AMD and GlobalFoundries Update: Orders Through 2024, Now Non-Exclusive | Hacker News Reader

AMD and GlobalFoundries Update: Orders Through 2024, Now Non-Exclusive

anandtech.com

153 pointsrbanffy67 comments

AMD could use that capacity to re-lauch older GPU at a lower price and make gamers happy. Low cost GPUs (even with higher watts) might be marketable again.

How old would those GPUs be? It might be great for those of us who just needs a reasonable GPU for older games and general desktop use.

The 75W GPU market is completely stalled at Rx 560 and NVidia 1050 Ti IIRC. People still buy those because:

1. 75W is the maximum the PCIe x16 slot can power without any additional power (see 6-pin or 8-pin connectors).

2. 75W is too little to really make a "decent" GPU for modern AAA games. But still a huge step up from integrated graphics. Playing 5-year-old or 10-year-old games on lower settings is the only real possibility.

Better 75W GPUs are possible, but they would be quite expensive on a perf per dollar basis.

An RTX 3070 Laptop GPU on a PCIe card would be quite nice, but the market for that isn't big.

Laptop 3070 would be sold out immediately. It's too good at mining and sometimes better than desktop part

AMD's last 12nm GPUs were the Radeon 7, Vega 56 and Vega 64, which were all pretty much flops. The rx580, their preceding model did quite well for a few years as a last gen budget model.

To my understanding, the radeon 7 and vega were expensive chips to produce. All of them used HBM, which contributed to this, and it's not clear how easy it would be to make a non-HBM variant. Would customers accept them if they needed the same MSRP as the 6xxx GPUs?

So if they were to relaunch and old GPU I think the rx 580/590 at $200 would do. In this climate, $200 for a 1080p medium GPU would probably sell.

Radeon VII was 7nm.

Re-releasing the Vega cards would be excellent for ROCm programmers. AMD needs to do something to get beginner programmers onto their system cheaply.

Radeon VII / Radeon VII Pro / MI25 seem like the cheapest option for cards still in production. Vega 56 / 64 still work if you can Ebay a used-card.

Radeon VII has comparable memory bandwidth to an RTX 3090 (or V100) and 7x the double precision performance of RTX 3090 for a quarter the cost. It has almost double the memory bandwidth of RX 6800 and 3x the double precision performance at half the price. It would be a compelling card today if it hadn't been EOL'd almost two years ago.

Radeon VII isn't really EOL'd. Its now called Radeon VII Pro and worth $2000. Or (EDIT: MI50: the MI25 is Vega64 IIRC).

As long as AMD is selling those cards out at the higher price, there's no reason for the $700 price point to come back...

It was a stopgap product that wasn't really profitable to reduce the gaming performance gap between NV and AMD. Its marketing wasn't that much targeted towards compute either.

And as soon as they had cards with better gaming perf out, it outlived its usefulness in the AMD product matrix.

Nvidia has the Titan V as a prosumer card with HBM2, 1/2 FP64 and all for $3k, but that seems to be out of stock nowadays too...

Agreed that it was EOL'd because it was priced too low, though the Pro also lifted the double precision throttle (3.5 TF to 6.5 TF) and added Infinity Fabric.

AMD APU or Intel Xe for older games?

A 7nm design uses less area, and is probably just going to get cheaper over time. A smaller, 7nm RDNA2 design might be easier to make than backporting RDNA2 to 14nm.

I think an older GPU might still be useful, but from a cost/benefit perspective, it seems to make more sense to just make a smaller chip on the latest process (especially if your latest technology, like RDNA2, already has lots of data on the 7nm node).

7nm capacity is likely to be tapped out for the new couple years. The OP's suggestion (I believe) was literally just to keep making older 12nm cards (no backporting) to supplement current supply.

Normally this would... not really work well. But in the current environment, it'd probably work just fine (just look at how much GT 1030s go on sale for...)

I think an older GPU would definitely be useful. Used GTX 1080 cards (I picked one up used for C$375 back in September) are listed for C$750-850 where I live. I'm not as familiar with AMD's products but I think 1-2 generation old GPUs would definitely have high demand until current gen ones are more available.

MooresLawIsDead suggested this as well and I think it's a good idea especially in the current climate, though it's hard to react situationally when it takes >12 months to launch a product even an iterative one. Take the 5600XT design and backport it to 12nm. You get something that performs pretty well at 1080p that you can sell for really cheap for a long time. Finally a replacement for the RX480/RX580.

One of the strange things with AMDs new GPU push is that their GPUs are less profitable in a supply constrained environment than their CPU/APUs. Ian Cutress recently talked about this (I think on YouTube @TechTechPotato). Even their highest end GPU dies selling to consumers at a price point higher then they planned for is much less profitable per wafer than a midrange APU. If there were no supply constraint they could just make as many of both as they can sell but right now they're just trading off profits against GPU market growth.

Can you “just” take the 5600XT and switch it to 12nm? Won’t it require more power and cooling?

Sadly no, because everything is closely tied to the node. For example, it might take 0.3ns for a signal to physically travel from one edge of the chip to the other on a 7nm chip. On 12nm, it might take 0.5ns. The chip was designed to compensate for a 0.3ns delay, not 0.5, so now the cache is sending data to the core at the wrong time, leading to corruption and garbage data being processed. This, plus a million other things in both manufacturing and design have to change to build a chip on a different node (manufacturing tolerances, clock speed, metallurgy, patterning issues, etc.).

You can't "just" do it, but it is possible as shown by Intel Rocket Lake.

I read great writeup https://news.ycombinator.com/item?id=25926211

By the time you actually get those products on the shelves, Ethereum will either be on Proof of Stake, in an ice age, or less valuable than now.

And thus the GPU shortage will end.

In 1-2 years we might see 5nm GPUs, and so the gulf between a 14nm GPU and the latest GPU (even if you can't find them on shelves) will be even greater.

Its more realistic for GlobalFoundries to consider opening a new factory with 10nm/7nm, especially if they can pickup established knowledge and hardware from other players.

I doubt it's so simple, there are many other profitable coins to mine.

Computer gaming moves in console generations.

The PS5/new XBox are targeting a ~5 year generation. This means that games will be mostly playable as long as you get a top of the line 14nm GPU for 5 more years.

Remember that graphics have stopped being a bottle neck for games for a few years now. The top played games: Fortnite, Apex, LoL, Dota, Fifa and the like are perfectly playable on mediocre graphics.

Nethack also runs pretty well on old GPUs. I consistently get >60fps even on integrated graphics.

Are you sure your terminal emulator is rendering that quickly? d= e.g. iTerm2 caps at 30fps unless you change the default.

The more important question is does it register key release events

Lol. Nethack is designed for the Unix console. All commands are key-down only.

But you had me going for a second. Nethack is also turn-based, so there's no real-time element at all. If you come across a difficult turn, its a common "strategy" to walk away from the computer, relax a bit, and then look at the screen with a different mindset. Maybe you can think of a better solution if you give yourself time.

Big deal. I was getting that on a VT-220 terminal connected with RS-232.

The VT-220 appears from quick Googling to max out at 19200 baud, or < 2 full screen (80x24) "draws" per second. ;)

(Although I assume nethack is more clever than just "re-send the whole screen for every change".)

Hey, but the frames are redrawn at AC line frequency.

No one said anything about the data changing ...

(and yeah, I believe it only sent data for the delta.)

PS5/new Xbox are 7nm though. 14nm was the PS4 shrink/Scorpio.

The top played games have pretty much always been playable on mediocre hardware. Otherwise they wouldn't be top played.

Proof of stake Ethereum is very similar to 5G, yes there is real technology, yes it's 'working' in some capacity now. But you really shouldn't plan on it reaching you any time soon.

The ETH2 'beacon chain' is a meta-chain that's supposed to checkpoint a bunch of 'sub chains' for scalability. They have launched this beacon chain. But it doesn't do anything other than mint ETH2 tokens and operate their proof of stake consensus system.

AKA it's another blockchain, different than ETH1 and it doesn't even move transactions yet. It's got a novel POS system to it's name, that's not nothing, but that is all it has at the moment.

What does a 'sub chain' actually look like to interact with? How do you coordinate with many of them? All these questions have answers in theory, but not answers in solid production ready interoperable code.

'Proof of stake Ethereum' only happens if these questions are figured out and then ETH1 is somehow brought under the beacon chain's governance. Which is a whole different kettle of fish than a greenfield subchain. Exactly how this is to happen... well I haven't even seen anything credible on this. As far as I can tell it's not even seriously being worked on yet.

So we may, if things go well, have another blockchain calling itself ETH2 that is proof of stake by the end of this year. But that won't cause ETH1 to stop wasting electricity or gracefully move it's economic activity to a proof of stake system.

Moving billions of dollars of economic activity out of the hands of miners (who have every incentive to screw it up) and into the hands of a proof of stake validator set while at the same time not significantly disrupting said system or opening up new vulnerabilities is not trivial.

tl;dr Proof of stake Ethereum is possible, but there definitely isn't a rush order on it. I would bet on Ethereum miners still being around 2-3 years from now.

All the more reason for crypto environment regulations. We can't keep destroying the environment just because "miners" want to make a profit.

I have a long laundry list of things that are destroying the envuronment just to make a profit.

OK, so add it to your list then.

This is not exact. The merge is definitely being prioritized and the current target is Q4’21 / Q1’22. Developers have chosen to perform an early merge by using the ETH2 consensus layer to validate ETH1 blocks, it’s a pretty straightforward process were both ETH1 and ETH2 clients co-exists with minimal changes. There is already a running testnet. More info https://blog.ethereum.org/2021/01/20/the-state-of-eth2-janua...

This is new since the last time I took a deep dive into things.

This does seem more straightforward, at the cost of essentially all of ETH2's advertised scalability and other features.

It still does not eliminate the challenge of actually performing such a delicate transition on a live economy. Which will be significant on a social front in addition to technical.

Not saying I don't believe these things can be done, or don't believe that they are good idea.

At this point I really wouldn't dare making any projection on the future of cryptocurrencies. You can't rationalize irrationality. Let me remind you that currently Doge coin, a literal joke fork of LiteCoin that traded for a fraction of a cent mere months ago, is the fifth biggest cryptocurrency by market cap and is being pumped by Elon Musk for some reason.

I have literally no idea what that space is going to look like one week from now.

I see it as no worse than gamestop or synthetics CDO's and some other stuff going on Wallstreet.

Same reason as any pump.

My 2014 Radeon HD 6850 was recently sold for a nice profit.

I don't think everyone in the world goes for only the latest and greatest. Or even have the budget to get a card more expensive than $200 USD.

I am a member of a simracing league with dozens of people in South America, and only 1/4 of them have machines capable of running games launched in the last three years.

Only one guy in the group has a GeForce RTX 3080. One, out of more than a hundred.

Everyone else in the group will definitely welcome an affordable mid-range card. Including me.

Why would AMD want to make less money?

How would selling even more GPUs make them even less money?

Right now their problem is that they literally cannot make RX 6000 series cards fast enough.Adding more GPU capacity in the form of reissued 14nm chips to fill some of the gap in demand would not result in fewer RX 6000 cards sold. People just want something, anything they can buy today to hold them over until they can get their hands on something good.

There are basically no cards available on the consumer market. Nobody is making money in this space.

I was looking this week for a replacement for the GPU I bought in 2014 for £80 (r7 260x). The only sub-£600 card available on any site when I checked was the GeForce GT 1030, which would give me slightly worse performance for £90. The situation is unbelievably dire.

Rather, everyone is making money, but the OEMs could make more if they'd be willing to raise MSRP. It's a shortage of supply, not a stoppage.

> and AMD in turn is required to pay for these wafers, whether they use this capacity or not. > AMD expects to buy approximately $1.6 billion [≈ box office sales of Bambi, 1942] in wafers from GlobalFoundries in the 2022 to 2024 period.

They'll pay for it, anyway, $1.6bi. I'm pretty sure that there's a way to spend $1.6bi to make GPUs and sell it in the long-tail consumer spectrum without bitting your sales. And create the desire, on those users, to upgrade to a better GPU in the future.

Or don't even make money about it, just ensure that your competitors will have to lose money too.

GPU's significantly increase performance between generations. My 1060 is already effectively outdated. The cost of production isn't less than it was originally, and $200 for a 4 year old GPU is far too much. Modern titles are going to require 6-8GB vram at a minimum.

Used 1080Tis are going for 600€+, so new ones for 400 would probably sell like hot cakes.

you would be right, if only the market is normal, or only slightly inflated.

Go to any used goods site, and you would realised $200 for 1060 is actually below the used market price.

Yes it would sell like hot cake

Engineering resourses are limited. It takes significant effort to transfer and verify a design for a new process. Why pull off engineers working on the next gen products to retapeout an old product on an old node? I doubt it would have ROI.

Note: "Now with that said, the net impact of this change is likely to be limited as AMD was already free to pursue other fabs for 7nm and smaller nodes – which will be the vast majority of AMD’s needs over the next three years."

What node is AMD IO die now a days? It always felt like they were keeping the IO dia one node backward to honor the GF agreement. Even with this they are able to give 128 PCIe lanes and Better Memory frequency. So when they are able to bring IO die also to leading node, it is going to be awesome.

The AMD IO die is using the GloFo 14nm process.

Not as awesome as you think. The IOD is on the older process because it benefits comparatively little from transistor shrinks. (Logic-level IO transistors have to be physically large, with thicker gate oxides.) But it will help some.

It might not help at all from a density perspective. The IO die has a lot of pads along the edges that can't be shrunk which puts a lower limit on die area. I wouldn't be surprised if the logic areas in the middle already have a lot of blank space at 14nm. You used to see the same thing with north bridge chips, and that's basically what the IO die is anyway.

> The IO die has a lot of pads along the edges that can't be shrunk which puts a lower limit on die area.

Are you just imagining things? This article includes an image of the 14nm die, along with the functional breakdown further down the article:

https://www.tomshardware.com/news/amd-64-core-epyc-cpu-die-d...

I'm not. Your citation shows extremely dense pads in two banks (even before taking power delivery into account which isn't shown in your functional breakdown), and decent amounts of empty space between the logic blocks.

Density-wise, no, probably not. But you'll get faster SERDES and should be able to improve the internal routers/crossbars, which contain more logic than you'd think. But, yeah, there isn't a lot there.

When the higher cost per area is considered (just count the number of bloody masks...) it's probably a net loss.

But Milan's IOD consumers a lot of power even if it's idle. Next EPYC will come with DDR5 and PCIe 5 so it getting much worse.

https://www.anandtech.com/show/16529/amd-epyc-milan-review/5

What fab(s) does AMD use instead? I skimmed the article but didn't notice anything about this.

Mostly TSMC, https://www.tomshardware.com/news/amd-tsmc-second-largest-cu...

TSMC. AMD and Apple is TSMC's core customers

Also NVDA, and by proxy Sony and Microsoft as they use AMD cpus/apus.

I think Sony and Microsoft are AMD's customers, not TSMC's. They buy their chips from AMD. I don't think they care much where they are manufactured.

TSMC and Global Foundries are the one's AMD uses right now .

Perhaps this will help auto-manufacturers in the U.S. by freeing up some supply that was previously committed to AMD?

Reply on news.ycombinator.com