The AMD Execution Thread [2007 - 2017]

[3dilettante]

Imagine the following scenario:
- AMD significantly ups its production of, say, R9 290. (Exchanging wafer capacity for a lower product to this one, because TSMC is pretty much always close to full capacity.)
- Silicon order to cards ready is 3 to 4 months.
- Meanwhile the coin craze dies.

This also assumes that the GPUs are the limiting component. Not every card shortage has been attributed to that, and a sudden spike in large memory video cards and those high-speed power controllers might hit someone else's supply constraint.

It may not require that the mining craze die down, rather if scrypt ASICs can become popular enough to make GPUs less favorable in terms of rig cost or power consumption, or in the worst case become dominant enough to force the mining difficulty too high for multi-kilowatt GPU rigs to pay for their electricity bill (possibly already the case in countries with higher power costs).

It seems like this starts for Litecoin around mid-2014, although the ramp may take some time.
Even without that impact, the GPU-amenable coins with some fixed total number of coins or slowly expanding monetary base still become more difficult over time, and return on investment is already much lower than those who got in when it was easy.

Miners with GPU rigs are already discussing which $coin they can hop to next, gambling that they can convert the seigniorage of the early part of the difficulty curve to a somewhat acceptable conversion rate to slightly less recent $coins all the way to Bitcoin, or directly.
It's a pyramid of illiquidity that really doesn't show a long future for them, and for all we know the next struggling generation of cryptocurrencies will be one of those being specifically tailored to gut both ASICs and GPUs.


I do fear that AMD may be facing a glut of used Tahiti and Hawaii cards late this year or the next that could depress the perf/$ ratio and hurt any new products coming out, unless AMD's next round of products is a much greater jump over its forebears than has been the case for quite some time.

The assumption that you are making is that AMD either didn't know about, understand, or care about mining and therefore didn't set initial or subsequent runs high enough. I continue to believe there is a fly in the ointment somewhere in the production and we just don't know yet. Until the 2014 WSA is released and we get more info on what is made and where...

I'm not sure when scrypt-based coins like Litecoin or Dogecoin were able to reach a level of acceptance, and even if AMD made a bet on a cryptocurrency, it could have bet on any of the many others that haven't.
The standard bearer Bitcoin is completely out of reach for AMD cards, so some of their data would have been indicating that it was too late anyway.

 

[silent_guy]

The assumption that you are making is that AMD either didn't know about, understand, or care about mining and therefore didn't set initial or subsequent runs high enough.

Of the three possibilities, my money is squarely on 'didn't care', because of the potential risk. That said, I was personally completely unaware of LiteCoin until just a couple of months ago.

This also assumes that the GPUs are the limiting component. Not every card shortage has been attributed to that, and a sudden spike in large memory video cards and those high-speed power controllers might hit someone else's supply constraint.

Yes. But I think that's not an unreasonable assumption. GPUs is the component that has the highest risk of becoming obsolete quickly, and that's the most difficult to manage in terms of SKU allocation. Memory and power controllers can be use interchangeably between different GPU (and other) products. But who knows...

 

[Dave Baumann]

Of the three possibilities, my money is squarely on 'didn't care', because of the potential risk. That said, I was personally completely unaware of LiteCoin until just a couple of months ago.

It started taking off in November; the Litecoin difficulty chart shows the difficulty significantly ramping in mid-November, which indicates when it really started getting popularized by GPU miners. If you were to have made any decisions based on this you are pretty much still in wafer out lead time, let alone channel availability lead time.

 

[silent_guy]

Wow, that's one steep increase in difficulty!

So the answer was probably "wasn't aware". Obvious follow question: what next?

 

[overclocked_enthusiasm]

Wow, that's one steep increase in difficulty!

So the answer was probably "wasn't aware". Obvious follow question: what next?

Here are the variables:

~ Maxwell script mining performance (if better than AMD will ease supply issues)
~ Maxwell launch/availability dates (if Q1-Q2 may push Hawaii refresh plans up depending on performance)
~ Maxwell performance vs R7/R9 (if not substantially better may push Hawaii refresh plans later)
~ AMD loyal gaming community frustrated about availability/pricing
~ Gamers switching to Nvidia due to performance per dollar (280x vs GTX 780 etc)
~ The trend of script mining with AMD hardware for the next 3-6 months

If I were AMD I would look to regain the lost gamer mind/market share in Q1-Q3 with an availability push on ALL R7/R9 skus. This would also allow many of the current R7/R9 card owners a mini-upgrade to a Xfire solution (which they can't afford right now due to pricing) before 2015.

This would run the risk of building inventory BUT the lost market/mind share with the core gaming market (especially with the recent emphasis placed on Mantle) is waaaay too important to cede to Nvidia for another day. The current R7/R9 skus should have plenty of performance legs (particularly with Mantle) to warrant MUCH heavier production runs to clear the supply constraint issue. At the launch date MSRPs the R7/R9 parts really do offer a compelling reason to buy AMD vs Nvidia.

This assumes there isn't a supply chain or fab issue that I have hypothesized to before. I still think there is something amiss in waferland...

 

[Andrew Lauritzen]

I still vote for remove bit rotate instructions in consumer cards

 

[3dilettante]

This assumes there isn't a supply chain or fab issue that I have hypothesized to before. I still think there is something amiss in waferland...

Rapid turnaround isn't what the manufacturing process permits, nor is it how the board partners, shipping, and retail contracts are structured.
Even if AMD managed to get a big rush started, that's a quarter or more out and other parts of the chain may not have contracted or planned on being able to ramp on AMD's whim.
Fixing the market share problem also assumes the board partners are willing to take on some potential risks of being caught with a downturn in mining, and the retailers in affected regions will fight to keep the prices as high as they can per unit, with some a potentially perverse incentive to constrain supply rather than have a surfeit of regularly-priced GPUs not flying off the shelves.

Everybody else using TSMC seems to be doing okay.

I still vote for remove bit rotate instructions in consumer cards

Driver-level detection of scrypt code that ramps the power consumption to TDP and/or imposes a clock ceiling would do quite a bit.
The activity seems to be very discernable.

 

[Dave Baumann]

Driver-level detection of scrypt code that ramps the power consumption to TDP and/or imposes a clock ceiling would do quite a bit.
The activity seems to be very discernable.

As far as I'm aware the element that makes us better isn't even natively exposed in OpenCL in the first place, I believe the application is using a backdoor / IL code to get access. But imposing any driver restrictions now would be futile for current GPU's as older drivers will just be used - the primary application that is used for GPU mining actually stopped development of the GPU path some time ago, people are just using older stable releases.

 

[overclocked_enthusiasm]

Even if AMD managed to get a big rush started, that's a quarter or more out and other parts of the chain may not have contracted or planned on being able to ramp on AMD's whim.

Everybody else using .

I have to assume they began addressing this as early as December when the availability issues surfaced. They must have channel checks that would have raised red flags very early on. That would mean some easing starting in March based upon the timelines put out here of 3-4 months IF AMD recognized and properly addressed the problem. I can't see how AMD would allow the market to get away from them for now 3-4 months and counting without responding in some way. They couldn't possibly risk a 6-12 month runaway market with something as big as Mantle and Kaveri on the field...it warrants decisive action to alleviate the supply constraint.

Wouldn't any crippling of AMD cards have to happen at the hardware level as opposed to software where drivers and BIOS can be rolled back? What would that even look like?

What are the odds GF has started to manufacture any GPUs? I think it is a given at least some (if not most/all) of the semi-custom APUs have joined Kaveri at GF to help satisfy the WSA and based on John Byrne's statements in December. With common IP in APUs, semi-custom APUs and GPUs can AMD leverage that commonality to drive costs down and yields up by using a single fab? This part of the industry I don't understand as well.

 

[3dilettante]

As far as I'm aware the element that makes us better isn't even natively exposed in OpenCL in the first place, I believe the application is using a backdoor / IL code to get access. But imposing any driver restrictions now would be futile for current GPU's as older drivers will just be used - the primary application that is used for GPU mining actually stopped development of the GPU path some time ago, people are just using older stable releases.

This was more of an idea for the future, as an alternative to removing the instructions.

 

[overclocked_enthusiasm]

http://www.digitaltrends.com/comput...e-hikes-due-chinese-new-year-litecoin-miners/

"When we spoke with AMD, they told us that the Chinese New Year wreaked havoc on supplies of AMD graphics cards, like the Radeon R9 290X. This isn’t like Memorial Day Weekend where the workers got a Monday off and that was it: AMD informed us that this was a two week long celebration, so those facilities that make AMD cards were essentially shut down for two weeks because of the holiday. According to AMD, this two week long period began on January 31, the day of the Chinese New Year."

How is this not known ahead of time and accounted for?

 

[silent_guy]

I think it's important to not reduce a complex global supply chain into a sound bite.

Everyone knows that Chinese New Year brings the everything in that region to a standstill. But what do you expect companies to do about it? Set up an additional temporary production line just a couple of weeks to compensate? That would be very cost effective...

 

[3dilettante]

I have to assume they began addressing this as early as December when the availability issues surfaced.

You can assume that.
They may well have not had clear visibility on where the cards were going, or they took what measures they had that weren't blindly reactionary.

They must have channel checks that would have raised red flags very early on. That would mean some easing starting in March based upon the timelines put out here of 3-4 months IF AMD recognized and properly addressed the problem.

This is going by your definition of a proper addressing of the problem.

AMD is not that nimble, and in this instance that may not be a bad thing. Trying to torque a mass supply chain to chase a fluke in the market founded on a get-rich-quick bubble has few upsides.
The initial extent of the miners' appetite was not clear, and it's not clear now.
We have no idea just how much AMD would need to flood the market with to get products spiking to $900 back down to MSRP. Whatever that amount is, AMD would be aware that the majority of its product would be in the hands of a very non-loyal customer segment whose behavior is out of the norm for the typical gamer in terms of consumption and upgrade cycles.
Whatever they have will be back on eBay in time to wreck the prospects for the next generation of cards.

Or they could wait a little while longer to see if the reintroduction of bit shifts in Nvidia's Maxwell makes GCN less compelling.

Wouldn't any crippling of AMD cards have to happen at the hardware level as opposed to software where drivers and BIOS can be rolled back? What would that even look like?

Existing models are already in the wild. There is no prudent way to put that genie back in the bottle.

What are the odds GF has started to manufacture any GPUs?

Manufacturing the GPUs you assert must flood the market? I'm skeptical.
Price hikes would be needed to make up for the extra cost of porting the exact same product to another line.

With common IP in APUs, semi-custom APUs and GPUs can AMD leverage that commonality to drive costs down and yields up by using a single fab?

Costs are likely wrapped up in the vagaries of WSA negotiation.
Putting a bunch of random and disparate products in the same fab doesn't give that much to improve yields if the process is already mature. Some of the choices that might improve things for one would have adverse effects on others.

http://www.digitaltrends.com/comput...e-hikes-due-chinese-new-year-litecoin-miners/
...
How is this not known ahead of time and accounted for?

It's called living with it.
The alternative is to establish manufacturing in a place whose cost structure is substantially worse 365 days of the year than one that takes two weeks off.

 

[cal_guy]

RBEs and memory bandwidth(but not latency) aren't important to mining so it might be prudent for AMD to look at their bins to see if any meet this criteria.

 

[Alexko]

Is it just me, or has AMD been uncharacteristically quiet lately? They haven't said anything about their GPU plans, they haven't said anything about Excavator since its original announcement years ago, they haven't said anything about mobile Kaveris or even future desktop SKUs, they haven't even acknowledged the existence of Carrizo.

They showed Mullins/Beema and touted some figures but they've been utterly quiet about that since, and it's unclear when actual devices can be expected.

I guess it's better than promising things they eventually fail to deliver, but it feels strange.

 

[Wynix]

Is it just me, or has AMD been uncharacteristically quiet lately? They haven't said anything about their GPU plans, they haven't said anything about Excavator since its original announcement years ago, they haven't said anything about mobile Kaveris or even future desktop SKUs, they haven't even acknowledged the existence of Carrizo.

They showed Mullins/Beema and touted some figures but they've been utterly quiet about that since, and it's unclear when actual devices can be expected.

I guess it's better than promising things they eventually fail to deliver, but it feels strange.

Hawaii launched only a few months ago and kaveri has been out less than 2 months, i do not expect them to start giving out solid information until 2H/2014.

Also with Carrizo almost certainly on 28nm they have no reason to broadcast the information several months before release.

 

[overclocked_enthusiasm]

IEveryone knows that Chinese New Year brings the everything in that region to a standstill. But what do you expect companies to do about it? Set up an additional temporary production line just a couple of weeks to compensate? That would be very cost effective...

No. I expect any global company who's primary supply channel comes to a predictable screeching halt for 2+ weeks to plan accordingly as any good procurement manager would do by having extra inventory to compensate for this supply disruption. Why would ANY publicly traded company sacrifice 1/6 of their quarterly revenue on a known event? The Chinese New year isn't exactly an unexpected event like adverse weather or other unexpected supply channel disruptions. I own a restaurant, and if my primary vendor went on vacation every year for 2+ weeks I can assure you I would order heavy in the weeks preceding to avoid any disruption for my customers. GPUs aren't exactly perishable...unlike the food & beverage I purchase.

For AMD to blame the Chinese New Year or any other expected and predictable event just seems amateurish or disingenuous at best. AMD continues to underwhelm me on their launches and execution. I understand the inventory pressure they are under and the INITIALLY unexpected script mining demand...but this has been known since very early December. AMD can and MUST do better. Do you see Apple, Nvidia, HP, Dell, Lenovo, Samsung, or anyone else blaming things on the Chinese New year?

 

[3dilettante]

No. I expect any global company who's primary supply channel comes to a predictable screeching halt for 2+ weeks to plan accordingly as any good procurement manager would do by having extra inventory to compensate for this supply disruption. Why would ANY publicly traded company sacrifice 1/6 of their quarterly revenue on a known event?

Per the Litecoin difficulty graph Dave linked earlier, the major rise in GPU mining started mid-November.
We've discussed the 3-4 months it might take to get things through the chain in the absence of potential confounding factors such as contracted schedules and the business needs of everybody not AMD.

So, with the Chinese New Year kicking off on January 31, let's do some time travel to our best-case day 0 in mid-November.
End of January to beginning is one month. So our 3-4 months is 2-3.
Beginning of January to beginning of December is one month, so 1-2.
December to middle of November is--wait don't just whiz on by mid-November!!!!!!!!!
Beginning of November, or minus .5-1.5.

Why didn't AMD do more to handle a problem with negative 3-8 weeks of warning?
Why didn't it increase production to match a secondary demand pool it had no evidence for and that still has no good depth measured?
Why didn't it just get a bazillion cards like the Hawaii chips? I mean, it's not like the beginning of November or October was significant for Hawaii in any other way, right?

The Chinese New year isn't exactly an unexpected event like adverse weather or other unexpected supply channel disruptions. I own a restaurant, and if my primary vendor went on vacation every year for 2+ weeks I can assure you I would order heavy in the weeks preceding to avoid any disruption for my customers.

The article discussed two reasons for the shortage, mind you. They would be operating in concert.
What if they did match their orders to compensate for the known event they had, per the Chinese Calendar, some centuries of warning for, but irresponsibly ignored the warning they wouldn't get for another month?

GPUs aren't exactly perishable...unlike the food & beverage I purchase.

Are you on the hook for multimillion dollar charges if those get written down?
(edit: What's your inventory cost when you need to store some thousands of cans of peaches?)

For AMD to blame the Chinese New Year or any other expected and predictable event just seems amateurish or disingenuous at best. AMD continues to underwhelm me on their launches and execution. I understand the inventory pressure they are under and the INITIALLY unexpected script mining demand...but this has been known since very early December. AMD can and MUST do better. Do you see Apple, Nvidia, HP, Dell, Lenovo, Samsung, or anyone else blaming things on the Chinese New year?

Is anyone using Samsung TVs for a Litecoin mining spike that occurs within the needed manufacturing lead time?
AMD has plenty to improve and be criticized for, but failing to do something we have no evidence it could do, and which would in the absence of clairvoyance appear to be literal insanity is not one of them.

This class of GPU is a product niche of a certain typical volume, and it has been impacted in an unintended and deeply ironic way, where AMD is accidentally massively competitive relative to the competition--a position I (and AMD, I think) cannot remember it ever being in.

If doing something that incurs millions of dollars of risk and makes no sense for weeks into the future is what you require to justify going long on AMD stock, I suggest not doing so.

 

[silent_guy]

If you can blame AMD of one thing, it's that someone had the bright idea to lay the blame on a completely predictable event. That was just dumb.

 

[overclocked_enthusiasm]

Are you on the hook for multimillion dollar charges if those get written down?
(edit: What's your inventory cost when you need to store some thousands of cans of peaches?)

I would risk ordering too much in order to protect my business...my core base...without a moments hesitation. My "inventory cost" would be a temporary reduction in cash on hand but would not negatively impact my margins as I can likely sell them later for my target asp. Guess what happens when I run out of single vineyard Burgundy, high elevation Napa cab, single barrel bourbon, craft beer etc....my customers get pissed and go down the street to my competition. These are high margin high asp items for me, just like 270x and above dGPUs for AMD, and they bring in the "right" kind of customer who will also buy other items...no different than maybe a GPU buyer deciding to go with an AMD CPU or APU as well while they are buying. That halo effect is very important for any business.

We are seeing this availability/pricing frustration in forum after forum...and in my own recent experience building a gaming rig...people who want to buy AMD are being driven to Nvidia either by high costs or lack of availability. I forgive AMD the initial surprise of Litecoin but if there isn't a meaningful response in the pipeline right now to be rolled out in March I think they are in bigggg trouble going forward. They risk losing initial builds...repeat buyers looking to Xfire...and the resultant backlash in market share, mind share, and their core AMD gaming base. How can Mantle gain traction without a large installed base of AMD GPUs? The risk to their core gaming base (which is fragile) and their potentially disruptive API (which is even more fragile) cannot be overstated...especially during this critical time in their restructuring (which is most fragile of all).

Let's take the mid-point of your 3-4 month production timeline for a supplemental order or increased second order of R7/R9 chips and first week of December as the likely date of "Houston...we have a problem". That would mean that by the middle of March, or in 2-3 weeks, we should see a meaningful channel response by AMD and their AIB partners to alleviate the supply issue. It's not like they have a 20nm refresh waiting to go and another big batch of 28nm parts are going to go stale on them or push any refresh back is it? I would argue that any large batch of chips received in March/April will sell through just fine as Maxwell isn't even out yet, we don't know that Maxwell is substantially better than R7/R9, and R7/R9 beat current Nvidia parts in most reviews across the stack.

The opinion that AMD was unfairly surprised, can't do anything about it, shouldn't risk doing anything about it, and just let things resolve themselves on their own is deeply flawed.