A1xLLcqAgt0qc2RyMz0y
Veteran
Revenue and margins are what counts for corporations as they drive profits.
Nvidia Post-Volta (Ampere?) Rumor and Speculation Thread
- Status
Revenue and margins are what counts for corporations as they drive profits.
But didn't NVIDIAs margins crash pretty hard after Turing launch?
Well of course, but that graph is a marketing piece, not internal or investor. It gives the impression that they sold 45% more Turing than Pascal and there are already articles with that type of wording because there are websites and people that merely spew whatever marketing provides.
The power of graphs.
DavidGraham
Veteran
FWIW, it was given to investors during an investor event after GTC.
A1xLLcqAgt0qc2RyMz0y
Veteran
Topic is about Turing only and no Turing margins did not crash.
What crashed was sales for mid-to-low end Pascal cards as they were awash in excess inventory at AIB's because said AIB's missed the memo that Crypto was crashing and ordered too many the previous quarters. Because of that they then didn't order any new cards the last quarter.
Last edited:
A1xLLcqAgt0qc2RyMz0y
Veteran
It is not a marketing piece as you state. It is directly from a NVIDIA investors day held for analysts.
So, let me get this straight: AIBs are to blame for the post-turing crash? Because they had flooded the channel with lower performing Pascal cards and they did not buy higher cost Turing card, which they allegedly would be able just to move through inventory? I fail to see any reason here.
A1xLLcqAgt0qc2RyMz0y
Veteran
That is not what I said at all.
Turing card revenues are higher than Pascal card revenues for the first 8 weeks of sales as shown in the chart above shown at the Investors Day recently.
Turing cards are selling at a lower volume than Pascal did because of the higher prices even with lower sales volume the Price times sales quantity is higher for Turing vs Pascal.
and yes AIBs are to blame for the revenue shortfall for Nvidia last quarter because they ordered way too many lower performing Pascal cards in the previous quarter.
How hard is it to understand too many widgets in inventory results in no more widget orders until widget inventory gets reduced.
And fits in nicely with next speed bump for HBM2.
They used to pride themselves on being able to tapeout a design with only a couple of metal spins, though it obviously failed a fair few times. Still, a year seems a long timeframe.
I'd just take their word for it and assume they are really avoiding 7nm for now, I fully expect it to still be a clusterfuck a year from now.
1 year is the normal timeframe for a big chip in a new process from tapeout to product. GP100 was also 1 year, vega was nearly 1 year on a known process. Polaris was faster, i think just 9 months, as was Turing. But Polaris was a simpler chip and Nvidia had very good process experience for Turing, which probably saved them some metal spins.
I'm assuming this is for pro market and no plans of consumer variants if at all?
Rootax
Veteran
I don't think they need to rush anything for the gamers (in the high end spectrum) before Intel get in the mix...
Frenetic Pony
Veteran
It appears to be in the same vein as other, interconnected "neuron" deep learning chips. The hypothesis being that learning is local (short range) and neural nets only connect locally, so you can scale up cheaply with chiplets (or similar) that only have local interconnects, because they don't need direct access to chiplets that are further away. Which has nothing to do with how GPUs, in the "Graphical" sense work as they need far too much universal bandwidth, and especially raytracing which needs the opposite of local interconnects.
So yeah that has zero to do with multi GPUs, but may have something to do with Nvidia's future GPUs. In that if they find a non monolithic design to be efficient, at least in terms of cost vs performance, they might ditch using GPUs for deep learning altogether. Which might be good news for video game consumers, as they won't need to buy questionably useful deep learning cores on their gaming GPUs anymore.
Last edited:
https://www.digitimes.com/news/a20190605PD200.html
Nvidia to adopt Samsung 7nm EUV process in next-gen GPU production....
Nvidia to adopt Samsung 7nm EUV process in next-gen GPU production....
D
Deleted member 13524
Guest
That's been rumored for quite some time.
Megadrive1988
Veteran
Will Ampere succeed both Volta and Turing, or just Volta in the HPC space. I hope Ampere will do both, because Turing is not that great and needs to be replaced as soon as possible.
It was a long gap in time between Pascal and Turing, and it was still a long time between Volta for HPC and Turing for GeForce/consumer cards.
It was a long gap in time between Pascal and Turing, and it was still a long time between Volta for HPC and Turing for GeForce/consumer cards.
And on Sale(shelves) in May-Jun...?
- Status