AMD crushes Q4 earnings estimates

[moinmoin]

FYI there will never be a 32 core threadripper without a socket redesign

der8auer thinks there is a chance (his conclusion at 18:19).

 

[Gideon]

I wouldn't be surprised at all if they'll release a 24 core or even 32 core Threadripper with the 2xxx series. Otherwise just up to 10% faster clocks doesn't seem all that exciting.

 

[Topweasel]

der8auer thinks there is a chance (his conclusion at 18:19).

My quote was from someone posting that same video.

That because he only saw the easy stuff. His CPU had four actual dies. That doesn't make TR4 able to be a four die platform. There are a dozen reasons that he could still be wrong about TR having 4 dies. Then there is a dozen reasons TR4 cant support four dies. The simplest is that TR4 specs for motherboard manufacturers won't have all the needed traces to support 4 dies. It would have made x399 boards even more expensive.

 

[french toast]

Margins are fine. It's the volume they need to worry about.

No margins are not fine.
I accept blog is skewed somewhat with console APUs which are high volume very low margins.
But AMD would really need to be hitting 45-50% margins to be considered stable and healthy imo.
Increase in volume on top of that would be welcome.

 

[realibrad]

No margins are not fine.
I accept blog is skewed somewhat with console APUs which are high volume very low margins.
But AMD would really need to be hitting 45-50% margins to be considered stable and healthy imo.
Increase in volume on top of that would be welcome.

I think if they can get there not including console then they are very good. Console is just extra revenue.

 

[CatMerc]

Margins are terrible. They need to increase these to become more profitable.

As long as they can fund their operations and keep up thanks to high enough volume even at relatively low margins, it doesn't matter for anyone but investors what the margins are. It's a stupid thing to chase. AMD could just stop doing console orders and suddenly their margins would jump to 40% or higher, but they'd be hemerrhoging money.

 

[scannall]

No margins are not fine.
I accept blog is skewed somewhat with console APUs which are high volume very low margins.
But AMD would really need to be hitting 45-50% margins to be considered stable and healthy imo.
Increase in volume on top of that would be welcome.

You don't go from 0% or worse margins to 60% overnight. Well, except in Hollywood movies. This is a good showing by AMD. And I'd expect first quarter next year will be 45ish%. It's a long steep hill, in an entrenched industry. This is about as close to a miracle as you can expect.

 

[CatMerc]

You don't go from 0% or worse margins to 60% overnight. Well, except in Hollywood movies. This is a good showing by AMD. And I'd expect first quarter next year will be 45ish%. It's a long steep hill, in an entrenched industry. This is about as close to a miracle as you can expect.

Your expectations are more optimistic than AMD's.

https://www.ultragamerz.com/wp-content/uploads/2017/12/amd-q3.png

 

[Kenmitch]

After a decade of AMD doom and gloom I'd imagine most end users are happy they're competitive on at least the cpu side of things for now. Arguing over the margins during the early years of the rebirth of the company seems silly at best. The biggest issue is the sheep herder philosophy of the internet fanboys which seem to steer potential sales to their company/brand of choice. A person on a 60Hz monitor most likely could care less if brand X gets 110fps vs brand Y getting 98fps in their favorite games.

 

[CatMerc]

Here's an interesting thing to note about these financial results. Their Q1 2018 guidance is actually an incredibly good sign.

Under the new accounting standard, revenue was 1.34bil for Q4 2017, while Q1 guidance under the same standard is 1.55bil. Q1 is usually the weakest quarter, while Q4 is either second or strongest quarter, meaning AMD's weakest is now up there with the stronger quarters of 2017.

The seasonality applies to the entire industry for the most part, it's not just consoles. It's quite unusual for NVIDIA or Intel to beat Q4 with the following Q1.

NV Q4:16 - 2.17bil
NV Q1:17 - 1.93bil
INT Q4:16 - 16.4bil
INT Q1:17 - 14.8bil
INT Q4:17 - 17.1bil
INT Q1:18 Guidance - 15bil

Now, not only is AMD not going to see a drop according to their guidance, but they're expecting 15% growth sequentially. This is a huge sign for AMD's trajectory in 2018 in my opinion. I expect by Q3 2018 we might see AMD's first 2bil revenue quarter since a damn long time, maybe even Q2 if quarter to quarter growth in 2018 follows a similar trajectory to 2017.

 

[Yotsugi]

Your expectations are more optimistic than AMD's.

https://www.ultragamerz.com/wp-content/uploads/2017/12/amd-q3.png

When was the last time AMD provided anything but the most conservative estimates possible?

 

[Topweasel]

When was the last time AMD provided anything but the most conservative estimates possible?

Yeah didn't their guidance for 4th quarter suggest that they would revenue in 4th compared to 2016? That didn't happen. But AMD is a company that has something like 15 years of nearly consecutive losses outside like one year in the mid 2000's. They are never going to be optimistic about their numbers. They are probably fine with the 10% shifts up and down on stock instead of chasing company valuation till they have solidified their market presence.

 

[CatMerc]

When was the last time AMD provided anything but the most conservative estimates possible?

That's still reaching higher gross margin than AMD predicts to reach in 2 years from now at the very least.

Unless EPYC sells monsterously well out of nowhere (talking something akin to taking 5-10% of the server market in a single quarter), they won't reach that lofty goal he predicted.

 

[scannall]

That's still reaching higher gross margin than AMD predicts to reach in 2 years from now at the very least.

Unless EPYC sells monsterously well out of nowhere (talking something akin to taking 5-10% of the server market in a single quarter), they won't reach that lofty goal he predicted.

Epyc is a great part. And it will likely make some headway. But, getting into server centers takes a *LONG* time. There was a lot of wishful thinking and wild speculation early on that didn't seem realistic.

 

[CatMerc]

Epyc is a great part. And it will likely make some headway. But, getting into server centers takes a *LONG* time. There was a lot of wishful thinking and wild speculation early on that didn't seem realistic.

I don't disagree. I think EPYC is quite attractive, and I expect 5% of the server market share to be AMD's by Q4 2018. But the chances of something like that happening in a single quarter are next to nothing.

 

[Udgnim]

No margins are not fine.
I accept blog is skewed somewhat with console APUs which are high volume very low margins.
But AMD would really need to be hitting 45-50% margins to be considered stable and healthy imo.
Increase in volume on top of that would be welcome.

AMD has covered their margin goals on Slides 36 and 37

AMD's goal is >36% for 2018 and >40% by 2020

http://quarterlyearnings.amd.com/static-files/fb47454d-ecb7-4462-92d6-7c9bf7143553

 

[moinmoin]

That because he only saw the easy stuff. His CPU had four actual dies. That doesn't make TR4 able to be a four die platform. There are a dozen reasons that he could still be wrong about TR having 4 dies. Then there is a dozen reasons TR4 cant support four dies. The simplest is that TR4 specs for motherboard manufacturers won't have all the needed traces to support 4 dies. It would have made x399 boards even more expensive.

He mentioned three possible reasons: Mechanical support (which I agree with him is plain silly), a way to increase the "yield" by using otherwise unused dies, or his conclusion that Threadripper is more of a cut down Epyc than AMD admits that electronically may allow for four full dies even if the platform remains locked to 4 memory channels and 64 PCIe lanes. What other "dozen reasons" are you thinking of that he could still be wrong?

 

[Topweasel]

We don't know how well if at all a die could run without memory connected to it. We don't know if retail CPUs actually have four dies. We don't know if the boards have the traces and probably don't for the extra pin usuage. We don't know what AMD specced the power usuage for. 4 dies clocked over a GHz slower would defeat the purpose of Threadripper. The list goes on.

 

[eek2121]

He mentioned three possible reasons: Mechanical support (which I agree with him is plain silly), a way to increase the "yield" by using otherwise unused dies, or his conclusion that Threadripper is more of a cut down Epyc than AMD admits that electronically may allow for four full dies even if the platform remains locked to 4 memory channels and 64 PCIe lanes. What other "dozen reasons" are you thinking of that he could still be wrong?

Honestly, I think it has to do with the way they build and test the chips. Before I get into this, let me say 2 things: I know only a bit about how CPUs are built these days, but I'm also an engineer and I'm pretty good at making educated guesses. That being said, feel free to take this with a grain of salt.

The cheapest way to build anything is to keep as simple and as straight forward as possible. This means keeping all of the components the same, making very few variances, etc. This would mean that every Threadripper CPU, whether it be the 1900x, 1920x, or 1950x would be exactly the same, and they are. So why did they use 4 dies instead of 2? To increase the overall yields of the chip (I believe this was one of the things he mentioned in this video). It gives them the most flexibility. You only need 2 dies, so if one or another is defective, you can turn them off. If you want to build a 1950x, you just need 2 fully functional dies. If you want a 1920x, you just need 6 cores on 2 dies to be working. If you want to build a 1900x, you only need 4 cores on 2 dies working. If they do things this way, that means that yes, you'll waste a bit more money, but your overall yields for the chip itself are almost guaranteed to be 99%, so it pays off. These chips aren't costing AMD a bundle of money to make. I guarantee you that even the 1950X costs AMD less than $200 to build. So wasting a few dies in each chip isn't a big deal.

Now here is the kicker: I believe that his point about a 32 core Threadripper CPU is also valid. However, they decided not to do so for Gen1 due to TDP and lack of a use case and limited boost technology. A 32 core threadripper CPU would almost certainly need much lower clocks in order to make TDP. As it is, without overclocking, a Gen1 Threadripper quickly drops down to 3400 MHz when placed under load in order to keep within TDP. If there were twice the cores, you would almost certainly need to set the base clock much lower, maybe even by half (1.7 GHz for 32 core) to stay within TDP. Now, for Zen+ on 12nm, the situation may be a bit different. They've reworked precision boost, and we will also see slightly better voltage/clockspeed curves. This would likely help raise the base clocks without raising voltages. If they can optimize things enough, we could potentially see a 24 core Threadripper or maybe even 32 core Threadripper. AMD actually has competitive reasons to release at least a 24 core Threadripper to one up the 7980xe.

Now, let's talk about the 7980xe for a moment. Excluding overclocking, here are the stats:

Cores: 18
Threads: 36
Base Frequency: 2.6 GHz
Max Single Core Turbo Frequency: 4.4 GHz
Max All-Core Turbo Frequency: 3.4 GHz

Now lets compare this first to my personal 1950X:

Cores: 16
Threads: 36
Base Frequency: 3.4 GHz
Boost Frequency: 4 GHz
Max stable all core OC frequency: 4.2 GHz (4.4 if I'm willing to kill my chip from 1.6V)

Keep the above in mind for this next part. Zen+ should bring around a 5-10% improvement in clock speed. In addition, it features a reworked precision boost that should operate a lot more like Intel's chips. Knowing this, I imagine we'll see something like this:

Zen+ 8 core (2900X) (and also the AM4 2800X)

Base Frequency: 4.0 GHz
Max Boost Frequency: 4.4 GHz (all core)

Zen+ 12 Core (2920X)

Base Frequency: 3.8 GHz
Max Boost Frequency: 4.4 GHz (all core unknown)

Zen+ 16 core (2940X?)

Base Frequency: 3.6 GHz
Max Boost Frequency: 4.4 GHz (not all core)

Zen+ 24 Core (2960X?)

Base Frequency: 2.6 GHz
Boost Frequency: 4.4 GHz (not all core)

Zen+ 32 Core (2980X?)

Base: 2.2 GHz
Boost: 4.4 GHz (not all core)

Now, you may be doubting the above, but it's really all about math, what we know about current leaks, and knowledge of current hardware. Ignoring TDP, Ryzen is already capable of hitting 4.4 GHz. It just requires obscene voltage to do it due to the LPP process (I can show you a screenshot of my 1950x at 4.4 ghz if you'd like). 12nm is supposed to be performance optimized, so the curves should be much better. That being said, we will probably still not see very high overclocks. The other improvements to Zen+ should mean a slightly higher IPC and much better gaming performance though. Overall, I expect that this next Ryzen release is going to bring MUCH more to the table, more than anyone has ever imagined.

Remember this post as we go through the year. AMD may be a bit more conservative about clocks, but I imagine these numbers won't be far off the mark. We'll see at least a boost clock of 4.2 GHz and I'm putting my money on 4.3-4.4 GHz as stated above. I'm usually pretty good at figuring these things out. I was mostly on the mark about the 1950X and people said I was full of shit. I claimed that the final chip would be around 3.6 GHz base with a 4.0 boost. The end result ended up being 3.4 base and a 4.0 (xfr 4.1) boost, but that I suspect was to make the chip easier to cool more than anything.

 

[raghu78]

I don't disagree. I think EPYC is quite attractive, and I expect 5% of the server market share to be AMD's by Q4 2018. But the chances of something like that happening in a single quarter are next to nothing.

I think AMD can get to atleast 10% server share (probably even more) by end of 2018. Its quite possible and very realistic given the Meltdown patch performance hit to Intel server CPUs. btw Lisa Su expects to hit 5% server share by Q4 2018 and she is very very conservative when giving out financial guidance like growth %, gross margin % or even share gain expectations. charlie believes Intel's server share will take a huge hit and he seems to know that AMD EPYC is ramping big at the cloud providers and OEMs. He even stated "I have a very good idea of what orders are placed and by whom, not to mention when. I am laughing at the low-to mid-single digit bit though." when i told him people still think EPYC will only get low to mid single digit share.

https://semiaccurate.com/2018/01/04/kaiser-security-holes-will-devastate-intels-marketshare/

https://twitter.com/ryanshrout/status/954015419832119297

 

[eek2121]

We don't know how well if at all a die could run without memory connected to it. We don't know if retail CPUs actually have four dies. We don't know if the boards have the traces and probably don't for the extra pin usuage. We don't know what AMD specced the power usuage for. 4 dies clocked over a GHz slower would defeat the purpose of Threadripper. The list goes on.

Yes, we do know this. If you would have watched the video, you'd see that all 4 dies are real, actual Zen dies. Multiple Threadrippers have been delidded and they all have 4 dies. Memory and PCIE lanes are not an issue, I don't think you understand how these things work. AMD specs CPUs based on TDP and safe voltage levels for their chips. Read my above post to get an idea of what a 32 core Threadripper would look like on Zen+. It is 100% possible, the socket supports it, and there is no real reason NOT to unless they don't see a market for it...but I'm willing to bet that with Intel releasing the 7980XE we will at least see a 24 core, and at that point, there is no reason not to release a 32 core as well.

EDIT: Oh and a 32 Core Threadripper AMD could get away with charging 2 grand for, that means the profit margin on that part would be insanely high. AMD released margin projections with the last quarterly financials I looked at, and they predicted much higher margins as we move towards 2020. This is one of the ways they will get there. People buy the 7980XE. Yes it's 2 grand, but people buy it. If AMD released a 32 core Threadripper and the reworked boost kept it competitive at gaming, hell I would buy it at 2 grand. The move to 16 cores was the best decision I ever made, even for mundane stuff like web browsing while playing a game or watching a movie, etc. Having the extra compute power for video encoding and game engine development is great as well.

 

[Yotsugi]

I think AMD can get to atleast 10% server share (probably even more) by end of 2018. Its quite possible and very realistic given the Meltdown patch performance hit to Intel server CPUs. btw Lisa Su expects to hit 5% server share by Q4 2018 and she is very very conservative when giving out financial guidance like growth %, gross margin % or even share gain expectations. charlie believes Intel's server share will take a huge hit and he seems to know that AMD EPYC is ramping big at the cloud providers and OEMs. He even stated "I have a very good idea of what orders are placed and by whom, not to mention when. I am laughing at the low-to mid-single digit bit though." when i told him people still think EPYC will only get low to mid single digit share.

https://semiaccurate.com/2018/01/04/kaiser-security-holes-will-devastate-intels-marketshare/

https://twitter.com/ryanshrout/status/954015419832119297

Charlie believes in a lot of things.
Can't wait for Snowy Owl tho.

 

[moinmoin]

@eek2121 thanks

We don't know how well if at all a die could run without memory connected to it. We don't know if retail CPUs actually have four dies. We don't know if the boards have the traces and probably don't for the extra pin usuage. We don't know what AMD specced the power usuage for. 4 dies clocked over a GHz slower would defeat the purpose of Threadripper. The list goes on.

Did you watch the video? He specifically bought retail Threadripper and Epyc CPUs. He specifically pointed out that the traces appear to be the same. That Threadripper wouldn't be able to use the additional Epyc pins should be obvious, none of the SP3v2/TR4 boards would be prepared for it anyway. But IF is already able to route any memory controller and any PCIe interface through any die anyway, so Threadripper's limit to 4 channels and 64 lanes doesn't require to disable dies, it would just naturally increase latency and bandwidth saturation.

 

[eek2121]

@eek2121 thanks

Did you watch the video? He specifically bought retail Threadripper and Epyc CPUs. He specifically pointed out that the traces appear to be the same. That Threadripper wouldn't be able to use the additional Epyc pins should be obvious, none of the SP3v2/TR4 boards would be prepared for it anyway. But IF is already able to route any memory controller and any PCIe interface through any die anyway, so Threadripper's limit to 4 channels and 64 lanes doesn't require to disable dies, it would just naturally increase latency and bandwidth saturation.

Right, this, and AMD has already mentioned in passing that they've improved latencies in Zen+. They could also easily do a chipset refresh and work with motherboard manufacturers to implement the full amount of PCIE lanes and 8 channel memory if they wanted...older boards would just have lower performance...or maybe they actually do support the spec already, who knows?

 

[Genx87]

Margins are fantastic where are your numbers?

34% margins in this space is horrible. Their competition is into the 60s.