Graphics Pascal and Polaris Discussion Thread

looking forward to seeing the Pascal on April 4 - 7th at GTC.
Really hope they give us a date. normally it only takes a month after unveiling for it to hit stores with nvidia, sometimes sooner.

 

hmmm I thought that 8GB double stacked HBM1 was possible (4-Hi or even 8-Hi) , at least Hynix said something along these lines last year - so we really could see 8GB polaris using HBM1 this year....

 

Acording to there own website on a 1024bit interface 8gb HBM1 is doable. Sounds expensive though.

 

Bindis post is the image that AMD has given to its investors as its official road map. If indeed as most news sites have reported that its 4gb limited again they could be at a significant disadvantage unless they hope dx12 and CFX will allow them to merge 2 together to give access to more memory and better 4K fps.

At 4K which is the point of new gpus 4gb is already touch and go.

 

that's 8GB per stack - fury uses 4 stacks of 1GB - so @8GB that's a 32GB card!

as for that image - that's why I asked about Hynix and 2GB stacks - they would likely be at that now....

 

If the leaks about 23xx shader counts are confirmed true, it looks like Polaris 10 and 11 are mainstream parts only, where 4GB could be acceptable, but it probably won't be HBM-based due to cost-sensitivity. Unless AMD found a way to put two Polaris on a single interposer and work as 1, both with 4GB. It's possible from a HW and SW standpoint, but technically a true breakthrough.

We should see something from Nvidia next week. Almost 100% sure they'll be using GDDR-based, single chip like usual.

 

SK Hynix said they were working on 2gb stacks of HBM1 last year - so why would they need to do what you suggest to go 8GB

 

1) Haven't yet seen example of this versus effort to get HBM2 done.
2) HBM cost is higher than equivalent GDDR5 so in anything but premium segment that's still a no-no. GDDR5 allows AIB partners to negotiate their own sourcing and price, which for the biggest players is an economic advantage.

 

it must be down to cost as :

https://www.skhynix.com/eng/product/dramHBM.jsp



shows 8GB 4-Hi of HBM1!

 

Shows 8Gb, not 8GB.

 

OK, it looks like Nvidia have gone completely bug**** insane and led their Pascal lineup with a balls-to-the-wall extreme compute card: the Tesla P100
- Make the jump to 16nm? OK, We'll fab an EVEN LARGER die then ever before at over 600mm^2, more than 4 times the size of any other 16nm die anywhere. So what the hell are we going to stuff this monster with?
- New architecture! Throw even more SMXs in, and add a whole bunch of FP64 back in too
- HBM2, 16gb of it!
- Power? 300W.
- Core frequency for a professional card (where Nvidia usually downclock for stability)? 1.4 freaking GHz!
- NVLink or PCIe? How about 4x NVLink interconnects AND 16x PCIe!


Well at least Nvidia is confident TSMC can get good yields from 16nm already. Then again, P100 has 4 SMXs disabled (56 of 60), so there's even room for a This Isn't Even My Final Form card later on.

 

Q117 - same as Vega. All because of HBM2.

The GP100 die is totally compute-centric, with a lot of die space dedicated to 1:2 FP64 performance and there's about ~48MB of caches + the extra IO from NVLink.
Anandtech's EiC tweeted that Maxwell arch on 16nm FF+, in 610mm^2 would be 8000+ shaders! So it could be that we'll see GP104 with 4096(64SM) AND it's still smaller because they drop all the compute related stuff not required for graphics. The core frequency is surprising for that size die, although at 300W it could be best case scenario depending on the quality of silicon. It also bodes well for the smaller cores having 1.5G+

GK104 = 294 mm²
GM204 = 398 mm²
~GP104 = ~300-400mm²

I'd strongly expect it to be close to the 300 mark, due to additional cost + size advantage of 16FF+, yield etcetc and of course the fact Nvidia will want to squeeze every $ profit it can.

OR (unlikely)

We could see tiny dies strapped together using an interposer (CoSoW) with NV link as the coherent link between them to make one pseudo-large die on single card. Then NV will just adjust the number of small cores per card.

 

I doubt we'll see anything about a consumer card until E3 now.

 

Certainly June and I agree with E3 rather than Computex actually (I didn't realize E3 was only two weeks after):

Nvidia will always do its own thing with a focus its own brand + US market first, which is E3 over Computex.
AMD will almost certainly show its hand at Computex because it loves using the AIB card partners to do their marketing push.
However,
AIBs will sooner cuddle up to Nvidia rather than AMD due to marketshare/sales potential. If Nvidia isn't an option that opens the door for AMD to be first and foremost.

Even if Nvidia says no to Computex we'll see some leaks of stock cards or official names/specs from a partner, we always do, because that info can't be traced back. Read: I will go find some graphics PMs at a few partners.

 

I only mentioned E3 as (unlike Computex), it's on Nvidia's calendar!

 

On one hand I'm not sure because it lists nothing in APAC, which is strange. The TW site is not maintained.

However on the other hand that says everything you need to know about Nvidia: If it's not in the US it doesn't happen. Most of its AIBs are China/Taiwan-based - it's two biggest are in Taipei (MSI + ASUS) but it's brand-centric id + almost monopoly marketshare means it doesn't need their support. It's bad to launch in this timezone. AIBs will bend the knee whatever.

 

Not sure about getting good yields, lets face it, so long as they earn enough to make money the yields will be 'good enough'. They're selling it in the DGX1, which consists of 8 P100s (plus other bits see link) for $129,000. Anyone know the rough costs of the other bits?

 

done a lot of reading into this and TBH I'm not that impressed at all - its very FP16 and FP64 centric , and 10.6 tflops? furyx is 8.6 already - single precision ; the extra transistors have gone into double precision - but FP32 (which is important to gamers ) takes a knock. Its reminds me of the AMD cpu`s with its dual FP16 into FP32.

this isn't a card for gamers. Deep learning for sure , but GP1xx for pc will be a different beast.


also - the 5 miracle breakthroughs - we already have 16nm products on the market and interposer-connected memory? Fury from last year would like to ask you about that!

no wonder he looked so nervous with those blatant lies!

 

A card that is not for gamers does not impress a gamer what's new lol.

Fury X will never see a server and the card is better compared to AMDs existing server gpus.600m2 board is bigger than most expected double my expectations.

Most 16nm products are for smartphones which is a entirely different performance envelope. Intel has its own products but they are not comparable to what he meant.

The biggest issue i have with Nvidias call was no consumer gpus mensioned, and trusting gloflo as AMD is doing is surely a receipt for disaster if history is anything to go by.

Could it really be August or September before we see a next gen card if so that's pretty disapointing from my point of view.

 

The titan X is the retail version of the Tesla M40 - they changed nothing for it when they sold it for `gamers`