1.) RX470 doesn't fit the power connector-less niche. This means it can't be used in New value desktops like
this one or one of the many used Core i5 Pre-builts out there.
Sure. Still a niche market, which they already have covered by the low-cost 460. But there are some potential sales there, if the cost isn't too high.
2.) As I mentioned earlier RX470 is already down to $159.99 without rebate. Taking the same Polaris die and making a 75W card out of it should cost less than making a 120W RX470 out of it (re: less VRMs and simpler cooling for any given card size)
Cheaper per card? Probably. But that's without counting R&D cost to actually develop this board, which would be significant. The 470 was essentially free, as it uses the same board as the 480. Not to mention the cost of fusing off parts of functional GPUs (to avoid people modding them to act as a 470/480, further killing AMDs margins) and the added cost of the larger memory bus compared to the 460. Or are you suggesting they disable some of the memory controller?
3.) 1792sp is 22% reduction from the full polaris die of 2304sp (used in the RX480). This along with around a 21% reduction in clocks (1000 Mhz vs. 1266 Mhz on the parent RX480) is not a huge chop in stream processors and frequency compared to past AMD releases:
It's still a larger chop than they were willing to make to get the RX 470. Considering how close that is to the 480, don't you think there might be a reason why they went that way?
Cape Verde GPU:
HD 7770 = 640sp, 1000 Mhz
HD7750 = 512sp, 800 Mhz (20% reduction in Sp and 20% reduction in clocks compared to HD7770)
Pitcairn GPU:
HD7870 = 1280sp, 1000 Mhz
HD7850= 1024sp, 850 Mhz (20% reduction in Sp and 15% reduction in clocks compared to HD7870)
(Though I do admit the TDP reduction 75W vs. 150W (RX480) is greater than the examples above which were 55W vs. 80W and 150W vs. 190W. But then again if you look at the chart in the opening post the voltage vs. frequency curve was much different on 28nm)
^^^^ Notice how 28nm voltage vs. frequency ramps up so much sooner on Bonaire (which is actually newer than the GCN 1.0 examples I used) compared to even this early FinFET (Ellsmere = Polaris 10).
Sure, P10 has a better voltage curve. But what else does that curve tell us? That there are no power savings to be had below ~1GHz. And again, you're only talking about the GPU, not the rest of the card. VRAM is a significant power sink on a GPU - between 10 and 20W on most GPUs. In other words, they won't reach 75W without a combination of reducing bus width, downclocking the RAM significantly, and/or reducing the amount significantly. And then we come into power delivery - sure, fewer VRMs bring down costs - but for them to operate efficiently (to keep within the ambitious power envelope), they'd need to be expensive, good VRMs. Same goes for anything else related to power delivery.
Let's do some math on actual Polaris cards:
The RX 480 is a 2304SP (144TU/32ROP), 1120-1266MHz card with 4-8GB of 8GHz RAM on a 256-bit bus, with a TBP of 150W.
The RX 470 is a 2048SP (128TU/32ROP), 926-1206MHz card with 4-8GB of 6.6GHz RAM on a 256-bit bus, with a TBP of 120W.
So to get a 20% reduction in power, you need a 11,1% reduction in SPs and TUs, 17,5% reduction in RAM clocks, and 17,4-5,0% reduction in clocks (base/boost). All on the same board. Thus, you can't simply extrapolate the core reduction and say "11% less SPs on the 470 gives a 20% power reduction, so a 22,2% reduction (1792 SPs) should give a 40% reduction." You'd have to lower everything else alongside this - given that power gains lower down in core clocks, RAM clocks and the like are linear (which your handy graph shows that they aren't). Your theoretical 1792SP card with 40% lower power than an RX 480 (i.e. 90W, not 75) would also require a base clock of ~730MHz (boost is harder to calculate, but would probably need a larger drop than the 470 relative to the 480), and RAM clocks of 5,2GHz. That's starting to look pretty unappealing already. And again, I'm ignoring the fact that your graph says there is
no power savings to be had below ~1GHz.
Disabling a memory controller (for a 192-bit bus) would probably be more efficient, but then you'd end up with some weird choices - either go with 3GB of RAM (cheap, but less RAM than the RX 460) or 6GB (adds cost, but better). Also, if 8GB of 8GHz RAM on a 256-bit bus consumes (say) 20W, 6GB of 8GHz RAM on a 192-bit bus would probably consume around 15. 5W less, sure, but we're still a ways away from 75W.
Getting Polaris 10 to 75W without binning would be difficult, expensive (in terms of R&D cost, which is significant), and would make for a pretty unattractive card. My guess is that this is why AMD isn't doing it.
Facebook
Twitter