Nvidia Ampere Discussion [2020-05-14]

Pascal's launch wasnt any better. It took months for most AIBs to be available and that prices would be under 700€ for the cheapest cards. And four years ago nVidia sold half dGPUs.

 

People sitting at homes due to a small issue with the virus are hoarding these cards, and all other home electronics too for that matter.

 

Well checking some threads from Overclockers.uk during the pascal launch.

https://www.overclockers.co.uk/foru...ders-edition-wanna-pre-order.18731877/page-13

some quotes from that thread from Gibbo:

Day before launch.

Evening of the launch day

Not sure if they shipped all of those at that point or if some were back ordered, but I think they might have actually had that amount of cards. or cleared that queue quickly, because about 3 weeks from the launch he posted:

also this is a factory model.

Everyone who wanted a card in their forum had a good chance to get one and they had sizable new shipments within few weeks of the launch. They were happily posting the pyramid pictures, this time they weren't allowed to do so.

The data from their forum members put the total shipment figures of Ampere cards into few hundreds as of today and lot of people didn't get a card. It's just one shop, but I think it tells a story.

Now I understand that the demand is higher now, but the volume of cards at least over here is suspiciously tiny.

 

Did anyone log the instruction mix of folding@home workloads yet? That should be a nice compute load for GA10x (x!=0) …
I currently see only minor improvements on a 3080 vs. a 2080 Ti, in the range of 10-20%, which makes me wonder whether or not F@H is predominantly special functions, which Nvidia did not scale up apart from higher #SMs and clocks. Power draw OTOH seems to indicate a very loaded GPU.

But maybe current Work Units don't scale up very good and FP32 monsters must wait for WUs with a higher base rating.

 

Not many reviews included F&H results, though did notice this a few weeks after 3080 reviews.
https://www.legitreviews.com/geforc...foldinghome-cuda-enabled-gpu-workloads_222389

Edit: VideocardZ does have results from different projects.
https://videocardz.com/press-release/foldinghome-gets-15-30-performance-boost-with-nvidia-cuda

 

Saw the legit site as well, without OC, I'm currently at around 5,7M ppd, but probably need to wait more in order to get a fitting WU. A not exactly overclocked 2080 Ti yielded up to 4,8-4,9M ppd.

The Nvidia-provided results from videocardz seem to make sense for small projects with 3080 not being significantly faster than 2080 Ti. Only on larger projects, they show an increase for very wide GPUs like A100, but did not include RTX 30 for unknown reasons.

 

Did I miss the anandtech review ?

 

Saw a twitter comment that said it was still being worked on, though think he might wait to include RTX 3070 results.

 

2070 super's seem to be running out of stock. At least newegg barely has any models under 600$. Same thing happened earlier with 2080 super's, the prices rocketed before 3080 launch as the stock started to be gone.

https://www.newegg.com/p/pl?d=2070+super

 

Looks like at the moment you can't buy any nvidia GPU's, except the 2060 and 2070 over here, and even the 2070 is scarce. Quite peculiar.

 

3070 was supposed to come out this week and replace 2070 super. 3070 got delayed 2 weeks. If 3070 supply is good that would be perfect situation for nvidia as no need to sell old product with discount price. On the other hand if 3090/3080/3070 supply is gimped that should show up in nvidia's q4 results in a really bad way.

 

Update from Proshop values:
https://www.computerbase.de/2020-10/nvidia-geforce-rtx-3000-serie-liefersituation/
2020 is really over

 

There was a rumor a while back that production was stopped on TU104 based cards. Of course that would be silly if you don't have anything new to sell.

 

Germany seems still full of RTX 2000 cards:
https://geizhals.de/?cat=gra16_512&...=de&hloc=pl&plz=&dist=&mail=&sort=r&bl1_id=30
(this is preselected for in-stock-cards)

 

https://hexus.net/qaeox2

There we go, somewhat backed / insider reports about RTX 30xx lineup being ported over to TMSC at some point in 2021.

 

Its unlikely that Nvidia can squeeze more performance out of Samsungs foundry so it was inevitable they moved back to TSMC for 30 Super series. But unless AMD shakes them up badly this quarter, i think those gpus are a year away

 

It's not too much of a stretch to assume Nvidia already has some 7nm RTX samples since they already use TSMC's 7nm FinFET process for Ampere A100.

 

What happened to SS8 is much cheaper and has 'infinite' capacity argument?

 

It will be interesting to see which architecture offers the best performance/efficiency once on the same node.

 

Im not sure i subscribe to the ”yields are bad!” theory. If i can speculate i think Nvidia wanted to announce first to steal AMDs thunder but they were nowhere near mass production in reality. Thats why some AIBs had to change MLCC setup at the last minute. And the current rumor is that production is ramping up big time now