Even 8 FP64 units per SM would be overkill for a '4'-series chip. GM204 had a 1/32 ratio - that would equate to a mere 2 FP64 units per SM if the modules still have 64 standard CUDA cores as they do on GP100.
I dont think we really "know" that, not to mention the GP104 cluster may not be the exactly the same as GP100 cluster. I mean, it can have the same 64 FP32 cores, but not those additional 32 FP64 units (but say only 8 instead per cluster)...that should save loads of die space, right?
I personally need max capacity of VRAM for my GPGPU purposes. Its not only about bandwith or speed of the RAM for me, thus better memory management does not really help my case. Therefore say 12GB GDDR5 with 384bit bus is better for me than 8GB GDDR5X with 256 bit bus, even though their bandwith is pretty much the same?
Either way, with 8GB of GDDR5X, 1080 GP104 will already be a tangible improvement over 980Ti/Fury X for anyone who skipped those cards.
I've been reading both... that it will match those and that it will exceed them. I can't really picture anything but the latter, because then NV would be holding back like crazy given the process advantage. But like many others I have the worry in the back of my mind that even with competition, performance this time will be held back to the annual average increase on 28nm despite a huge advantage with the 16/14nm jump.
Will they? They will have to deliver steady sales numbers to their stockholders on 14/16nm for 4+ years probably. They have plenty of motivation to hold back and not make it more difficult to themselves to come up with products in 2017/18/19/20.
Will they? They will have to deliver steady sales numbers to their stockholders on 14/16nm for 4+ years probably. They have plenty of motivation to hold back and not make it more difficult to themselves to come up with products in 2017/18/19/20.
AMD won't hold back with Vega, it'll be the last big GPU they make. Whether that's 450mm2 or 550mm2 is another question though.
If they didn't hold back it would be 600mm2. And Polaris certainly wouldn't be the size it is.
Currently, all NVIDIA AIB partners of the GeForce GTX 970/980 / 980Ti series production has been discontinued, no longer supply the chipset from NVIDIA.
...The initial product launch are all the reference product and the mainstream (GTX 1070) Class AIB partners of the cooler only designs products that are released about two weeks late.
Nvidia ceases deliveries of Geforce GTX 970, GTX 980 and GTX 980 Ti VGAs
www.hwbattle.com/bbs/board.php?bo_table=news&wr_id=18732
Will they? They will have to deliver steady sales numbers to their stockholders on 14/16nm for 4+ years probably. They have plenty of motivation to hold back and not make it more difficult to themselves to come up with products in 2017/18/19/20.
Nvidia ceases deliveries of Geforce GTX 970, GTX 980 and GTX 980 Ti VGAs
www.hwbattle.com/bbs/board.php?bo_table=news&wr_id=18732
AMD won't hold back with Vega, it'll be the last big GPU they make. Whether that's 450mm2 or 550mm2 is another question though.
Why?
- SLI/CF support requires even more developer work and driver work from AMD/NV. Seems in the last 6 months a lot more AAA games have poor support
- NV isn't going to stop making large die flagships
- Mid-range die strategy failed with HD4000-6000 series
- Since 14nm node is here to stay for a long time, we should expect die sizes to grow as the node matures (7970->290X -> Fury X)
- As the node matures, it will become more affordable to produce 550-600mm2 die
All of these reasons suggest AMD should go through a similar transition of increasing die sizes (as outlined above from Tahiti to Hawaii to Fiji). Then with 7/10nm node, it starts all over again.
Even if Vega is only a 450-480mm2 die, it doesn't mean it will be AMD's last large die flagship, unless you have inside sources?
Unless you believe AMD's new strategy is to put a 450-500mm2 die against NV's 300mm2 GP104, how do you expect them to match or beat a 600+mm2 Big Pascal? Don't tell me dual-chip flagship Vega card priced at $699? Radeon Pro Duo is $1499, R9 295X2 before it was $1499. So this strategy will not work.
----
970/980/980Ti going EOL was reported a long time ago. Not sure why this is suddenly news again. Also, cards launching for June is a consistent theme. Again, not new information.
Even if NV low-balls early GP104 with a crappy 40% increase in perf/watt, 1080 will already beat 980Ti. I think after-market 1080 GDDR5X will deliver 50-80% performance increase over 980 at 1440/4K. The time is running out to sell those 980Ti cards for close to market prices....
Why?
They'll go multi die acting as one larger gpu. Not impossible if designed properly from the getgo.
It seems you missed out on all the talk of multi-dies on an interposer communicating as one-large die. To software/drivers, it will be indistinguishable because the inter-chip communication will appear the same as intra-chip communications from various blocks, CU groups and such..
Remember that GPUs are already quite modular, there's no reason they cannot exists apart but linked via an interposer.
Raja stated this twice in public, one in PCPER's interview with him, he talks about the future where node yields will be much lower than today and one way to solve that is exactly with multi-smaller dies on the same chip. The second time, some tech journalist asked him on twitter whether interposers can be used to connect multiple GPUs together into one and his response was a solid "Yes".
This will be the norm at 10 and 7nm.
Btw, the usual folks who are anti-AMD having been very dismissive of this approach as usual, saying it can't be done etc... so it will be true.
It seems you missed out on all the talk of multi-dies on an interposer communicating as one-large die. To software/drivers, it will be indistinguishable because the inter-chip communication will appear the same as intra-chip communications from various blocks, CU groups and such..
Remember that GPUs are already quite modular, there's no reason they cannot exists apart but linked via an interposer.
Raja stated this twice in public, one in PCPER's interview with him, he talks about the future where node yields will be much lower than today and one way to solve that is exactly with multi-smaller dies on the same chip. The second time, some tech journalist asked him on twitter whether interposers can be used to connect multiple GPUs together into one and his response was a solid "Yes".
This will be the norm at 10 and 7nm.
Btw, the usual folks who are anti-AMD having been very dismissive of this approach as usual, saying it can't be done etc... so it will be true.
I'm quite skeptical about the feasibility of this myself. Someday? Sure. Foreseeable future? I don't think so.
Why? You can get 4096bit HBM2 to communicate with a GPU.
All that's important are lanes of communications.
When you look at GPU uarchs, they have blocks communicating with cache & each other. Within each block, there's TMUs, ROPS, and cores. Thus they often cut a few blocks out to make a lower end SKU.
A multi-chip approach wouldn't work without the interposer giving it the lanes of communications that it needs.
We'll see it with Navi, late 2018 or whenever 10/7nm comes around early on.