The TRUTH About the GPU Shortage
Introduction
Chances are, if you clicked onto this video, you probably have some idea of the ongoing chip “crisis” that’s affecting everyone’s ability to get their hands on the latest and greatest GPUs
It’s something we all hoped would be a brief issue around the launch of the first RTX 3000 series and RX 6000 series cards – much like a typical spike in demand for a new product
And yet, here we are 9 months later, with (at time of filming) no obvious end in sight, and news stories coming out constantly showing more impacts beyond the PC gaming world
No matter what retailers, manufacturers and famous tech Youtubers do, the majority of gamers can’t get their hands on a GPU at MSRP
This leads to either settling for expensive second hand cards, or succumbing to ludicrous scalper priced new cards (or, continuing to have hope and wait)
So today I wanted to dive deeper into some of the finer details of this dire situation, and try to provide the best estimates for when “normality” may resume in the GPU world
GPU Examples
Unfortunately, what we have right now though… is far from normal; GPUs are usually a depreciating asset – you buy it, and it immediately loses value… that… isn’t the case anymore
Take last gens RTX 2060 6GB (prop/b-roll optional) – which launched about 2 and a half years ago in January 2019 for an MSRP of £270
If you want to get your hands on one of these today… you’d need about £470 – £200 more (https://www.ebay.co.uk/itm/393386933753?hash=item5b97b03df9%3Ag%3AcO4AAOSwKypgwMoK&LH_BIN=1&LH_ItemCondition=3000)
How about a GTX 1060 6GB (again, prop/b-roll optional) – a now nearly 5 year old card – which launched in July 2016 at an MSRP of £239
Second hand price today? It does vary a little, but try roughly £289 – £50 more expensive than when it was brand new. (https://www.ebay.co.uk/itm/265180578535?hash=item3dbdfe82e7%3Ag%3AyRUAAOSw%7EVFguMnf&LH_BIN=1)
Used market for the current gen RTX 3060 (prop/b-roll) then? If you bought one at the £300 MSRP and used it… well the “good” news is you can still double your money on the second hand market for… £600 or more (https://www.ebay.co.uk/itm/224489635311?epid=12044882484&hash=item34449ffdef%3Ag%3ACWcAAOSwmDtgvnrT&LH_BIN=1&LH_ItemCondition=3000)
If you haven’t managed to get your hands on one for MSRP though… then I’m not surprised – and you definitely are in the majority
But wait, there’s more
These kinda shocking examples are just the start though, and as I alluded to earlier, GPUs are not alone in being affected
The automotive industry has been another high profile victim, with widely reported issues relating to the low supply of chips; which has led to a serious knock on impact for manufacturers
Audi for one, made 10,000 fewer cars in Q1 2021 and had to furlough 10,000 workers, all as a result of not having sufficient chips to use in their new vehicles (https://www.bbc.co.uk/news/technology-55704936)
Ford on the other hand had to cancel shifts at two of its car plants, with their predictions indicating a potential 2.5 billion dollar hit to their profits this year, and that’s after posting a 1.2 billion net loss last year (https://www.theguardian.com/business/2021/mar/21/global-shortage-in-computer-chips-reaches-crisis-point , https://sec.report/Document/0000037996-21-000012/)
In perhaps the worst case I’ve found yet, General Motors has had to outright stop production of the Chevy Camaro and two models of Cadillac, diverting supply of what chips it has to it’s higher volume models (oh, and they closed an entire car plant too) (https://drivetribe.com/p/chevrolet-halts-all-camaro-productions-RCmB8MnlTTib4sOaZPG4fA?iid=NudH2Y1HT7S5duLOzSLBCg)
The chip shortage as a whole has been estimated recently to cost the automotive industry 110 billion dollars – that’s 80% higher than the original prediction that was made in January this year, indicating that problem has got worse over the course of 2021 (https://www.cnbc.com/2021/05/14/chip-shortage-expected-to-cost-auto-industry-110-billion-in-2021.html)
Luckily for many of us, the used car market is relatively plentiful, although there is some anecdotal evidence that prices for used cars are creeping up as more buyers opt not to buy new vehicles, owing to the supply chain problems being experienced (https://www.vox.com/the-goods/21507739/coronavirus-car-market-used-expensive)
And as one final “honourable mention” for a non-GPU victim, we have the next gen of consoles. Many parents would have hoped that the console shortage at Christmas was usual “hype” over a new shiny thing, which would die down before their kids birthday time came around
But… as you may have already guessed, it’s near enough as hard to get a console as it is to get a GPU which after everything you’ve heard so far, shouldn’t be too surprising (https://qz.com/1996287/why-you-still-cant-get-a-ps5/)
And with that, PC gamers and console gamers have finally been united by a common enemy
Context
But… who or what should we actually be sharpening our pitchforks for? Well… here’s where it gets complicated as there’s not really one easy culprit to pin down for the chip shortage
First off, we have the global elephant in the room: the pandemic.
COVID hit the world like a freight train, disrupting supply chains as factories were forced to close due to lockdowns, employees off sick with COVID or quarantining, and incredibly volatile demand spikes.
China was the first region impacted heavily by the virus forcing businesses to halt production while everyone quarantined.
The trouble is, China manufactures around 30% of the entire world’s products so no output from them means not only no stock of end user products, but also knock on effects where companies that do build products elsewhere can’t get the components they need.
The other obvious cause, at least for us gamers, is crypto mining.
A perfect storm of new hardware and surging values of the main coins means profitability skyrocketed.
Everyone and their dog heard about bitcoin back in 2018 with the last price surge, so anyone with even slight technical knowledge used this spike as an entry point and started mining, often on an entire industrial scale.
An example of that is the story from Birmingham of a Bitcoin mining farm that was raided by the police as they thought it was a cannabis farm, but turned out to be 100 ASICs mining away on stolen electricity, ran by “three nerds”.
(https://www.bbc.co.uk/news/uk-england-birmingham-57280115)
Lastly, the problem with purchasing products that we are all painfully aware of, bots and scalpers.
Writing a bot to buy up stock the very instant it becomes available isn’t overly hard these days, so unscrupulous scalpers looking to make easy money or crypto miners looking to score a batch of new cards are using bots to snap up all the stock in milliseconds.
These scalpers are also helping to restrict supply, creating artificial scarcity driving even legitimate sellers’ automatic pricing models to skyrocket.
Then the vicious cycle ticks over, as the scalpers can then increase prices even more… and well, you know the rest.
Nuance
What you might not know though, are some of the more obscure factors beneath the surface which not everyone talks about when discussing this situation.
A less obvious cause, not only of GPU shortages, but of the whole ‘global chip shortage’ is the unprecedented demand for chips. Everything these days from your phone to your fridge needs microcontrollers, NAND flash, and SoCs.
Cars in particular use a lot of lower spec chips for the hundreds of modules in each vehicle, and are now using more powerful CPU+GPU units to run their digital dashboard and infotainment displays. The number of silicon dies needed for all of these new pieces of tech is growing massively year on year; at a rate seemingly faster than chip fabs can keep up with.
On top of that, COVID caused a ‘gooseneck’ in demand on top of the natural increase in growth. At the start of the pandemic, company’s predictions were that COVID would significantly reduce demand so many cut down their orders of everything from raw materials to chips.
However, their predictions turned out to be wrong, and demand increased significantly as many people found themselves with more disposable income for material goods, given their usual foreign holidays and leisure venues were no longer available.
Inevitably this included items needing (you guessed it) chips, such as cars, big TVs and shiny new gaming computers. All of the manufacturers that had cut their orders with fabs like TSMC then found themselves at the back of the queue trying to increase their supply to meet demand.
Some firms were savvy and delayed product launches long enough to accrue plenty of stock to fulfil demand, like Apple who slightly delayed the launch of the iPhone 12 ensuring bountiful stock even at launch.
That increased demand then has a knock on effect of longer lead times between orders and fulfilment across the board. These increased steadily over 2020 as demand naturally rose and production capacity didn’t, but it skyrocketed at the start of 2021.
The average lead time in January for a microcontroller was around 16 weeks, but by April it was 44 weeks – that’s nearly a year, making annual product cycles basically impossible. (https://levadata.com/how-the-semiconductor-shortage-increases-lead-time-generates-supply-chain-risks)
Bottlenecks in the supply chain can also extend those wait times, for example ASML Holding NV who brand themselves as “The world’s supplier to the semiconductor industry” hold a virtual monopoly on the cutting edge photolithography equipment needed to manufacture modern, small process node designs.
Another example is Shin-Etsu Chemical Co, who supply the majority of chemicals used in semiconductor manufacturing, so if either of these firms have delays, there is a massive chain reaction slowing the entire world down.
(https://www.bloomberg.com/graphics/2021-semiconductors-chips-shortage/)
To top it all off, the largest chip fabrication company in the world, Taiwanese Semiconductor Manufacturing Company or TSMC as it’s more commonly known, is facing the worst drought the nation has seen in 56 years.
TSMC uses 156,000 tonnes of water PER DAY, and the Taiwanese government has asked them to reduce their usage by as much as 15% compounding the shortage issues for obvious reasons.
(https://asia.nikkei.com/Business/Tech/Semiconductors/TSMC-tackles-Taiwan-drought-with-plant-to-reuse-water-for-chips) (https://www.bbc.co.uk/news/world-asia-56798308)
Even if you can build your own products without any bottlenecks, there is a good chance you won’t be able to ship them anywhere once they’re made. As demand for physical goods has increased, space on cargo ships has vanished.
This is partially also due to medical equipment taking priority, and, thanks to a sharp decline in passenger air transport, air freight capacity has plummeted too as much of the space in a passenger plane’s hold is sold as cargo space but less passenger planes flying means less capacity.
And again, what capacity there is is often being utilised for life saving medical equipment anyway so… good luck getting affordable transit to get your products in consumers hands.
The Suez canal blockage also didn’t help, causing billions of dollars worth of goods to be delayed for weeks or months as the ships had to go the long way round Africa. And on top of all of that, there is a shortage of shipping containers. Containers aren’t being loaded back onto ships that are heading back to Asia, instead piling up in already congested ports. A UK firm quoted 6-8 week delays just on the shipping containers.
(https://www.ship-technology.com/features/global-shipping-container-shortage-the-story-so-far/)
All of these issues slow things down, but the real nail in the coffin for manufacturing is a technique called Just-In-Time production. This was first pioneered by Toyota and it works like this:
The factory that makes rubber for the tyres should only ship a batch to Michelin when the wheels have been made and are ready to have tyres fitted, and those wheels should only be made when the rest of the car is ready to be assembled, which should only be built when dealers are ready to have more vehicles on site. Each step in the process runs “just in time”, so all the components you need should arrive exactly when they are needed as holding stock costs money and slows your ability to pivot to market demand.
The trouble here is that most companies do it wrong without much extra thought. Toyota saw its reliance on chips and held between 3 and 6 months inventory, knowing they’d still be useful no matter what happened, which is the reason they are the only automaker that hasn’t seen much impact from the shortage. They haven’t had to shut plants or cut production, unlike most other automakers. Everyone else held no stockpiles, meaning they were fully at the whim of their suppliers as I touched on earlier.
(https://www.reuters.com/article/us-japan-fukushima-anniversary-toyota-in-idUSKBN2B1005)
On top of that, there are so few chip fabs in the world, and even fewer that can produce the bleeding edge process node sizes modern tech, including GPUs require. TSMC is the main source, with your other option being Samsung if you want a 7nm chip.
If you want 5nm, again your only options are TSMC or Samsung. Even for larger process nodes SMIC, UMC and Global Foundries are your main choices, but by market share TSMC controls some 54% of the foundry market.
TSMC’s main customers are Apple with 25% share, AMD with 9%, Mediatek and Broadcom with 8%, Intel and Qualcomm with around 7% and NVIDIA with 6%. If you aren’t one of those brands, or buying from them, you are going to have a rough time trying to get your chips built.
In my best attempt to channel my inner TechTechPotato, I thought it’d be interesting to look at the production capacity of TSMC and Samsung, which for the 7nm node NVIDIA is using from Samsung for their RTX 3000 cards, is about 40,000 wafers per month.
Taking a look at an image of both an RTX 3080 and RTX 3070 die, using at GDDR6X package as a measurement – they are 14mm by 12mm – we can measure the pixels of the GDDR6X, then the die size, so we can get the X and Y dimensions.
For an RTX 3080, a GA102 die, a rough estimate puts it at 27mm by 23.5mm. That nets roughly the same die size quoted only a couple mm2 off. For an RTX 3070, the same process offers 20mm x 19.5mm.
If you plug those numbers into the caly technologies die per wafer calculator, with 300mm wafers and a defect density of 0.09, you’ll see that including defect dies, you can get a maximum of 79 GA102 dies per wafer, and 137 GA104 dies.
(https://caly-technologies.com/die-yield-calculator/)
I’ve included the defect dies here as the benefit of using the same die for different SKUs is that you can absorb those defects and just sell that as a lower end model with cores disabled. Of course some will be fully broken, but many can be tiered down.
That’s the first piece of data, the second is how many were made. Using the Steam hardware survey, RTX 3080s account for 0.9% of GPUs using Steam, with 3070s holding 1.48%. Rough napkin maths says that there were likely 500,000 and 1 million RTX 3070’s produced, versus between 250,000 and 750,000 RTX 3080s. Now putting those two data points together…
Best case, where you can get the maximum useful dies out of a wafer and they sold fewer cards, NVIDIA bought 3,650 wafers from Samsung for the 3070. Worst case, where defect dies weren’t useful and they sold the maximum, try 10,000 wafers. For the 3080? Best case is 3,200 wafers, worst case is 16,000.
Taking a look at Samsung’s 7nm capacity again, even the lower estimates would take 15% of Samsung’s monthly capacity just for these two cards. In reality it’s likely somewhere in the middle of my estimates meaning an even larger share.
Of course these cards weren’t all built in a month, they will have been built over 6-12 months in advance of the launch – but therein lies the issue. Lead times at TSMC are reported to be between 6 and 12 months for these sorts of chips right now, and assuming the same can be said for Samsung, NVIDIA won’t be able to dramatically increase supply any time soon.
They also don’t want to overcorrect, creating far too much stock for the market, so are likely hedging their bets with steady increases in supply for the next few months, rather than trying to go all out and double the number of units produced in a short space of time.
(https://www.tomshardware.com/news/report-tsmcs-lead-time-for-7nm-chips-is-still-6-months-or-more)
And, specifically for NVIDIA, they outright said they didn’t build enough on launch. Quote, “The demand for the GeForce RTX 3080 was truly unprecedented. We and our partners underestimated it.” So few people bought the first generation RTX cards that NVIDIA under forecasted demand and manufactured far too few dies, meaning board partners couldn’t build enough cards.
(https://www.nvidia.com/en-gb/geforce/news/rtx-3080-qa/)
It was even claimed they rushed the launch, so much so that their AIBs didn’t have enough time to fully test their designs so issues like the Crash-To-Desktop crept in, caused by incorrect capacitors, something their testing would (and eventually did) catch and correct.
(https://videocardz.com/newz/manufacturers-respond-to-geforce-rtx-3080-3090-crash-to-desktop-issues)
Legacy
So, what does the future hold?
Well, according to Intel’s CEO Pat Gelsinger, the shortage is likely to “take a couple of years” to resolve itself, mostly through foundry capacity increases. Intel themselves are building a new fab in Arizona and will be offering their services to others, much like Samsung already does.
(https://www.reuters.com/technology/intel-reiterates-chip-supply-shortages-could-last-several-years-2021-05-31/) (https://www.intc.com/news-events/press-releases/detail/1452/intel-announces-major-expansion-in-arizona)
TSMC is looking to increase production capacity too, investing $100 billion to increase their output over the next three years. That kind of investment, coupled with a slowly saturated market, should allow supply to increase and help decrease demand too.
(https://www.livemint.com/industry/manufacturing/tsmc-to-invest-100-billion-to-increase-semiconductor-output-11617337683007.html)
Unfortunately for us though, it’s all a waiting game. Thanks to the longer lead times, there is literally a shortage of machines that build machines that make products, so even to get small increases in production capacity is still incredibly difficult.
Wrap up
With all that in mind, where does that leave the humble PC gamer in 2021?
Well, right this second, if you want a new GPU, buy a prebuilt PC. I’ve got affiliate links to OverclockersUK in the description, who list which items in their builds are in stock and they have GPUs available. Of course there are plenty of other SIs you can choose from too, but that is your best bet to get a new card right now.
If you can wait a little, stock levels do seem to be slowly returning. Scan currently lists 112 preorders outstanding for various RTX 3080 models, which is down from the thousand plus at launch. So, in theory, getting in a queue might net you a card… eventually.
(https://www.scan.co.uk/shops/nvidia/rtx-30-series-faqs)
Hopefully companies like AMD and NVIDIA have learned a difficult lesson from this experience. It was a perfect storm of knock on effects, but at least now there’s a chance they’ll be more prepared in the future… once current issues are actually resolved…
