AMD Polaris Rumors and Discussion

Aren't these new cards for neural networks exclusively?
If so, is there any advantage in talking about DP?
Some of the very large science labs will want to do both training and FP64 modelling.
DGX-1 in general allows smaller labs to really broaden what they can do with a node in terms of modelling/training/inferencing with premium performance, which is ironic when one considers Nvidia is also pushing the Int8 accelerated based funtions for inferencing as the way forwards and requires GP102 instead of the P100.
I think the P100 is more of a forerunner though for the evolution jump they intend with Volta and especially Tesla market (1st massive scale HPC contracts are for mid-late 2017), instead of one big tech jump they broke it into milestones.

Cheers
 
AMD didn't talk about Bistrol Ridge's GPU either, and then it comes with half rate DP. Hmm.
Does seem rather odd to add it in and then do away with it on your next major architecture launch.

A native FP64 high performance scalar to absorb the special functions and MADD, MUL, DP in the SIMD I'd think could get away with it without significantly increasing logic. Similar concept may work for 4xINT8. Disable the SIMD and use all the decoding logic to feed it. Not quite the same as 4xINT8 across the SIMD, but it probably wouldn't bottleneck as easily. Might be possible to consider INT4 at that point.
 
Since this is brought up, the 2016 MBP thingy is not really about the support of LPDDR4, but the lack of 64Gbit x32 LPDDR3 (and apparently Apple isn't interested to call for a part that would have a super low economies of scale).
Gotcha, thanks for the correction. Are 64 Gb x32 LPDDR4 chips expected in the future?

At least when I last checked the public product catalogues, you can't even achieve 16GB in 128-bit DQ with what Micron and Samsung are offering in LPDDR4 (except SK Hynix, perhaps).
I'm not familiar with the x[number] nomenclature and related properties, and you mention Hynix in your post, but is 8 Hynix 32 Gb x16 LPDDR4 chips a valid possibility for 32 GB? (Assuming 128-bit bus)
 
Nvidia is looking to still sell the K40 and K80 for some time, so AMD can still keep the Hawii DP products, and worth remembering the last Hawaii DP FirePro was upgraded to 32GB in April 2016, albeit the Workstation model.
Yeah none of those models are ideal relative to modern efficiencies in latest GPUs but are cheap comparatively to latest gen HPC-Pro.
Cheers
 
Last edited:
Gotcha, thanks for the correction. Are 64 Gb x32 LPDDR4 chips expected in the future?

I'm not familiar with the x[number] nomenclature and related properties, and you mention Hynix in your post, but is 8 Hynix 32 Gb x16 LPDDR4 chips a valid possibility for 32 GB? (Assuming 128-bit bus)
Yeah. x16 usually means the package has 16-bit data bus. So eight of them would get you 32GB. But I am not sure if LPDDR4 is really available in x16. AFAIK the spec is two indepedent 16-bit channels per die though. Moreover, the ballout count seems to suggest otherwise. It looks more like 24Gb and 32Gb are available only in x32/x64 packages.
 

And early benchmarks are pointing to lower-than-P11 performance:

http://videocardz.com/64787/amd-polaris-12-spotted-in-linux-patches

Looks like I'm going to eat my virtual hat on this one.


OEM demands are a strange thing. I guess the best part of staying on top of the food chain like apple or Microsoft is that you don't really have to deal with this bullshit.

Maybe they're going for a Shield TV 2 competitor with P12.
 
And early benchmarks are pointing to lower-than-P11 performance:

http://videocardz.com/64787/amd-polaris-12-spotted-in-linux-patches
I think it was quite expectable. AMD already confirmed it in the Polaris architecture whitepaper (27th July 2016), here is direct quote:
Polaris-based GPUs have 1-4 geometry engines, depending on overall performance targets (e.g. the Radeon™ RX 460 GPU has two, while the Radeon™ RX 480 GPU has four).
;-)
 
okay thats just weird, its a desktop cpu and the gpu is performing worse than anything in a very long time. is it possible that this is some kind of prototype cayman card, that some guy got a hold of and posted online hd6980 maybe lol.
 
And early benchmarks are pointing to lower-than-P11 performance:

http://videocardz.com/64787/amd-polaris-12-spotted-in-linux-patches

Looks like I'm going to eat my virtual hat on this one.


OEM demands are a strange thing. I guess the best part of staying on top of the food chain like apple or Microsoft is that you don't really have to deal with this bullshit.

Maybe they're going for a Shield TV 2 competitor with P12.

This may be crazy but my future father in law runs his own business and we set him up with 3 monitors per pc. We are using really old radeon 5x00 series cards that are no longer supported. IF Polaris 12 sips power and runs 3 monitors esp new 4k ones we will get them for those pcs and ditch the old ones. Not everyone needs gaming power
 
This may be crazy but my future father in law runs his own business and we set him up with 3 monitors per pc. We are using really old radeon 5x00 series cards that are no longer supported. IF Polaris 12 sips power and runs 3 monitors esp new 4k ones we will get them for those pcs and ditch the old ones. Not everyone needs gaming power

R7 240 with an Oland chip?
$10 per card after rebate.
 
non-gaming HTPC's with HDMI 2.0, 4k HEVC decide, 2020 support etc is another perfect use for those cards.
 
that's not bad at all either. Might grab those for now but they are surely getting close to end of life
It's GCN 1.0. It's no spring chicken (can you believe 7970 came out 5 years ago this week?!), but AMD won't be dropping 1.0 support for a while yet. I reckon it gets a few more years before it gets shuffled off to legacy, in part because AMD sold Pitcairn for so long.
 
It's GCN 1.0. It's no spring chicken (can you believe 7970 came out 5 years ago this week?!), but AMD won't be dropping 1.0 support for a while yet. I reckon it gets a few more years before it gets shuffled off to legacy, in part because AMD sold Pitcairn for so long.

And it's friggin' $10.
No-brainer IMO.
 
Do you think a custom version of raven ridge is a realistic possibility for project scorpio?
(8 logical ryzen cores / 6 TF vega)
 
Last edited:
Do you think a custom version of raven ridge is a realistic possibility for project scorpio?
(8 logical ryzen cores / 6 TF vega)
Hard to tell. It could also be an Arm core going by the recent MS moves of getting Windows on Arm. Will be interesting to see what if any features get implemented on Scorpio that miss the Vega window.
 
Back
Top