Nvidia Ampere Discussion [2020-05-14]

Do we know how many of these orders are legit, vs "bot"-order ?

 

No two ways about it, this is a joke of a launch.

 

No, but I think the more relevant part is how much the shop has ordered cards (a lot more than they have standing orders) and how many of them they've actually received.

 

This stat would be a lot more interesting if the same shop would've provide similar data for Turing and Navi launches. Without it it's impossible to say if Ampere is any different to previous launches even in this shop specifically.

 

https://www.prnewswire.com/news-rel...nerally-available-in-the-cloud-301135956.html

At the same time the A100 GPUs produced at TSMC, not Samsung, are shipping to cloud providers and other HPC applications as scheduled, and in decent quantities.

Shall we play a game of guess? How many GA102 chips have been successfully been produced so far? Is it even 5000 pieces in total?

At a 500 bucks per piece, that's how much revenue for NVidia? Just shy of 2.5 million USD? At an R&D investment of just how many billion USD? And how much of a loss is this going to be if demand tanks in a month or two?

 

GA100 went in production ~5 months prio GA102.
And demand for GA102 will be huge. The workstation and virtualisation market is locked. A6000 is >2x faster than RTX8000 with compute or raytracing.

 

So? You would assume that whoever needs a GPU for raytracing in a professional setup had already bought into Turing series. For a theoretical factor 2 uplift, best case, it's a heavy investment and the reality is that this and next years IT budget has mostly been blown out already with the need to send tons of employees into home office quickly, plus a generally extremely weak economy.

GPU (not pure compute) virtualization is something NVidia certainly hopes to become a big player in one day, but a couple of high-profile but low-volume (and probably highly discounted) projects like Stadia etc. aren't going to make much of a difference there. NVidia made sure to make virtualization completely unattractive compared to sticking a dedicated consumer GPU into a (still rack-mounted) workstation where needed. And plenty of solutions just don't rely on virtualization at all, but go by a simple client-server model in user space which works just fine with consumer GPUs in a multi-GPU setup as well.

In realistic terms, there may be a global demand for about 50k GA102 chips, at the current point. That's still a poor joke. And a number NVidia might not even be able to source till the end of this year. Demand just looks so horribly inflated at the moment due to a complete lack of supply.

 

The pandemic will speed up cloud virtualisation for workstations. No company will allow their employees to work from home with a desktop pc. These companies will put more money into their own private cloud infrastructure.
And before Covid-19 nVidia's Quadro business was on the path to a run rate of $1.2 billion a year.

 

LOL. Missing some zeros there.

 

I've heard that it's 5. Just 5. And there won't be any more till 2022.

Seriously, what is the point of such guesses?

Why? Everyone will die or something?

 

Overclockers UK had 1000+ Navi in stock for the launch of 5700 series. I've already brought this up and linked the post on that forum, about 2 weeks ago, if you care to scroll back. Might be in the other Ampere thread...

This store ordered more than 8000 3080s, before it opened up preorders (again, I've already posted about this and linked the numbers a few days ago).

 

They’re one of the bigger fish. How many 3080’s did they get?

 

Unique timespy or similar benchmark submission scores would give indication of minimum number of sold units assuming 100% of owners would have tried the benchmark. Not sure how to figure out unique results but simple search gives 26k hits. There is bound to be a lot of duplicates?

https://www.3dmark.com/search#advanced?test=spy P&cpuId=&gpuId=1338&gpuCount=0&deviceType=ALL&memoryChannels=0&country=&scoreType=overallScore&hofMode=false&showInvalidResults=false&freeParams=&minGpuCoreClock=&maxGpuCoreClock=&minGpuMemClock=&maxGpuMemClock=&minCpuClock=&maxCpuClock=

 

I last visited that forum on the Saturday (morning I think it was) after 3080 went for sale and counted between 20 and 25 people who either had the card already or had delivery confirmation. Some people had the card on launch day, as Overclockers allows people to visit and collect orders.

At the time Overclockers made it clear to everyone:

• NVidia disallowed the store from publicising the numbers of cards that were in stock at launch time or sold (which is why I'm surprised by the very detailed numbers that were posted this week)
• NVidia also disallowed on-going reports of quantities being delivered to the store
  
• no customers would be told their position in the "preorder queue"

There's an owners' thread for 3080 owners (there is usually a thread like this for each of the major graphics card launches - organised by members of the public) so it would be possible to get an up-to-date count of people. Obviously not everyone who purchases actually posts on the forum.

Note that the store made all 3080 product listings "pre-order only", even though cards were in stock. A public thread announced that this was a tactic to "favour forum members". The idea was that forum members would know that "pre-order" was a lie, at least when launch time arrived, and that at least some non-forum members would choose not to order when faced by "pre-order". Then it was just a matter of how many forum members piled in.

Shortly after the time that 3080 sales went live the site crashed - error messages indicated that about 400 site visitors were present at the time of the crash.

There then followed several days of people complaining about the price changing when they checked out and discovering that their order really was nothing more than a pre-order. It seemed pretty angry in there. As I posted at the time (on this forum), Overclockers seemed to be doing some real damage to its reputation.

I haven't been back as the signal to noise ratio will be extremely low.

 

So on one side we have these guys saying they spoke to some vendors who claimed they got shipped similar numbers as Turing cards during this launch (ie it's all just conspiracy theory and the demand is just astronomical).

OTOH we have Proshop making public what seems to be ridiculously low amounts of cards received, and Overclockers UK saying nvidia isn't even letting them show their numbers.

Humm...

 

Your estimation is the joke here...

 

Nvidia’s fiscal quarter runs through Oct 31st and they report financials mid-Nov, so in about 5 weeks we should find out whether they managed to ship the 5,000 GPU yet, thus satisfying 10% of lifetime demand for the GA102.

 

Looks to be around 100-200 since the start of October according to this: https://docs.google.com/spreadsheet...dTeJ2iLAOBR-yVURCgbsTn5LSU/edit#gid=761772307

From this post: https://www.overclockers.co.uk/forums/posts/34050370/

The preorders seem to be in the 3000+ range. I'm number 264 in the queue for a TUF (non OC) but judging by the lack of shipments I don't expect to get my card by the end of this month really.

 

I'm just going to throw out another angle to consider for availability and shipments. What would be interesting to know is what the situation is like in China. The allocation split and priority in 2020 could be very different compared to just a few years back in say 2016 due to the growth of the Chinese market. A compounding effect would be the current logistics situation with respect to Covid and proximity (place of origin). Also I believe that China roughly at the moment is going through a holiday sales period?

I'd also wondering what the situation is with respect to ocean freight shipments. There are indicators that the lead times currently are rather large, so follow on higher volume bulk shipments which would come from ocean freight might be later compared to past launches.

 

It was the golden week in China last week which affects shipments for next week afaik. Nvidia should've factored all of this and delayed the launch until they had adequate stock. They just wanted to launch a product before AMD so that they couldn't use "fastest GPU" or whatever for RDNA2 marketing. If anything, they ended up pissing people off and potentially driving them to RDNA2 if that's not a paper launch as well. I don't expect any decent stock for Ampere until December with the way things are going, if AMD have decent stock for RDNA2 I can see a lot of people jumping ship.