Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

AMD Confirms RX 480 At $199 USD, Other APU & Polaris Announcements

IMO it will look like this:
DX300 - 2048 GCN cores, 4 GB of GDDR5
DX500 - 2304 GCN cores, 8 GB of GDDR5
DX700 - more than 2304 GCN cores, less than 4096, 8 GB of HBM2

So far Zauba shipping manifests indicate that first GPU with C99XXX moniker will have 3 times higher cost than RX 480, and that the second GPU with C99XXX moniker will have 4 times higher price than Polaris 10.

Note that there is no difference in cost of GPUs using the same die, because you pay for die, regardless if it is fully functioning or not, at least thats what we can draw from looking at shipping manifests on Zauba.

So there are 4 GPU dies coming from AMD currently made on 14/16 nm process: Polaris 10, Polaris 11, Vega 10, Vega 11.

And every GPU can be custom made, semi-custom design, or fully functional normal GPU in upcoming Mac Pro. It is up to Apple what they will decide.

Interesting thing in this context however is the fact that none of the RX 480 is working fully under macOS Sierra so far, but RX 470 - is.

So either Apple will go for RX 470 as base model, and higher tiers are based on Vega, or we will have huge surprises on release.

P.I. I genuinely thought that difference between RX 480 and GTX 1080 will be much bigger in compute benchmarks. It is not bad, at all.
 
koyoot said:
IMO it will look like this:
DX300 - 2048 GCN cores, 4 GB of GDDR5
DX500 - 2304 GCN cores, 8 GB of GDDR5
DX700 - more than 2304 GCN cores, less than 4096, 8 GB of HBM2

So far Zauba shipping manifests indicate that first GPU with C99XXX moniker will have 3 times higher cost than RX 480, and that the second GPU with C99XXX moniker will have 4 times higher price than Polaris 10.

Note that there is no difference in cost of GPUs using the same die, because you pay for die, regardless if it is fully functioning or not, at least thats what we can draw from looking at shipping manifests on Zauba.

So there are 4 GPU dies coming from AMD currently made on 14/16 nm process: Polaris 10, Polaris 11, Vega 10, Vega 11.

And every GPU can be custom made, semi-custom design, or fully functional normal GPU in upcoming Mac Pro. It is up to Apple what they will decide.

Interesting thing in this context however is the fact that none of the RX 480 is working fully under macOS Sierra so far, but RX 470 - is.

So either Apple will go for RX 470 as base model, and higher tiers are based on Vega, or we will have huge surprises on release.

P.I. I genuinely thought that difference between RX 480 and GTX 1080 will be much bigger in compute benchmarks. It is not bad, at all.
Click to expand...

IMHO, RX 470 will go to iMac (Pro). And what comes to Mac Pro, I am expecting the "huge surprise".
 
In the GM build of macOS 10.12 Sierra, AMD9500Controller.kext contains these PCI IDs: 0x67E01002 0x67EF1002 0x67FF1002 0x67C01002 0x67DF1002.

67E0, 67EF, and 67FF are Baffin Polaris 11.

67C0 and 67DF are Ellesmere Polaris 10.

67DF is definitely RX 470 but I'm not sure for which GPU 67C0 is. I have not found any reference as to what this 67C0 might be but it's definitely a higher end card than RX 480(67DF:C7) and RX 470(67DF:CF). This could very well be the one going into the new Mac Pro.

Polaris 67DF:C7 is the full chip so this mysterious card could be a dual GPU setup.

Source: http://pciids.sourceforge.net/v2.2/pci.ids
 
I probably missed the message 50 pages ago stating this, but do the RX 4x0 series run at 2.5 or 5 GT/sec on cMPs?
 
http://radeonprorender.com

AMD does pretty good job lately at forging the OpenSource software initiatives.

theitsage said:
67DF is definitely RX 470 but I'm not sure for which GPU 67C0 is. I have not found any reference as to what this 67C0 might be but it's definitely a higher end card than RX 480(67DF:C7) and RX 470(67DF:CF). This could very well be the one going into the new Mac Pro.

Polaris 67DF:C7 is the full chip so this mysterious card could be a dual GPU setup.

Source: http://pciids.sourceforge.net/v2.2/pci.ids
Click to expand...
What do you base this? 67C0 is just different name for Ellesmere die. There is only one die, and one configuration for standard Ellesmere die: 36 CU's from which you can cut down other versions of it.

67C0 can be however custom made chip. Will it be? It is highly unlikely because custom design for each chip is around 90 mln USD. It can be cut down version of the initial Polaris 10(Ellesmere) design.
 
koyoot said:
It is highly unlikely because custom design for each chip is around 90 mln USD.
Click to expand...

I'm not current on wafer cost, but years ago some friends developed a bitcoin miner asic ant each prototype on 28nm wafer was on about 10K$ and they need to wait in row about 90 days (due other labs demand), they got about 200 chip on these waffers less than 3% defective, at end they spend about 500K on theirs first production batch (or at leas that was what they told).

I dint think so, this maybe the cost to develop a an all new chip from 0, but a variation on a current chip cost its much more less expensive, it depends the kind of extra R&D work required, the minimal foundries production run cost, once you have a GCN processing core defined(this is the really expensive r&d), build a custom chip on it its somewhat like cut'n paste, as many other logic components can be recycled on a custom chip r&d work its much more less.

Another question is the production setup, once your design is stable to get economical mas production you need a minimal setup at the foundries, a minimal # of wafers you can order on a production setup.

http://www.sigenics.com/custom-asic-cost-calculator/

of course NRE cost vary way long not the same as in the US, Germany or Israel as on Korea, India or China.
 
DeviceIDs don't tell much. Much more you can predict from frambebuffer personalities in appropriate kexts. Regarding mythical nnMP of course :D
 
Mago, He bought Silicon wafer. Not designed the chip from the ground up ;).

Buying silicon wafers are part of silicon design, and the costs of silicon wafers are part of silicon design costs ;). 28 nm silicon design cost was around 30 mln USD, 14 nm is 90 mln, 7 nm will be 270 mln. All of this is because of increased prices of silicon wafers. And the ball is rolling more and more...
quote:
At 5nm, it will cost $500 million or more to design a “reasonably complex SoC,” Johnson said. In comparison, it will cost $271 million to design a 7nm SoC, which is about 9 times the cost for a 28nm planar device, according to Gartner.
Click to expand...
http://semiengineering.com/going-to-gate-all-around-fets/
666sheep said:
DeviceIDs don't tell much. Much more you can predict from frambebuffer personalities in appropriate kexts. Regarding mythical nnMP of course :D
Click to expand...
Have you spotted anything new, Black? ;)
 
koyoot said:
design costs ;). 28 nm silicon design cost was around 30 mln USD, 14 nm is 90 mln, 7 nm will be 270 mln.
Click to expand...
these are empiric numbers, not the same design an all new chip, than a variation on an existent chip (on those most of the issues you alreade solved it when you developed the original), while those numbers seem related to USA/GERMANY/ISRAEL RET Korea/China has RET much lowers .
 
koyoot said:
Have you spotted anything new, Black? ;)
Click to expand...

I didn't even update to Sierra GM. Lack of time currently.
But if one will find nMP specific personalities in new AMD Controller kexts, it would indicate that Apple is working on something. Otherwise I would be skeptical.
 
666sheep said:
if one will find nMP specific personalities in new AMD Controller kexts, it would indicate that Apple is working on something
Click to expand...
This only indicates a product is ready to be launched, Apple could keep Kext specific for new products under developments and just after announcement integrate them into macOS release then upload the "update".

But after seeing an iPhone 7 plus "just competitive" against a Mid Range Android, i don't have hopes on Apple ot be hurry to update nothing, maybe they wont release this year the nnMP and waits for next year (we are safe to see updated rMBP, iMacs, MBA? and Retina Thunderbolt Display), new Mac Pro and mini to wait until Q2/17.
 
Mago: but you aren't sure, that's why you said "could keep" instead of "is keeping". That is what this (and few other) threads is all about. Illusions, dreams and wishes instead of proofs.
 
theitsage said:
RX 470(67DF:CF). This could very well be the one going into the new Mac Pro.

Polaris 67DF:C7 is the full chip so this mysterious card could be a dual GPU setup.
Click to expand...

I tried to decode this info, then I guess how it could fit int the nnMP:

The tcMP (nnMP not Tim Cook Mega P.), is constrained about PCIe lines to drive a lot of stuff requiring PCIe3 lines each Alpine ridge controller rq 4, a typical NVMe rq 4 too, on the other side current nMP keeps its 2nd GPU dormant most of the itme bein useless due macOS intrinsic lack of suupor for dual GPU setup, whit this in mind I relized the less traumatic solution is to include a powerfull dual GPU card (240-300W TDP) into a single slot and use the available PCIe lines (20) to drive 3x AplineRidge for 6 Thunderbolt 3 ports (12 lines total), 2x NVMe PCIe X4 SSD (8 lines total) = 36 PCIe that's left 4 pcie3 lines + 8 pcie2 available for other peripherals as dual 10G Lan, WiFi, USB 3.1

666sheep said:
That is what this (and few other) threads is all about. Illusions, dreams and wishes instead of proofs.
Click to expand...

I want to believe, things exist before we have proofs, proof are for UN-bleievers with no faith.
 
Mago said:
I tried to decode this info, then I guess how it could fit int the nnMP:

The tcMP (nnMP not Tim Cook Mega P.), is constrained about PCIe lines to drive a lot of stuff requiring PCIe3 lines each Alpine ridge controller rq 4, a typical NVMe rq 4 too, on the other side current nMP keeps its 2nd GPU dormant most of the itme bein useless due macOS intrinsic lack of suupor for dual GPU setup, whit this in mind I relized the less traumatic solution is to include a powerfull dual GPU card (240-300W TDP) into a single slot and use the available PCIe lines (20) to drive 3x AplineRidge for 6 Thunderbolt 3 ports (12 lines total), 2x NVMe PCIe X4 SSD (8 lines total) = 36 PCIe that's left 4 pcie3 lines + 8 pcie2 available for other peripherals as dual 10G Lan, WiFi, USB 3.1



I want to believe, things exist before we have proofs, proof are for UN-bleievers with no faith.
Click to expand...
You can achieve this by using coherent fabric on GPUs and running them on dual x8 PCIe bus.
 
macOS Metal has improved its multiple GPU support. Not very difficult anymore really. Hopefully app developers will use it more.

"Some macOS devices feature multiple GPUs. If you need to work with multiple GPUs, call the MTLCopyAllDevices function to obtain an array of available devices. Create and retain at least one command queue for each GPU you use." - Metal Best Practices Guide
 
Zarniwoop said:
macOS Metal has improved its multiple GPU support. Not very difficult anymore really. Hopefully app developers will use it more.

"Some macOS devices feature multiple GPUs. If you need to work with multiple GPUs, call the MTLCopyAllDevices function to obtain an array of available devices. Create and retain at least one command queue for each GPU you use." - Metal Best Practices Guide
Click to expand...
Erm, isn't this similar to Microsoft Explicit MultiAdapter in its idea?

Now things are getting more interesting, for my inner 10 year old :D
 
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back