News Proshop RTX 3070 Inventory Only 2.5% fulfilled

Shadowclash10

Prominent
May 3, 2020
184
46
610
Looks like Nvidia is getting clobbered by its decision to use Samsung 8nm, just can't seem to get the chips in meaningful quantities.
Isn't this the best case scenario though? I thought TSMC had a lot less capacity than Samsung? Meaning if they went with TSMC, the situation would be worse.
 

InvalidError

Titan
Moderator
Isn't this the best case scenario though? I thought TSMC had a lot less capacity than Samsung? Meaning if they went with TSMC, the situation would be worse.
Samsung may have lots of fab capacity but most of it is for DRAM and NAND manufacturing, not performance-oriented 8nm logic. Samsung has a long history of struggling to get new fab processes up to speed and losing contracts for cutting-edge process to TSMC, so its 8nm may be in a similar state to Intel's 10nm - more work still needed before high volume production.
 

Blacksad999

Reputable
Jun 28, 2020
70
48
4,570
We'll see, but I'd bet AMD could have rushed a release out the door like Nvidia did, but decided not to. They've been dealing with TSMC for quite awhile now, and with all the GPU's already made for the console releases, I'd wager that AMD probably stacked their deck as far as the PC Gpu release.
 

Olle P

Distinguished
Apr 7, 2010
720
61
19,090
Looks like Nvidia is getting clobbered by its decision to use Samsung 8nm, just can't seem to get the chips in meaningful quantities.
It's a rushed launch. Not certain that the current bottleneck is at GPU production since the card manufacturers were also invited late to the party. I do agree that GPU production most probably isn't stellar though...

Isn't this the best case scenario though? I thought TSMC had a lot less capacity than Samsung? Meaning if they went with TSMC, the situation would be worse.
There was production capability available when Nvidia began their price negotiations with TSMC.
Once they realised TSMC wasn't willing to go as low as they desired AMD and others had stepped in and purchased a lot of the capacity at a higher price point.
  • So the current production level is a "best case" given Nvidia's miscalculation on the demand for 7nm.
  • From a consumer point of view a better follow-up would have been to delay the sale start of these cards by a couple of months to give the AIB-partners time to design better cards and produce them in larger quantities before the release. By then the production of GPUs would (will) also have matured, resulting in better yields. A pre-mature sale start is by no means a "best case".
  • Nvidia is obviously interested in getting some advantage over AMD by getting first to market, and then also making more money by having a high demand paired with low availability so that customers are willing to pay more than the MSRP just to get any card. (See for example the user reviews on Newegg. Each $700 RTX 3080 model has far fewer reviews than the >$800 models.)
 

bigdragon

Distinguished
Oct 19, 2011
1,107
547
20,160
Ouch. Looks like the Ampere supply problem is an even bigger problem than I realized. I was excited about the idea of upgrading from a 1070 to a 30-series card, but these supply issues and the capacitor stability design flaw has me thinking Ampere is tainted. Nvidia really botched this launch.
 

InvalidError

Titan
Moderator
It's a rushed launch. Not certain that the current bottleneck is at GPU production since the card manufacturers were also invited late to the party. I do agree that GPU production most probably isn't stellar though...
The most recent numbers I could find about Samsung's 8nm yields were from analysts estimating it to be around 30% at the beginning of the year. Combine that with 400-600sqmm dies, you burn through wafers pretty quickly. If yields are still anywhere near that bad, production may be further limited by incremental tweaking between batches slowing things down in hopes of improving yields over time.

Since fabs usually sell capacity by the wafers-per-month, I could imagine Samsung and Nvidia come to a mutual agreement to limit the number of fab lines assigned to GA102-106 until yields improve: Samsung gets spared possibly having to compensate Nvidia for sub-par yields, Nvidia gets to have more usable dies per wafer at no extra cost in the long run by not having to pick up as many sub-par wafers.