power on a much smaller and hopefully cheaper die. Die size raises costs exponentially, or so I've heard, so smaller and denser is better, especially early on.
It is probably the GPU that AMD was talking about "bringing VR required GPU performance for lower price".
That card is probably like 150-199$ and should be around 290X +-5% levels of performance and consume like 100W~. I have nothing official, but I feel like I am not to far off with my estimations.
The die size has pretty much no indication on the performance range of a certain GPU coming out. This is merely proving that AMD is using 14nm for the next major GPU line-up.
The 380 is a rather low performaning card (relatively seen as we have 2016 and what gains we have seen the last couple of years).
The R9 380 is basically just a slightly stronger HD 7950. A card that was release early 2012. 4 damn years ago.
Why do people have so low expectations? For crying out loud how could a 232mm² Polaris 14nm FF card be as bad as a 366 mm² 28nm card that is cut down [28 compute units (1792 stream processors)] ?
The 380 is a rather low performaning card (relatively seen as we have 2016 and what gains we have seen the last couple of years).
The R9 380 is basically just a slightly stronger HD 7950. A card that was release early 2012. 4 damn years ago.
Why do people have so low expectations? For crying out loud how could a 232mm² Polaris 14nm FF card be as bad as a 366 mm² 28nm card that is cut down [28 compute units (1792 stream processors)] ?
199-230, like the 380-380X, as we can asume that we will have two SKUs out of the chip.
The thing is, if the CES chip is the little one (Polaris 10), this has to be the "big" one in the naming scheme, the Polaris 11. I will be very disappointed if AMD doesn't release a bigger one to replace the 390/390X.
199-230, like the 380-380X, as we can asume that we will have two SKUs out of the chip.
The thing is, if the CES chip is the little one (Polaris 10), this has to be the "big" one in the naming scheme, the Polaris 11. I will be very disappointed if AMD doesn't release a bigger one to replace the 390/390X.
It's a possibility that this thing will end up faster than the 390 in some usage cases, so it might replace it pretty seamlessly.
edit: If you've seen the material AMD has been releasing around polaris, it's pretty clear they want to paint a picture of the architecture being better, not necessarily much faster per dollar, on release.
199-230, like the 380-380X, as we can asume that we will have two SKUs out of the chip.
The thing is, if the CES chip is the little one (Polaris 10), this has to be the "big" one in the naming scheme, the Polaris 11. I will be very disappointed if AMD doesn't release a bigger one to replace the 390/390X.
Thing is, if the small chip already performs at least as well as the 960 (which is roughly ~10-15% slower than a 380) it makes absolutely zero sense for the bigger chip to be right on top of that. The smallest difference should be in the 25% range for the cut down version which would put this in line with the 390.
Quote:
Originally Posted by Tivan
It's a possibility that this thing will end up faster than the 390 in some usage cases, so it might replace it pretty seamlessly.
edit: If you've seen the material AMD has been releasing around polaris, it's pretty clear they want to paint a picture of the architecture being better, not necessarily much faster per dollar, on release.
The 380 is a rather low performaning card (relatively seen as we have 2016 and what gains we have seen the last couple of years).
The R9 380 is basically just a slightly stronger HD 7950. A card that was release early 2012. 4 damn years ago.
Why do people have so low expectations? For crying out loud how could a 232mm² Polaris 14nm FF card be as bad as a 366 mm² 28nm card that is cut down [28 compute units (1792 stream processors)] ?
We have nothing really to compare but die size. A "14nm" 232mm² die should pack roughly the same transistors as a 28nm 375mm² die would. You would think a 14nm should be exactly half the size of a 28nm, but that isnt really true because much of what is called 14nm is actually closer to the 20nm sizes on transistors.
Anyway, just because of die size estimations people are thinking of R9 380 performance level because the sizes would be similar. Realistically, we should be a bit higher since we are also going to see some good architecture improvements this time around as well.
Basic maths would make you think that a chip ported to 14nm, should be exactly 1/4th the size of a 28nm chip, because surface area is a quadratic measure.
Thing is, if the small chip already performs at least as well as the 960 (which is roughly ~10-15% slower than a 380) it makes absolutely zero sense for the bigger chip to be right on top of that. The smallest difference should be in the 25% range for the cut down version which would put this in line with the 390.
Perf/W doubled from 6970 to 7870. This should be an even bigger difference, since the difference in nodes is bigger.
To be fair, the 6000 series was also very cheap, and AMD was in a good enough position with their move to 28nm that they could price the way they did. Lack of a performance increase is expected. Usually, last generation's flagship or an equivalent is the not-quite-highest-end GPU the next one. So the 7800 chips replaced the 6900 chips, and were replaced by the 270 chips.
Thanks to AMD's small die strategy, that is correct. Cheaper and easier to produce than Nvidia's monolithic dies. But it's clear that that strategy has gone out the window, ever since Hawaii confirmed it. The Fiji die found in the Furies and the Nano is a massive 596mm2, one of the biggest GPUs ever made and pushing TSMC's fabs to the limit. That's the biggest reason neither AMD no Nvidia included much in the way of FP64 computing in Fiji or GM200; it would take up too much precious die space.
Come on guys.. this thing is going to blow the 380 out of the water. Even if Polaris doesn't include a per core performance boost, you can cram a lot more SPs on 14nm than on 28nm.
This is probably going to have around 290x level of performance and it ought to sell for around $250 tops.
I am not all that confident in this project F being an actual production chip. The source doesnt give much context but as of right now, projects C, D, and E dont appear to be existing 28nm products. I suppose one or two could be eventual R7 equivalent replacements.
Quote:
Originally Posted by EniGma1987
We have nothing really to compare but die size. A "14nm" 232mm² die should pack roughly the same transistors as a 28nm 375mm² die would. You would think a 14nm should be exactly half the size of a 28nm, but that isnt really true because much of what is called 14nm is actually closer to the 20nm sizes on transistors.
Anyway, just because of die size estimations people are thinking of R9 380 performance level because the sizes would be similar. Realistically, we should be a bit higher since we are also going to see some good architecture improvements this time around as well.
According to TSMC, their 16ff+ has 2 times the density of their 28hpm.
This is an older thread, you may not receive a response, and could be reviving an old thread. Please consider creating a new thread.
Related Threads
?
?
?
?
?
Ask a question
Ask a question
Overclock.net
27.8M posts
541.2K members
Since 2004
A forum community dedicated to overclocking enthusiasts and testing the limits of computing. Come join the discussion about computing, builds, collections, displays, models, styles, scales, specifications, reviews, accessories, classifieds, and more!