Nvidia Ampere Discussion [2020-05-14]

Is your shop one of the bigger ones? I assume smaller shops are getting shafted even more than usual.

 

Small one, but anyway i know manufacturers (AIB) numbers in my country too, and i can assure this launch is the worst ever, not sure even if we can consider this a launch, definitely paper launch fits much better this time, cause even if the product exists, it only does on paper for most.

6 weeks since 3080s supposed launch, and more than 90% of the customers are still waiting for it, some of those since more than a month ago. 3090 could be better cause demand is low, and 3070 maybe a bit better, but also very bad even after 2 weeks delay.
Demand is high no doubt, but i wouldn´t say higher than for Pascal in 2016/2017, supply is much much worse, maybe even 1/10th of pascals on launch..

 

Better, but not that much better. Proshop operates in Sweden among other countries and even with the two weeks delay they've gotten under 500 cards from over 5600 they had ordered already ages ago.
https://www.proshop.de/RTX-30series-overview

 

The same people are probably just signing up on pre-orders everywhere until they get their card. As soon as they get the card they can just cancel all other preorders.

 

MicroCenter Manager says it seems like a purposeful slow down in distribution. Channel is dry.

Either there is a manufacturing issue or perhaps new cards on the way and NVidia wants to slow-walk RTX 30 until a new stepping is out? Channel has zero feedback of when, or why. (Like NVidia is holding back, on purpose)


I wonder if NVidia is doing this, because they are waiting for 3070 saturation, or maybe waiting on AMD to fully release Navi2?

 

How does Nvidia benefit from not selling stuff on purpose?

 

If you don't make good margin but have to have a product out there, one solution is to not sell a lot (until you find a way to make money on what you sell) ?

 

Very curious indeed. The very few 3070 cards in Finald sold out in minutes. No Nvidia cards available at all unless you want a 2060 or 2070.

 

Well even if we assume that to be true, selling at some margin is better than not selling. And what magic bullet could fix it in a quarter?

Nvidia has provided certain forward looking revenue estimates to the stock exchanges and their shareholders, and they have a fiduciary duty to their shareholders. It's very unlikely (and bordering on illegal) they are artificially limiting production for whatever reason. As someone already mentioned, the earnings call is due soon, we'll have some answers then I'm sure.

 

I don't think Ampere would really significantly impact Q3 even if supply is as bad as it seems. It's not like they stopped selling everything else.

 

Meaning professional and data center products? Gaming is around half their revenue and TU102 and TU104 parts are EOL and sold out everywhere. If they have no product to sell I don’t know how they will have a strong quarter.

 

Products below $400 exist.

 

Yea TU102 and 104 supply would have stopped by around mid Q3 or so and they'd have started shipping GA102 to partners. There is the laptop segment in addition to professional and HPC, where Nvidia does have near market dominance but I don't know if that's enough to make up for the shortfall in desktop. Agreed that it's probably not going to be a strong quarter.

 

7 days. We’ll see.

 

Can GA102 support GDDR6 (non-X)?

 

RTX A6000 will use GDDR6.

 

I thought it was interesting that in the Ampere whitepaper Nvidia clearly noted that GA104 can use GDDR6X as well, so it only makes sense that GA102 can go the other way. Sounds like they designed the memory controller blocks so that they're capable of both. Makes sense given that GDDR6X is so similar to regular GDDR6 - probably required very little area to make it backwards compatible with GDDR6 too.

Also means that if GA104 is even a little bit bandwidth bound, it's very easy for Nvidia to add another SKU above the 3070 using GA104 but with GDDR6X to combat AMD's launch.

 

It probably doesn't make sense to think of it in those terms anymore. Modern games are probably "everything" bound.

I was playing around a bit more with the Nsight profiler on my 1650 super in Doom. Flops per byte is about the same as the higher end cards so would expect similar behavior. This was just one frame from one game and there wasn't a lot happening on screen at the time but my high level observations were:

- the bottleneck shifts many times during the rendering of a single frame.
- the bottleneck is almost never FP32 ALUs.
- when shaders are running full tilt the texture units and L1 cache are usually also maxed out.
- peak integer ALU utilization was ~30%.
- L2 hit rates are pretty good (over 70%) for most of the frame.

Here are a few interesting samples to help demonstrate the point.

Spoiler: Early in the frame, might be the Z-pass?

Spoiler: Heavy bandwidth usage, viewport culling and vertex attribute fetch. Not sure what it's doing.

Spoiler: Heavy viewport culling, perspective correction and vertex attribute fetch near the middle of the frame.

Spoiler: High ALU usually accompanied by high TEX, maybe a compositing or post processing pixel shader.

Spoiler: Rare high ALU without TEX.

Spoiler: Peak INT usage.

 

Amazon EC2 P4d instances deep dive
November 2, 2020

https://aws.amazon.com/blogs/compute/amazon-ec2-p4d-instances-deep-dive/

 

When you say this are you referring to the "L2 Throughput" number that appears in the graphs or some other number?