More detailed AMD GPU roadmap for 2016 to 2018 published

[Corky34]

Why? All of nVidia cards are also codenamed with lower series numbers being more powerful (GM100 > GM104 > GM106 etc.)

But AMD have clarified it's a chronological numbering, not performance based. They aren't user numbers either - it'll presumably be 480, 470 etc.

Yea their just as bad when it comes to my persnicketiness.
Both of them need to learn that bigger numbers are meant to mean more not less, at least they're just codenames though so i don't have to be subjected to that nails down a chalkboard feeling too often.

[kalniel]

Yea their just as bad when it comes to my persnicketiness.
Both of them need to learn that bigger numbers are meant to mean more not less, at least they're just codenames though so i don't have to be subjected to that nails down a chalkboard feeling too often.

Is it better to finish 1st in a race, or 10th?

[chj]

Stupid question but is it likely Polaris and Vegas will be using current pcie standard?

[Dayks]

Stupid question but is it likely Polaris and Vegas will be using current pcie standard?

I'm not an expert but I believe so, I had a look around for this myself and the general consensus was that pcie 3.0 has plenty of headroom left.

As far as I am aware it will be the sever versions that don't use pcie.

[Xlucine]

There'd be a lot more warning if PCIe motherboards were about to become obsolete. For one, we'd need MB's with the replacement to be available at the same time

Or perhaps they put two Polaris 10 dies on a carrier, using the same tech as used to mount HBM next to a Fiji die. If the yield drop off at high die size is bad enough, then that could pay for itself in increased yield across the wafer.

With the exponential drop in yields with die size I'd be very surprised if it didn't make sense financial sense - the question would be sorting the communication between the dies IMO

[lumireleon]

The current GDDR5 memory still an awesome performance headroom, adding HBM1 or HBM2 is not everything when it comes to card performance, the next gen core from AMD has to be a performer while minding its own energy consumption, I was disappointed with the Fury-X in many ways it has many core counts than any nvidia GPU out there but has not managed to beat a 980Ti & Titan X on many tests. The 4GB HBM1 performance was adequate & awesome as it was found to be equally comparable to the 980Ti with 6GB of GDDR5 which is more GBs for nothing in the field of gaming. What AMD needs to do is not to lie to us with 4096 cores (that maybe some lie idle) but to see their high end cards top on Withcer 3 @ 4K, 3DMark @ 4K, Total War Rome 3 @ 4k and other compute performance tests.

[kalniel]

With the exponential drop in yields with die size I'd be very surprised if it didn't make sense financial sense - the question would be sorting the communication between the dies IMO

PCIe Same as the X2 cards but on the interposer rather than the PCB. However I think there are other problems like power delivery and heat sink design that would make a two chip on interposer less viable than two chip on PCB.

[TooNice]

The current GDDR5 memory still an awesome performance headroom, adding HBM1 or HBM2 is not everything when it comes to card performance, [...]

This is something I have been wondering. I haven't kept up with tech news, but it seems like there isn't a single graphic card related thread where someone doesn't mention HBM2. Judging by the amount of time it is mentioned, I would expect it to be something that solve some major bottleneck in current graphic cards. Does it? Are we so memory bandwidth limited right now that we could expect a more significant boost from HBM2 than simply having a faster GPU?

[kalniel]

This is something I have been wondering. I haven't kept up with tech news, but it seems like there isn't a single graphic card related thread where someone doesn't mention HBM2. Judging by the amount of time it is mentioned, I would expect it to be something that solve some major bottleneck in current graphic cards. Does it? Are we so memory bandwidth limited right now that we could expect a more significant boost from HBM2 than simply having a faster GPU?

It depends on the task and inefficiencies of the underlying chip. AMD stood to gain more than nVidia anyway since their mem compression was a bit behind and they seem to be having to up the bandwidth on big parts to compete (see 290X etx. having 512bit interfaces vs the 256bit on the 980).

But the Fury range of chips which remove the bottleneck, aren't completely knock out of the park kind of increases, but it did seem to allow AMD to scale up the rest of the chip. So rather than being a boost by itself, HBM seems to allow for chip expansion without hitting a bottleneck.

There are other, mostly power, related gains for HBM too - which in turn helps you turn up the wick on the rest of the chip.

[DanceswithUnix]

PCIe Same as the X2 cards but on the interposer rather than the PCB. However I think there are other problems like power delivery and heat sink design that would make a two chip on interposer less viable than two chip on PCB.

But we were discussing smaller die, so overall thermal problem would be no worse than one large die and possible better as the heat is slightly spread about.

The GPU seems to funnel all memory traffic into the L2 and then back out into the memory controllers. There must be a big crossbar there to coordinate all that, if the crossbars on two chips could be tied together across an interposer then that could make them appear as one GPU enough to avoid having to do alternate frame rendering and see a shared memory pool.

But the Fury range of chips which remove the bottleneck, aren't completely knock out of the park kind of increases, but it did seem to allow AMD to scale up the rest of the chip. So rather than being a boost by itself, HBM seems to allow for chip expansion without hitting a bottleneck.

There are other, mostly power, related gains for HBM too - which in turn helps you turn up the wick on the rest of the chip.

The FuryX is basically a double sized R9 380X (Fury a double sized R9 380), and the 380 only has a 256 bit bus. So Fury could happily have worked on a 512 bit wide GDDR5 bus with no obvious bottlenecking on the memory. Forget the 290, that doesn't have the compression of GCN 1.2

[QuorTek]

I think that Radeon wins this round against Nvidia and probably a few years ahead.

Brave call! What's that based on?

For starters look at the price tags and that they already have experience with HBM... 1 year test run with the Fury X... yes it might not actually beat 980gtx/titan cards completely but it is up there and still performing very well and at a lower price and in general have a better hardware base compared to nvidia which is way more dependent on the software side... hey now I am not a fanboy if either GPU brand whenever it get to it... been running with Nvidia for a few years... before that it was Ati (now AMD) and before that it was 3DFX + Matrox... I think that AMD is the next one to rule the world of GFX is all, it look more promising... I am still running on a Nvidia 690 card myself and seen how the Nvidia cards evolve over the years... for starters I am not really very impressed compared to the ancient card I have so losing faith in Nvidia.

[kalniel]

For starters look at the price tags and that they already have experience with HBM... 1 year test run with the Fury X... yes it might not actually beat 980gtx/titan cards completely but it is up there and still performing very well and at a lower price

But this round isn't going to be HBM at all - that's the following round.

and in general have a better hardware base compared to nvidia which is way more dependent on the software side...

That's a very good point - DX12 may help a bit, but I don't see anything else that's really saying that nVidia won't be ahead on the software side again, especially for non-close to metal coded titles.

[DanceswithUnix]

That's a very good point - DX12 may help a bit, but I don't see anything else that's really saying that nVidia won't be ahead on the software side again, especially for non-close to metal coded titles.

My only thought there is that AMD seem to have had silicon in test for months now, and no sign of anything still from Nvidia. There is still time for Nvidia to get product into production for the summer, but how polished are the drivers going to be?

Could this be the first time for ages that an AMD launch doesn't seem rushed? I think they had a way to go to catch up on driver efficiency so that might still not be enough, but it could help.

[scaryjim]

... I think they had a way to go to catch up on driver efficiency so that might still not be enough, but it could help.

I think that depends on whether they're going to bother with DX11 driver optimisations, or if they're going all out for a better DX12 experience. AMD's DX12 drivers already seem to be very efficient, while their DX11 drivers seem to have a much higher CPU overhead than nvidia's, and if they don't bother doing anything about that they'll still look comparatively poor in most benchmark comparisons, because the number of pure DX12 games out there is so low...

[DanceswithUnix]

I think that depends on whether they're going to bother with DX11 driver optimisations, or if they're going all out for a better DX12 experience. AMD's DX12 drivers already seem to be very efficient, while their DX11 drivers seem to have a much higher CPU overhead than nvidia's, and if they don't bother doing anything about that they'll still look comparatively poor in most benchmark comparisons, because the number of pure DX12 games out there is so low...

I think that is what I meant, but better put

There have been a few articles like http://wccftech.com/amd-r9-fury-x-pe...atest-drivers/, but driver updates are too late as people seem to make their minds up at launch and when Fury launched it was most definitely the slower option. OTOH, you can't put launch off forever, at some point you need to ship stuff.

[DanceswithUnix]

Whilst browsing on there I see the latest rumour for Nvidia estimates as many transistors on a 1080 as you get in a Titan X but using the new 16nm process to halve the die size.

http://wccftech.com/nvidia-pascal-gp...ctured-leaked/

Now GPU performance largely comes down to clocking those transistors, so if the number of transistors is roughly the same then performance all comes down to how well that beastie clocks.