As I noted in my last Sidenote post about my own system build, I am at the point a lot of builders are currently – waiting for a CPU and GPU to become available and to fight off the hordes of bots, scalpers, and other fellow desperate PC enthusiasts begging to get something to finish their current system builds.
As things have progressed from the iffy rollout of the Ampere-based Geforce RTX 3000 series cards to AMD’s Ryzen 5000 and Radeon RX 6000 rollouts all being similarly impacted, and coupled with the incredible constraints of console supply that sees systems being snapped up before most people can see the box in person, there are a few trends that are obviously affecting everyone. Rather than simply blame the obvious squiggly globe virus in the room (not my room and hopefully not yours either!), let’s discuss the factors it caused or exacerbated, and then the company-specific ones, as there is a mix of business priorities, iffy decision making, and a global pandemic and the knock-on effects that caused! So we’ll start at a company level for Nvidia and AMD, then drill into general trends, and then note an exception in Apple and how their product launches for late 2020 have been relatively unimpeded.
New Process, Unexpected Process Node: In 2019, Nvidia was reported to have inked a deal with Samsung to manufacture their newest GPUs using their 7 nanometer extreme ultraviolet lithography process node. It was a node that Samsung had spun up, but was marred by reported poor yield rates making it a difficult process to build a business upon. Yet, Nvidia was doing something that was fairly common in the semiconductor business at that point in time – threatening publicly to leave the TSMC manufacturing nodes behind for lower-cost Samsung silicon manufacturing. As TSMC has been strongly booked up as-is, the bluff was called and Nvidia was indeed left with manufacturing on Samsung.
However, as we all know now, Nvidia isn’t building gaming GPUs on Samsung 7nm at all. Instead, they’re using 8nm – a branch of Samsung’s 10nm process that makes further density and power characteristic improvements, but not nearly to the extent of the 7nm process, their next real node advancement after 10nm. Why is that?
Well, the yield rate must have been a concern from the beginning, as the silicon die process design tools used for Samsung are incompatible between the 10nm and derivative nodes and the 7nm family, meaning that early in the process of designing Ampere, Nvidia had to have known that they weren’t using 7nm. 8nm seems to be marginally better yielding, at least to the extent that such information is available in the public sphere – but it also offers far worse power characteristics and thus, performance. Not to say that Ampere is a slouch (it certainly is not) but to simply say that Nvidia could have had a far higher performing part on 7nm with Samsung, much less the TSMC node they attempted to haggle on.
Indications seem to be that Nvidia simply isn’t getting enough chips to fill their demand, and while not all of that rests on Samsung (we’ll discuss more on this point later), some of the blame certainly does. It seems to indicate some things about yields at Samsung, given that the lower-tier, smaller die 3070 and 3060Ti have been far more supplied already to retail than the 3080 and 3090, from some reports, and while demand is slowly being satiated, all 4 Ampere GPU models remain instant sell-outs at any restocking.
GDDR6X – The X is for Xtremely Low Availability: GDDR6X, while not entirely the cause of Nvidia’s supply woes, seems to be at least a factor. Both the 3080 and 3090 use a large number of maximum capacity chips, and Micron only spun up production over the summer by most reports. While DRAM is far easier to yield and manufacture compared to a full GPU die, the amount of silicon used by the VRAM on both cards is more than the GPU itself, and mid-summer manufacturing spin-up, during a global pandemic, means reduced workforce, reduced material availability, and less time to stockpile for a major launch. While it doesn’t seem to be the single or even most important bottleneck to the Ampere card availability, it is worth noting that the more-available 3070 and 3060Ti both use existing GDDR6 supply instead of the new 6X standard, and I have to imagine that is at least a part of the reason that Nvidia remains supply constrained on the high-end cards.
Competition Makes Things Harder: A lot of the rumor mill (and these are rumors, mind you) suggests that a part of why Nvidia has been short on supply of the 3080 and 3070 is that both cards are due for refreshes already. Rumors continue to swirl about a 3070 Ti and a 3080 Ti or the return of the rumored 3080 20GB card that was whispered about and then squashed before making it out. The reasoning is simple – while AMD doesn’t establish a clear lead over the 3080 with the RX 6800 XT, there are some titles where they do, and some cases where even the RX 6800 does the same over the 3080. Meanwhile, the RX 6800 does cleanly beat its RTX 3070 competition. While AMD does not have a strong footing in the market and their competition isn’t a clear win for AMD, Nvidia would probably like to counterpunch for when market conditions start to stabilize. The rumors have it that this remains a part of why 3080 and 3070 restocks are less common with the 3060Ti launched and the 3090 remaining an immediate sellout at over twice the MSRP of the 3080 – Nvidia is potentially harvesting dies that meet the requirements of a refreshed lineup in order to launch with some volume. Given that CES (virtually this year) is less than a month away, it doesn’t seem that farfetched.
TSMC Capacity and AMD’s Portfolio: AMD is suffering on multiple fronts from product availability issues, because their supply of wafers on TSMC 7nm is forced into a many-way split, including at least two contracts with likely minimum requirements and milestone bonuses. Currently, while AMD is TSMC’s biggest 7nm customer, they have to split their wafer allocation at the foundry between the following products:
Zen 2 CPU Chiplets
Zen 3 CPU Chiplets
Ryzen 4000/5000 Mobile CPUs
RX 5000 series Navi GPUs that remain in production at the lower tier
RX 6000 series Navi GPUs that just launched
Playstation 5 SoCs
Xbox Series X SoCs
Xbox Series S SoCs
AMD Instinct Datacenter Accelerators
So…yeah. That is, by my count, 10 different products that all need to be manufactured at one place on their supply of wafers. Some of these can obviously be low-volume products (the Zen 2 chiplets are surely near EOL on manufacturing and remain in only for XT refresh desktop CPUs and Rome server chips, RX 5000 has small dies that seem to remain in stock but the RX 5700 series is said to have already ceased GPU production, and Instinct datacenter accelerators were just announced around a month ago and don’t seem to be listed as shipping just yet with any datacenter vendor I checked with), but there are some titans in the mix. Firstly, the major scourge of the PC gamer in all of this is the console SoCs, with some rumors and reports suggesting as much as 70% of AMD’s wafer allocation went to these parts, likely to maintain supply agreements or to meet a bonus milestone from Sony and Microsoft. While they won’t remain in 70% manufacturing mode forever, they at least are doing so during a time where they have also launched two high-demand PC DIY parts – the Ryzen 5000 series CPUs and the Radeon RX 6000 GPUs.
Some rounded-up math using my napkin math for the Navi 21 RX 6000 dies suggests that 70% of AMD’s 7nm capacity in a month could yield as many as 5 million working console SoCs, which would be a lot – not enough to satiate current demand, but enough that if Sony and Microsoft could manufacture final machines for each of those, then that would be a boon to supply.
And that leaves the mainline PC DIY market. If we assume that the remaining 30% of wafer supply goes fully into Ryzen 5000 and Radeon RX 6000 series parts, it isn’t pretty on the GPU front, as it signals that AMD could manufacture maybe 10,000 graphics cards in a month. That’s not a small number, but for a launch with potentially a million or more customers? Yikes. The CPU front is much better, helped by the fact that manufacturing yields substantially more Zen core chiplet dies than it does GPU dies, by a lot, and half of the Ryzen 5000 lineup only needs 1 of those dies compared to two in the higher end parts, but this can also be a weakness. If you’re AMD, and facing a big opportunity to compete against Nvidia and push margins up on higher-end graphics cards, it seems like a shot you take. However, taking it comes with an opportunity cost – you have to decide to short manufacturing somewhere else to make more GPUs, but all of your products right now are selling like hotcakes!
We’ve already seen one victim of this – the “Milan” codenamed Epyc server CPUs have been delayed into next year, which means that a large number of the Zen 3 chiplets made this year can go to desktop consumer CPUs (or that a ton of chiplets are being binned for volume launch on Epyc), but my suspicion is that AMD is playing moment to moment right now. Milan being delayed means Zen 3 needs less production capacity today, because you can still make an absolute truckload of desktop CPUs off of even just 1 or 2 wafers worth of silicon, which means you can allocate more to Navi 21 manufacturing, while keeping Instinct, the laptop silicon, and the end-of-life stuff at low or even no volume while the massive majority of your supply is eaten up by console SoCs, a situation which likely eases in Q1 2021 and allows AMD to start pushing supply of other parts, at least some of which should even be available to buy easily!
Either way, that silicon supply situation is a mess, and I don’t envy the decision makers who have to fix that up, even as I acknowledge that AMD themselves chose to stagger their launches this way and force an artificial silicon competition!
AMD and Hype: The Radeon team did a much better job with this GPU launch of managing hype and expectations – putting out fairly solid internal benchmark numbers, making a solid reference card (with a couple of baffling choices that aren’t showstopping, like the suffocating baseplate design), and shipping drivers in generally better state than they’ve been in at a Radeon launch in a while. However, newest member of the executive team, Frank Azor, got AMD into hot water with fans by claiming the launch would not be a paper launch thusly:
Now, the idea of a paper launch is generally that there is no availability at all – that the company says you can buy cards on X date, no one buys anything, and everyone is sold out anyways. I think the term is overused – I wouldn’t claim that either Nvidia or AMD had a paper launch, but I do think that Azor left his tweet open to broad interpretation, not helped by the fact that indeed when both Ryzen 5000 and Radeon RX 6000 parts launched, seemingly no one could get them. Yes, we can see confirmations that a number of non-reviewers do have cards and CPUs in hand, but that doesn’t salve the wound of being really excited for a launch only to get out-F5ed by someone else, or to lineup at a retail store in the middle of a pandemic only for them to have 12 total cards available (which was, indeed, the situation at seemingly every Microcenter store in the US).
It then salts the wound to go back to Twitter and show your purchase receipt for your employer’s product as a “see, I told you so!” which Azor definitely did do:
So…yeah. Kind of a dick move, and exacerbating it further – this dude works for AMD! He could (and probably does have!) access to an employee card of some sort. Buying a card online out of the initial retail sales takes that card from someone else to prove a stupid point that most people don’t even believe you proved anyways – because who knows if you had the link early as an employee, or were on their network and able to order easily ahead of the rush, or hell, maybe the receipt is a fabrication!
I don’t find it a coincidence that his last day Tweeting was on the Radeon RX 6800 launch day. For the good of the company image, maybe it ought to stay that way?
Global Factors on Supply
Reduced Supply of Materials: This one is simple enough, but also gets messy. While it is true that manufacturing for most technology was only slowed during the early part of 2020 as COVID-19 was largely contained through highly-enforced stay-home orders and being in regions where people aren’t, well…Americans…it is the case that these small gaps in production have led to huge knock-on effects down the chain. Silicon wafer supply stayed steady, but production of chip substrates, capacitors of varying types, printed circuit boards, raw materials like metals for various uses, and a myriad of other inter-related shortages or reductions in supply cause problems. Nvidia’s CFO has now publicly indicated that the company is short both on chips for their RTX 3000 series but also on core manufacturing components, specifically substrates, which means that even if they got an increased supply of usable GPU dies, they might still not be able to manufacture that through to a finished product. When Nvidia (or AMD on the GPU side) sells to a board partner, they typically sell a BOM (bill of materials) kit which includes a packaged GPU on substrate ready to be soldered to the graphics card and the VRAM allocation for that product. If Nvidia or AMD can’t get substrates, they can’t really sell the BOM kit – they instead have to wait for supply, and there is no financial incentive for AIB partners to buy substrates for themselves – most chip supply for product lines like GPUs, motherboards, and other cards, either come in flip-chip packaging or pre-mounted to a substrate of a varying size and pinout requirement.
If you’re AMD or Nvidia, you might be stuck waiting on a substrate, raw copper or other materials for pins and traces, surface-mounted capacitors for power supply and input filtering, or any number of other small pieces. Both Nvidia and AMD GPUs this generation use metal shims around the GPU packaging to reduce mounting stress, and while that seems trivial, it requires specific materials engineering to work in the role it is assigned, meaning that reduced supply can also impact that very simple aspect of the design.
Even assuming you have a nice supply chain feeding the needed parts in, there is another logistical concern looming…
Getting Cards Out of Asia and Into the Global Market is Harder Now: In the wake of coronavirus, a number of travel restrictions and cleaning requirements have caused delays and reductions in capacity. Crews onboard these boats must be socially distanced and isolated, screened at all ports, and wear masks when performing their duties. Boats that disembark in the United States require following CDC guidelines on sanitization of areas where crew work, recommend additional sanitation protocols in the event of a confirmed case, and while they offer options to continue operation in the event of a confirmed case, none of the three options on offer indicate an immediate return to capacity shipping. Boat shipping remains the most common method of PC part delivery from manufacturing in China and Taiwan into the United States and much of the world, as with proper preplanning, a launch can be supplied in larger numbers at a lower cost. However, around the holidays, it is not uncommon for manufacturers buy up unused cargo capacity on consumer flights, flying smaller volumes in alongside existing high-volume boat and air cargo operations. However, yet again, COVID strikes here – with far fewer consumer flights and especially far fewer flights into or out of the United States, options to buy up that cargo space are few and far between. Couple this with savvy operators who already plan ahead to supply millions of units for their launches and have the cash on hand to make such moves easily (we’ll talk more about Apple in a minute), and you have Nvidia, AMD, Sony, and Microsoft all fighting for tiny amounts of added capacity on consumer flights, jockeying to get consoles and DIY PC components onto cargo airplanes, and boats scheduled out with regular arrivals, albeit in smaller amounts due to reduced capacities onboard, on journeys that take multiple weeks of planning and travel time to produce results.
The whole situation is a hot mess, frankly, but it meets an exceptional market condition on top of that.
Everyone is Buying New Stuff Because They’re Stuck at Home and Their Old Shit Sucks: This market is a boom for PC manufacturing, with some inside saying the sales volume this year has been like Black Friday on a recurring basis almost daily. With work-from-home office professionals being stuck with their own anemic home PCs (or provided work computers with underequipped monitor and desk setups), learn-from-home requiring a stronger PC than most households give their children, and people looking to console and PC gaming as means to stave off boredom and connect with people, and the market obviously grows in size. Just this year in my own household, we’ve bought my wife a new monitor, new desk, new chair, new sound system for her PC, and I’ve bought new monitors, a new PC (mostly), new headphones, and then my current desktop will be my wife’s once all is said and done, completing yet another “new!” cycle, with her old desktop going to my in-laws, and the myriad of Dell 27″ monitors we’ve replaced being doled out to family and friends.
The market conditions mean that no matter what is released, it sells out almost instantly. Nvidia has launched 4 Ampere graphics card families and all remain in constant sellout. AMD has launched two new lineups and added to the RX 5000 lineup and nearly all remain sold out, with another new line on the way this week! AMD CPUs have a new lineup and remain sold out. Intel’s tippity-top CPUs remain sold out and I won’t hurt their feelings talking about the rest of their stack. This even goes back – RTX 2000 cards sell out and get marked up high by unscrupulous sellers, RX 5000 series cards continue to sell out or remain in low availability, GTX 1600 cards are hit or miss, AMD’s Ryzen 3000 series CPUs are starting to sell out and be marked up for demand spikes as well, and all flavors of console, from the newest and hottest to the value-conscious refreshes of the 2013 launched PS4 and Xbox One are selling out in most places. Want a popular monitor with great reviews for your WFH setup? It’s probably sold out right now, or has been sold out in the last week somewhere. Wanna get some nice headphones for your learn from home kids? Good luck!
However, at least on the PC DIY and console markets, we can speak to a more common factor…
Product Launches for PC DIY and Consoles Always Suck, at Least A Bit: When I got my consoles last, in 2013/2014, I was able to get an Xbox One in late December 2013, the first time I saw one in person since it launched a month prior. I was not able to get a PS4 until February of 2014, and sure, yes, I was out of the country for most of January 2014, but even in stores overseas, I didn’t see consoles for sale. Last year in 2019, when AMD launched Ryzen 3000 and Radeon RX 5700 XT cards the same day, both sold out pretty quickly, and while the Radeon RX 5700 XT wasn’t even expected to be a high-end contender! Availability constraints remained in place for the top of both product stacks for around a month. Intel has had the Core i9-10900k on the market for a while now, and it has remained difficult to find – so much so that Intel launched the i9-10850k to take its place, more or less. Nvidia’s RTX 2080 Ti, the shitty Turing card that was marked up almost double the prior flagship card for only 30% more performance at best, was sold out everywhere for nearly two months in mid-2018 at launch!
Under even ideal circumstances and a seeming lack of demand, launches of these types of hardware nearly always sell out and remain hot items for a while. COVID has amplified this such that even parts that might not normally push such high volume continue to do so – the RTX 3090 sells at mystifyingly high numbers for the cost-to-benefit ratio – but it is still fairly accurate to what a traditional product launch in these markets will do under better conditions, just longer and more painful!
The elongation of availability concerns does exacerbate one last problem, though…
Fucking Scalpers and Bots, Man: Retailers are in a no-win situation with the current market. Ampere and Big Navi GPUs are hot ticket items that have droves of eager customers, as does the AMD Ryzen 5000 lineup and both next-gen consoles. If you announce a drop of stock ahead of time, you run the risk that a scalper can clean up inventory easily, or that a bot can make some purchases. Sure, you limit to 1 per customer, or you do what NewEgg has been doing and make a stock of items available in combo form only, and that sort of helps, but then it also means someone is discouraged from buying. For example, I could have probably purchased a pre-waterblocked Gigabyte Aorus RTX 3080 this week if I was willing to also buy an 850w power supply with it. Two problems – I already have the power supply for my new build and no use for an additional one, but also, the power supply included in the bundle was an awful no-name brand and the markup meant I couldn’t afford the bundle!
So okay, pre-announcing retail drop date and time is a challenge that scalpers and bots can meet, and a scalper making one sale at ridiculous markup can offset being forced to buy a turdbox power supply for an extra $100 to get the GPU you want to sell, so why not do random drops?
Well, to the average user, a random drop is even worse. If you’re on the toilet when a random drop occurs (is that bad phrasing? Probably!) then you might not be able to immediately jump into the app of your preferred retailer to buy. If you work during the hours restocks happen at most often, it probably sucks to not be able to immediately jump over to Amazon and buy the part you need to finish a build. On the other hand, bots can scour and return results quickly, so even if a drop is a short-notice announcement, the bots are probably already there waiting, pounding on the virtual door.
In short, there’s not an ideal solution. Bots are not 100% preventable, and can only be slowed or delayed in most cases, or pushed to a path where an admin has to put them back on the trail, and scalpers have similar resources at hand. Sure, there is a growing use of bots for good (stock trackers that send alerts via Discord, Telegram, or other social media, and bots that place unfulfillable bids on scalper auctions to discourage scalping) but none of these measures guarantees that the end-purchaser is a gamer wanting the part for their own personal reasons. Nvidia gave up on selling their own Founder’s Edition cards directly because of this and has outsourced the sales to retail partners in each region, while AMD almost never restocks their own store with Ryzen 5000 or Radeon RX 6000 cards, instead seemingly pushing almost all volume to retail partners and asking through letters and launch kits for them to implement various anti-scalping and anti-bot measures. And sure, retailers might have a sort of customer service concern with wanting their buyers to be actual end-users, but at the end of the day, a sale is a sale. NewEgg’s combos with RTX 3000 cards isn’t necessarily full-on shitty, but it gets pretty close because the component selection is often pretty meh – get your dream GPU and one 8GB stick of shitty value RAM, or an 850w no-name PSU, or a motherboard that we picked which you might not ever use!
And that’s why, I believe, that while a system like an SMS two-step authentication for a sale might make people like me happy, I can’t even be sure that such a system is bot-proof, and at the same time, that would be a cost to implement for a retailer, who is likely just fine selling as is to whomever may get the cards from them, regardless of why that person is buying.
That leaves my honorable mention!
Apple: Doing Things…Right?: Frequent readers know that I’m not a huge Apple fan or advocate, and short of having had an iPhone in the past, having an iPad Pro currently, and having a 7-year old Macbook Air that I never use, I really don’t much care for the company or their products. However, I have to say, that in the current conditions, Apple really planned ahead for their big launches this fall, and they have made a ton of them.
Apple in the last few months (since the Ampere unveil on 9/1/2020) has released a new iPad, 4 variants of iPhone 12, a new Macbook Air and Macbook Pro, a new Mac Mini, two new Apple watches, two new iPads, and if we stretch back to August, they also launched a new iMac. None of these have been in short supply, and generally, if you want one, you can get one fairly simply at most retail outlets around the world.
Why is that? Well, Apple works to ensure relatively smooth launches through supply chain management. While they fall victim to the same issues as everyone else described here (manufacturing SoCs that need substrates and other hardware to complete assembly, selling complicated devices with multiple components sourced from different manufacturers that must be shipped to a single location for final assembly and then shipped for distribution, relying on TSMC manufacturing capacity for the SoCs powering all of the above devices save for the August 2020 iMac), Apple has a very strict supply chain and is willing to part with money to ensure higher sales volume. What is interesting is that CEO Tim Cook actually has an almost contradictory mantra to this – a famous quote attributed to him is that “inventory is fundamentally evil.”
Now, this actually works in the company’s favor for things like launches. Apple has a lot of data about what to expect in terms of sales volume and builds very well to account for it. They offer pre-order windows that are short but allow the company time to vet purchases to ensure validity, and their ability to supply their products in sufficient quantities helps reduce scalping and bot activity. Even while some of their products have been hit by supply constraints, their lineup has other options to pivot to. If a Macbook Air isn’t available, you can probably get a Pro. If a certain iPad form-factor isn’t in stock, there’s probably one that is. iPhones now come in 4 different size/performance envelopes, so you can pick one that fits your current usage or shift to a smaller or larger device easily enough, and multiple size options have remained available in the 12 lineup.
Apple has gone to great lengths during COVID but also prior to and in the long-term to ensure supply chain readiness. Tim Cook’s background in inventory and logistics has made this his focus as CEO, but before even that, Apple had strict supply chain management in place, works with preferred partners and has built their own sort of supplier ecosystem such that even with companies that do volume orders to multiple customers, Apple is always given the red carpet. At TSMC for example, while AMD has a lot of wafer capacity on 7nm, they only gained additional allocation once Apple moved off the node, and Apple now has a near-total monopoly on the 5nm TSMC process for their Apple Silicon parts for both new Macs and mobile devices. Apple has been willing to pay for consumer cargo space more than many other tech companies and to work diligently to ensure priority for their shipments, often using their financial horsepower to outbid others for shipping capacity.
For Apple, this makes sense because their products are all very high margin per sale, and that margin has only grown as they’ve priced higher, designed and produced more of the components on their own, and built their own direct sales operation through retail stores and online in a way that allows them to retain more of the value of their sale. From the beginning, you could only buy iPhones at Apple stores, and even when they extended to being offered by more carriers and through the carriers directly, Apple often still sold a sizable chunk of iPhones on their own. To the end-user, an iPhone is an iPhone, and whether that $1,000 purchase goes to Apple directly or Apple via AT&T, it doesn’t make a difference, but to Apple, the money saved on not shipping or providing large amounts of inventory to carrier partners adds up from pennies per purchase to millions, if not billions, of dollars at a global scale. Apple was the first company to offer monthly payment and trade-in upgrade options directly to consumers around their carriers, a practice that their biggest competitor in mobile, Samsung, now also does.
So for AMD or Nvidia, does it make sense to micromanage their supply chain to this extent? Eh…maybe not. Sure, I’d love for them to really push hard to get graphics cards and CPUs into every wanting hand, including my own! – but I don’t expect it. AIB partners have been reported to be making some pushes to get more inventory shipped in through what little capacity they can pick up outside of normal avenues, but it is likely to remain anemic until COVID restrictions ease globally, likely not until the holiday period has ended and a vaccine is in clearer focus. For AMD and Nvidia, chip supply will likely open up based on individual factors – for AMD as console sprint manufacturing draws to a close and for Nvidia as Samsung refines their 8nm process further and delivers better yields.
But in the meantime? Expect getting any of the mentioned devices to be a war, and prepare accordingly.