Did AMD miss a big opportunity in 2017?

[cbn]

For Pre-built desktop, how about workstations? (EPYC)

P.S. I see the specs mention a "default TDP":

http://www.amd.com/en/products/epyc-7000-series-1-socket-models

http://www.amd.com/en/products/epyc-7000-series-2-socket-models

So maybe a higher TDP option as well for workstation usage?

 

[Spartak]

The desktop is actually ~40% of the PC shipped in 2016.

Also, OEMs have been stuck with Intel because, before 2017, AMD hasn't had competitive processors since 2011.

Only a small slice of the shipped desktop PC market are enthusiasts / gamers PCs since enthusiasts / gamers tend to build their own PCs.



Dell, HP, Lenovo, and Acer are all selling PCs with AMD processors (Summit Ridge, Bristol Ridge, Carrizo, Kaveri, etc.)

You are splitting hairs when many people already answered your question. There was no missed opportunity since they need time to roll out Zen, the APU's add complexity, globalfoundries needs to ramp production capacity, and they need time to get the OEMs onboard.

We're not talking one factor but a whole set that explains why. You seem to live in a fantasy world where at their executive meeting deciding on the zen rollout a majority voted for "let's take it easy and focus on our fans first".

 

[Vattila]

Question is, what will be the strategy for 7nm Zen 2? Will it be the same? I.e. start with a pure CPU die with no integrated graphics first, followed by a shrunk APU later on?

The OP points out the disadvantage, compared to Intel, of not having integrated graphics in the mainstream desktop space. I have argued that AMD could attack many markets with a single 7nm APU die (3 CCX + 1 GCX), assuming a modular graphics design is realised and ready.

So, at 7nm, perhaps start with rolling out this hypothetical 12-core APU as the successor to Ryzen 7 in the desktop market, ramping it as fast as possible for Ryzen 5/3 and Ryzen Mobile, while at the same time validating the server APU based on the same die. Then follow up with a 16-core pure CPU die for that rumoured 64-core EPYC successor, to maintain a lead in the x86 core-count war.

 

[Yotsugi]

I.e. start with a pure CPU die with no integrated graphics first, followed by a shrunk APU later on?

Of course.

Gotta hit Intel before they can roll EMIB products out.

 

[Flash831]

No, they couldn't do it any other way. Ryzen/Epyc/Threadripper all entered areas where AMD had little presence before. They replaced old Vishera products that AMD's partners had like 2 years to clear out their inventory (2015 & 2016).

Raven Ridge is directly replacing Carrizo/Bristol Ridge. Products that probably do not sell like hot cakes (but still not completly obsolete).
Pushing Raven Ridge out the door would instantly make the channels (AMD & the partners) inventory of Bristol Ridge parts unsellable. Millions of dollars flushed down the toilet.

Or they could simply let the channel clear old inventory and continue improving RR and release it when it is most economically wise to do so.

 

[cbn]

Question is, what will be the strategy for 7nm Zen 2? Will it be the same? I.e. start with a pure CPU die with no integrated graphics first, followed by a shrunk APU later on?

The OP points out the disadvantage, compared to Intel, of not having integrated graphics in the mainstream desktop space. I have argued that AMD could attack many markets with a single 7nm APU die (3 CCX + 1 GCX), assuming a modular graphics design is realised and ready.

How many stream processors in a 3CCX + 1 GCX APU?

I am assuming it would be pretty low for an enthusiast, but then maybe as Workstation/FirePro APU with capability compared to some entry level cards:

And if the iGPU had 1/2 rate double precision FP (Ideally with ECC enabled HBM type memory) it could be used for FP64 Workstation tasks accelerated by GPU. (re: 1/2 rate double precision needs a high amount of bandwidth)

 

[StinkyPinky]

Is Vega really a dumpster fire? Sure it won't compete against the 1080ti but that's not a huge market anyway.

Vega 64 could end up like the 290 and 290x. No one really liked them at release but they ended up having quite long longevity.

 

[Yotsugi]

Is Vega really a dumpster fire?

Currently? Yeah.

Sure it won't compete against the 1080ti but that's not a huge market anyway.

It should, there's no way around it.

Vega 64 could end up like the 290 and 290x.

These performed well launch day. And they were released feature complete too.

 

[Vattila]

How many stream processors in a 3CCX + 1 GCX APU?

It needs to at least beat Intel. A GCX about the size of 11 CUs seems to fit nicely with the size of a CCX. Here is my mock-up based on the Raven Ridge die.

https://forums.anandtech.com/thread...will-all-be-apus.2523446/page-2#post-39139778

 

[Yotsugi]

Why would you waste die area for GPU?

Datacenter folk that need GPUs will attach discrete boards anyway.

 

[Mockingbird]

No, they couldn't do it any other way. Ryzen/Epyc/Threadripper all entered areas where AMD had little presence before. They replaced old Vishera products that AMD's partners had like 2 years to clear out their inventory (2015 & 2016).

Raven Ridge is directly replacing Carrizo/Bristol Ridge. Products that probably do not sell like hot cakes (but still not completly obsolete).
Pushing Raven Ridge out the door would instantly make the channels (AMD & the partners) inventory of Bristol Ridge parts unsellable. Millions of dollars flushed down the toilet.

Or they could simply let the channel clear old inventory and continue improving RR and release it when it is most economically wise to do so.

I don't know if you realize how pathetic Bristol Ridge really is.

A12-9800, which is the fastest Bristol Ridge processor, loses to Pentium G4560 in multi-threaded tasks.

AMD doesn't have a processor with integrated graphics that can compete with Pentium, let alone, Core i3 or Core i5.

https://www.techspot.com/review/1486-amd-a12-9800/

 

[DrMrLordX]

I don't know if you realize how pathetic Bristol Ridge really is.

A12-9800, which is the fastest Bristol Ridge processor, loses to Pentium G4560 in multi-threaded tasks.

AMD doesn't have a processor with integrated graphics that can compete with Pentium, let alone, Core i3 or Core i5.

It was late to market, and they let it langour in OE space for too long. It is sad that AMD chose to bloat up the channel with those APUs this late in the game, making it financially injurious for them to just go ahead and launch Raven Ridge now. Stupid decision making there. Bristol Ridge should already be on its way out of the retail channel.

 

[Cloudfire777]

Raven Ridge is a fantastic design by AMD.
The biggest issue for AMD now is getting products out there with Raven Ridge
AMD have been the second choice for the majority of notebook OEMs, if not everyone, for so many years now they have a pretty big challenge ahead of them with that. And its gonna take a while to normalize

 

[Yotsugi]

Raven Ridge is a fantastic design by AMD.
The biggest issue for AMD now is getting products out there with Raven Ridge
AMD have been the second choice for the majority of notebook OEMs, if not everyone, for so many years now they have a pretty big challenge ahead of them with that. And its gonna take a while to normalize

I think it'll be easier than you imagine, considering how easily AMD concieved OEMs into adopting EPYC.

 

[cbn]

Why would you waste die area for GPU?

I don't think it wastes much area (or maybe any area) when we consider the 12C iGPU-less alternative would be a rectangular die (which wastes die area in a different way-- increased wafer edge loss)

 

[Yotsugi]

I don't think it wastes much area (or maybe any area) when we consider the 12C iGPU-less alternative would be a fairly rectangular die (which wastes die area in a different way-- increased wafer edge loss)

Why would you waste die area on it instead of slapping more cores?

Like, really.

 

[cbn]

Why would you waste die area on it instead of slapping more cores?

Like, really.

Remember it is still dual channel memory though.

With that mentioned, I think 16C should work with dual channel DDR5.....but I believe we are talking about a DDR4 system here.

 

[Yotsugi]

Remember it is still dual channel memory though.

Throw some more caches (which is exactly what they are going to do according to rumors).

 

[Mockingbird]

It was late to market, and they let it langour in OE space for too long. It is sad that AMD chose to bloat up the channel with those APUs this late in the game, making it financially injurious for them to just go ahead and launch Raven Ridge now. Stupid decision making there. Bristol Ridge should already be on its way out of the retail channel.

I am pretty sure that the real reason that AMD launched Bristol Ridge (along with compatible chipsets/platforms) back in 2016 is to provide OEMs with platforms that have drop-in compatibility with Ryzen.

Look at HP ENVY x360 with Ryzen Mobile processor.

The reason HP is able to launch it so quickly is that it is the same laptop that HP launched back in 2016 except with Raven Ridge processor instead of Bristol Ridge (and a few minor changes).

Now, I am not sure why AMD launched Bristol Ridge into retail channel because it makes no sense.

 

[cbn]

Throw some more caches (which is exactly what they are going to do according to rumors).

That would grow the die size further.

I am not saying this couldn't happen, but I am guessing even with extra cache it would limit each die (with the full 16C) to a lower level of TDP.

 

[cbn]

Regarding Bristol Ridge one strength it has a very high amount of double precision floating point on the iGPU.

So it would have been interesting perhaps as FirePro APU on desktop.

 

[cbn]

Throw some more caches (which is exactly what they are going to do according to rumors).

That would grow the die size further.

I am not saying this couldn't happen, but I am guessing even with extra cache it would limit each die (with the full 16C) to a lower level of TDP.

In this post here I mentioned the use of HBM stacked on the iGPU part of the processor die.

How about doing this with SRAM (ie, processor cache) for the CPU cores on a pure CPU die.

Think of this as something like "Tiles", but connecting upwards rather than outwards as with Intel EMIB.

 

[DrMrLordX]

Now, I am not sure why AMD launched Bristol Ridge into retail channel because it makes no sense.

Exactly. It only makes sense as a low-cost part to fill out the AM4 lineup, yet it does so rather poorly by modern standards. It's really not worth it especially if it makes it harder to get Raven Ridge out the door.

Regarding Bristol Ridge one strength it has a very high amount of double precision floating point on the iGPU.

So it would have been interesting perhaps as FirePro APU on desktop.

While I would like to see that happen, it appears that almost nobody is making use of that feature except perhaps some semicustom buyers who were apparently the ones who got that thrown into the design in the first place.

 

[Yotsugi]

How about doing this with SRAM (ie, processor cache) for the CPU cores as well.

You're suggesting stacking a separate SRAM die on top/underneath the CPU?
Intel did it in Polaris.
It will be seriously thermally limited.

 

[cbn]

You're suggesting stacking a separate SRAM die on top/underneath the CPU?
Intel did it in Polaris.
It will be seriously thermally limited.

Yes, that is what I am thinking about.

But according to following sources that didn't appear to limit the chip thermally:

http://xtreview.com/addcomment-id-1624-view-intel-polaris-80-core-processor.html

https://en.wikipedia.org/wiki/Teraflops_Research_Chip#Statistics_.5B21.5D

http://blog.shubhamsaxena.com/intel-polaris-80-core-cpu/

The prototype of processor polaris has the design concept LGA with 1248 contacts, it operates at the frequency 3.16 GHz with vcore 0.95 v, TDP level in this case is equal to 62 W. At this frequency the performance of processor exceeds one teraflop. To increase the processor performance intel has to increase the clock frequency. For example, at frequency 5.1 GHz processor demonstrates performance on the order of 1.63 teraflop, in this case the energy consumption reaches 175 W. At the frequency 5.7 GHz the performance reaches 1.81 teraflop, the energy consumption grows to 265 W.