What is going to happen to AMD?

[Ferzerp]

That's what I used to think, but it can't be true. Why would the process node and PDK be designed to handle such high clocks - neither mobile nor server need to hit 4 GHz (or possible higher). You can't produce those kind of bins 'by accident. It takes work, time and dollars - so clearly, while not it's main emphasis anymore, Desktop CPUs matter to Intel.

Single threaded performance is still a HUGE factor in server performance. If it weren't, the AMD solution would be far more attractive and have more than the almost non-existent market share that it does.

The embarrassingly parallel workload is just not as common I think you think it is. It takes some specialized software, and other than the big players who are running entirely custom code, most of us still need great single threaded performance, along with "enough" cores in our server environments.

Remember, scale out requires specific support, but scale up always benefits everything. When it comes to CPUs, a faster core is a scale up solution whereas more cores is scale out.

 

[AtenRa]

At 15W it would be absolutely insane to use 896 shaders. 15W now is badly held back by tdp, especially when the CPU is heavily loaded.

2-3x current 15W performance is on the order of 384-512 shaders at low clocks. AMD is better served by bringing two dies to the market, dual core Zen (with HT) + 384 shaders and a desktop 4 core zen + 896 shaders.

AMD already is using 512 shaders at 19W TDP with a 28nm planar process on Kaveri and 512 shaders at 15W TDP with 28nm HDL in Carrizo.

14nm FF will have 60% power reduction over 28nm planar and more than 2x the density. Add HBM Gen 2.0 and I dont see why 896 shaders will be too much for 14nm FF in 2017.

So, in 2017 an APU with all those features (14nm FF, HBM 2.0, mArchitecture, ZEN CPU Cores) could have 2-3 times the performance of 19W TDP Kaveri. That is within the performance capabilities of the 2017 technology.

 

[AtenRa]

I know what you said. And I said there is no way to get an 80% power reduction, which is what you are proposing would require (approx 150 watts to 30). Restating what you said will not make it realistic. And mobile is actually going away from even 30 watt TDP toward 15 watt and lower. At the *very best* one might be able to get HD7770 level or lower performance in a 35 watt TDP. Even then with 640 shaders and no memory bottleneck I think you would have to downclock it (edit: a lot) to get within 30 watts for cpu plus gpu.

I never claimed the performance of R7 260X in 15W TDP in 2017.

Again read carefully what i said, a 14nm FF with 896 shaders and HBM Gen 2.0 could have 2-3 times the performance of current 15W TDP products (like 19W TDP Kaveri).

Edit: Just to remind everyone, 2016 Intel will have 15-28W TDP Skylake GT3e and GT4e.

 

[DrMrLordX]

Absolutely. Stacked memory like HBM, HMC and Wide I/O will replace conventional memory as time goes. The biggest issue in the long run is of course with servers. But in the future you will buy a CPU with fixed amount of stacked memory without any option to expand.

Maybe this thread isn't the perfect place to discuss it, but I'll throw out the question anyway: what's to stop a motherboard OEM or chipset manufacturer from putting together a hybrid DDR4 + HBM/HMC/Wide I/O solution? Even if it's some kind of ghetto DRAM cache for the storage subsystem . . . what do you think takes a hit on a desktop that goes over its system memory limits? That's right, the system hits the disk and performance tanks.

If I had an unexpandable HBM memory system, I'd love to have 16-32 gigs of DRAM storage cache to fall back on before I had to swap to an SSD or *gasp* platter-based drive. There are other things you could probably do with that cache besides have superfast swap.

 

[ShintaiDK]

Maybe this thread isn't the perfect place to discuss it, but I'll throw out the question anyway: what's to stop a motherboard OEM or chipset manufacturer from putting together a hybrid DDR4 + HBM/HMC/Wide I/O solution? Even if it's some kind of ghetto DRAM cache for the storage subsystem . . . what do you think takes a hit on a desktop that goes over its system memory limits? That's right, the system hits the disk and performance tanks.

If I had an unexpandable HBM memory system, I'd love to have 16-32 gigs of DRAM storage cache to fall back on before I had to swap to an SSD or *gasp* platter-based drive. There are other things you could probably do with that cache besides have superfast swap.

Nothing prevents it. And its most likely such a way servers will be for a long time. However in desktop and mobile I dont see it happening outside a possible short window. OEMs etc will go directly for the "kill". Its also messy and costly to support both.

 

[Yuriman]

so you are comparing an APU that might be reality in 2017/2018 and comparing that to a current gen intel cpu?

even with 40% IPC bump i am not expecting AMD to beat skylake performance let alone whats available in 2017/18. we are already seeing EDRAM coming to lga and intel is only going to make things better

With a 40% IPC bump, in most cases Zen would still be behind Sandy Bridge. However, that may be enough if they can fit it in a small enough power envelope.

I feel the market for large-iGPU chips is small, but I'm excited at the idea because I'm part of that small market.

 

[RampantAndroid]

Patents don't last forever. I don't know when the filing dates were for those specific patents, but the products based on them came out over a decade ago.

20 years, iirc. There are some extra complications on it, but it's valid long enough to put Intel's main CPUs out of production if they lose access to the patent. But if it's a sure thing, then there's no concern about the patent.

As Microsoft kept alive Apple in the 90s, Intel will keep AMD alive... they need to.

MS floated Apple as a result of a court case. Intel has no incentive to do the same; Intel is competing with ARM these days, with PCs on the decline.

 

[Ajay]

Intel has 47 W Haswell mobile parts that can turbo to 4 Ghz. It's probably not for very long and only for one core, but it's there.

Yes, and it's a definite edge, but it's there because Intel's process and uArch can handle 4 GHz. If we took desktops out of the equation, Intel wouldn't likely put in the extra effort to hit that high a clock. So my point isn't that Intel isn't thinking about their growth markets first (Enterprise and mobile), but that they are still optimizing their process and PDK for high clocks mainly for the DT market. It's not a Zero sum game. It still pays to have the best DT performance.

 

[2is]

The odds of Zen landing as a Conroe-esque counterpunch are quite slim. The process gap won't be quite as severe as AMD has had to deal with the last few years, but it will still be a very real thing. I think what we'll see is a jump back to Phenom II era competitiveness with the top end Zen trading blows with top end i5 models and some specific workloads nipping at the heels of some i7 ones. Additionally, I'd guess that they will run a little hotter and draw a little more power but come at attractive price points. AMD will be fine in the long run, they have nice safe business in the consoles and graphics divisions and who knows, maybe Zen and K12 will round the CPU division back into shape.

Their market share in GPU's continues to decline vs nVidia though, not sure how "safe" I'd feel relying on that metric, and while their volume of PS4/XB1 APU's might be large, i'm willing to bet the profit margins are quite slim.

 

[Silverforce11]

so you are comparing an APU that might be reality in 2017/2018 and comparing that to a current gen intel cpu?

even with 40% IPC bump i am not expecting AMD to beat skylake performance let alone whats available in 2017/18. we are already seeing EDRAM coming to lga and intel is only going to make things better

It doesn't need to beat Skylake, it just needs to be sufficient for a balanced design, because the purpose of an APU is to provide a great CPU/Graphics option that's capable. Currently, none of them are able because they can't game at 1080p (mainstream resolution) at acceptable performance, heck, Intel's can barely game at 720p in modern games at 20 fps or less. Thus for a gaming rig, these APUs waste half their die space doing nothing while dGPUs handle all the graphics. This is the reason why AMD's APU aren't popular because they cannot serve their purpose, limited by DDR3 bandwidth.

The problem with edram is its small, cache-like that gets swamped at higher resolutions & better texture quality.

Intel will need HBM to compete. We're talking about APUs with 4/8GB HBM2 delivering insane bandwidth (matching dGPUs). On the server side, if the software takes advantage of HSA & OpenCL, nothing Intel makes would come close to that potential.

For gaming, it will obliterate Intel's iGPU combo.

 

[Ajay]

Single threaded performance is still a HUGE factor in server performance.

Yes, but the clocks are lower on server CPUs because of the many cores on die. High IPC is still a must, and that were the BD variants were lacking (among other problems as well).

 

[Ajay]

It doesn't need to beat Skylake, it just needs to be sufficient for a balanced design, because the purpose of an APU is to provide a great CPU/Graphics option that's capable. Currently, none of them are able because they can't game at 1080p (mainstream resolution) at acceptable performance, heck, Intel's can barely game at 720p in modern games at 20 fps or less. Thus for a gaming rig, these APUs waste half their die space doing nothing while dGPUs handle all the graphics. This is the reason why AMD's APU aren't popular because they cannot serve their purpose, limited by DDR3 bandwidth.

I don't expect Zen to match Skylake, but it would be much better if it did. I guess the key will be gaming performance when coupled with a dGPU. If Zen stays within 5 fps on demanding games - things will go much better for AMD. The likely reason that there is no iGPU is that leaving it out cut down design & verification time a fair bit** - and AMD can keep the die size smaller and open the way for hitting higher clock targets (which will be necessary to compete with Intel).


**Zen will go from blue sky to a shipping CPU in something like 4 years, Haswell took 5. AMD needed to drop something to hit this aggressive timeline.

 

[Enigmoid]

AMD already is using 512 shaders at 19W TDP with a 28nm planar process on Kaveri and 512 shaders at 15W TDP with 28nm HDL in Carrizo.

14nm FF will have 60% power reduction over 28nm planar and more than 2x the density. Add HBM Gen 2.0 and I dont see why 896 shaders will be too much for 14nm FF in 2017.

So, in 2017 an APU with all those features (14nm FF, HBM 2.0, mArchitecture, ZEN CPU Cores) could have 2-3 times the performance of 19W TDP Kaveri. That is within the performance capabilities of the 2017 technology.

And 19W kaveri throttles like mad, especially when the game is CPU heavy.

Which Kaveri APU is 512 shaders at 19W? The only mobile 512 shader Kaveri chip on the market that I can find is the FX-7600P.

 

[cbn]

R7 260X at 28nm has a die size of 160mm2(including 128bit memory). It has 896 cores with a TDP of 115W and 2GB GDDR-5 memory.

At 14nm (lets take 2x density only) it will have a die size of ~80mm2, add 4GB of HBM Gen2.0. In 2017 add four 14nm ZEN cores

(snip)

That APU at 15-35W TDP will have 2-3 times the performance of the current APUs. This is a perfect Laptop product

A 15-35W Zen quad core with 896sp (even on GF 14nm LPP) would be clocked pretty low.

I am thinking dual channel DDR4 3200 would probably be more cost effective than 4GB of HBM.

With that mentioned, I don't know what the price on HBM will be like in 2017. I'm guessing it will be expensive and warranted only with larger, higher clocked iGPUs (like the many core Server APUs)

 

[erunion]

Their market share in GPU's continues to decline vs nVidia though, not sure how "safe" I'd feel relying on that metric

That's the elephant in the room.
Their discrete GPU marketshare was 24% last quarter!


http://www.gamesindustry.biz/articles/2015-04-17-amd-lost-USD180-million-and-gpu-market-share-in-q1

 

[cbn]

Deleted.

 

[Boze]

Dear God, this thread is proof that hope springs eternal for AMD lovers.

Folks, AMD is so far behind Intel in terms of raw computational research for CPU designs, and so far behind NVIDIA in terms of market share, that bankruptcy is inevitable at this point. Its not a matter of if, its a matter of when.

Zen comes out in Q3 2016? Maybe worst case scenario, Q4 2016? Guess what'll be out then... Skylake-E. Zen doesn't seem to be positioned to be a budget CPU either... it looks to be squarely in the $200 - $300 range. If I'm going to drop $300 on a CPU, I might as well drop $375 for a Skylake-E that will almost certainly be a 3.3 to 3.6 gHz hexacore processor, and could very well be an octacore. I would not one bit be shocked to find out that the Core i7-7820K is an 8-core 16 thread processor running at a base of 3.3 gHz.

For Zen to keep the company alive, would be the mother of all Hail Marys. Like Fox Mulder... I *want* to believe. But the situation is dire... is the 4th quarter, 4th down and 10, and there's 3 seconds left on the clock... gonna take a real miracle to keep this company afloat come 2017.

 

[cbn]

That's the elephant in the room.

Their discrete GPU marketshare was 24% last quarter!


http://www.gamesindustry.biz/articles/2015-04-17-amd-lost-USD180-million-and-gpu-market-share-in-q1

I know the article mentions the launch of 980 GTX and 970 GTX causing a decrease in AMD market share, but I wonder how much of AMD's market share is also caused by a decrease in their low end dGPU (eg, R7 250,etc)?

The reason I am asking is because Nvidia seems to have a good line-up at the low end and these cards went on sale far more often than the AMD R7 240 and R7 250 cards ---> http://forums.anandtech.com/showthread.php?t=2403091

P.S. For whatever reason the far more powerful R7 250X GDDR5 (my own daily driver card) goes on sale for a lower price than the much weaker R7 250 DDR3/GDDR5. It is almost as if the R7 250 is offered only in very limited production (which keeps price relatively high relative to its performance). In fact, not one time did the R7 250 even make it into that Anandtech entry level hot deals thread I linked. Only the R7 240 made the price level at that time which was $50 and under.

 

[2is]

Dear God, this thread is proof that hope springs eternal for AMD lovers.

Folks, AMD is so far behind Intel in terms of raw computational research for CPU designs, and so far behind NVIDIA in terms of market share, that bankruptcy is inevitable at this point. Its not a matter of if, its a matter of when.

Zen comes out in Q3 2016? Maybe worst case scenario, Q4 2016? Guess what'll be out then... Skylake-E. Zen doesn't seem to be positioned to be a budget CPU either... it looks to be squarely in the $200 - $300 range. If I'm going to drop $300 on a CPU, I might as well drop $375 for a Skylake-E that will almost certainly be a 3.3 to 3.6 gHz hexacore processor, and could very well be an octacore. I would not one bit be shocked to find out that the Core i7-7820K is an 8-core 16 thread processor running at a base of 3.3 gHz.

For Zen to keep the company alive, would be the mother of all Hail Marys. Like Fox Mulder... I *want* to believe. But the situation is dire... is the 4th quarter, 4th down and 10, and there's 3 seconds left on the clock... gonna take a real miracle to keep this company afloat come 2017.

So you're saying there's a chance?

 

[VirtualLarry]

Dear God, this thread is proof that hope springs eternal for AMD lovers.

Folks, AMD is so far behind Intel in terms of raw computational research for CPU designs, and so far behind NVIDIA in terms of market share, that bankruptcy is inevitable at this point. Its not a matter of if, its a matter of when.

Zen comes out in Q3 2016? Maybe worst case scenario, Q4 2016? Guess what'll be out then... Skylake-E. Zen doesn't seem to be positioned to be a budget CPU either... it looks to be squarely in the $200 - $300 range. If I'm going to drop $300 on a CPU, I might as well drop $375 for a Skylake-E that will almost certainly be a 3.3 to 3.6 gHz hexacore processor, and could very well be an octacore. I would not one bit be shocked to find out that the Core i7-7820K is an 8-core 16 thread processor running at a base of 3.3 gHz.

For Zen to keep the company alive, would be the mother of all Hail Marys. Like Fox Mulder... I *want* to believe. But the situation is dire... is the 4th quarter, 4th down and 10, and there's 3 seconds left on the clock... gonna take a real miracle to keep this company afloat come 2017.

Well, the Pats have won the SuperBowl several times, under practically just that condition.

I wouldn't count AMD out just yet.

 

[2is]

Well, the Pats have won the SuperBowl several times, under practically just that condition.

I wouldn't count AMD out just yet.

How do you figure that?

One of the most talented coaches in the league and arguably in history
One of the most talented QB's in the league and arguably in history
One of the most talented tight ends in the league and arguably history

And that's before we even get into the whole cheating thing.

What alternate sense of reality are you experiencing where this is analogous to AMD?

 

[Enigmoid]

P.S. For whatever reason the far more powerful R7 250X GDDR5 (my own daily driver card) goes on sale for a lower price than the much weaker R7 250 DDR3. It is almost as if the R7 250 is offered only in very limited production (which keeps price relatively high relative to its performance). In fact, not one time did the R7 250 even make it into that Anandtech entry level hot deals thread I linked. Only the R7 240 made the price level at that time which was $50 and under.

At this point in time the yields on Cape Verde are high enough that a huge number of the chips are functional. Therefore any 250 cards are perfectly fine cards that must be cut down and sold at a lower price which is a loss compared to not cutting down the cards (I would say the same thing with the 970 given the huge numbers of 970s sold vs 980s but these are high margin parts and this cutdown is much less troublesome). AMD doesn't want to sell 250 cards at a low price vs. possible 250X sale.

In canada this is not the case.

In the US the difference is ~$9 between the cheapest 250X and 250.

 

[cbn]

At this point in time the yields on Cape Verde are high enough that a huge number of the chips are functional. Therefore any 250 cards are perfectly fine cards that must be cut down and sold at a lower price which is a loss compared to not cutting down the cards

The R7 250 (384 GCN sp) is based on a different die than the R7 250X (640 GCN sp).

R7 250X = Cape Verde
R7 250= Oland

http://www.anandtech.com/show/7751/amd-announces-radeon-r7-250x-shipping-today

Since the R7 250 was already a fully enabled Oland part, and the R7 260 was already a cut-down Bonaire part, to fill the gap AMD is calling their GCN 1.0 based Cape Verde GPU back into service.

 

[bigboxes]

My take as a very AMD friendly consumer- either they become truly competitive with their next release, or I write them off for good. And there isn't much keeping me from buying a 4790k right now.

I am a 4790K owner, but I made the switch to the 920. I love AMD, but when won't buy inferior tech when I'm an enthusiast. I hope that Zen can compete. Whatever the case, I will choose the best for me.

 

[frozentundra123456]

At this point in time the yields on Cape Verde are high enough that a huge number of the chips are functional. Therefore any 250 cards are perfectly fine cards that must be cut down and sold at a lower price which is a loss compared to not cutting down the cards (I would say the same thing with the 970 given the huge numbers of 970s sold vs 980s but these are high margin parts and this cutdown is much less troublesome). AMD doesn't want to sell 250 cards at a low price vs. possible 250X sale.

In canada this is not the case.

In the US the difference is ~$9 between the cheapest 250X and 250.

I noticed that on Newegg. 250x is GDDR5 also, and 250 can be either DDR3 or GDDR5, but usually DDR3 I think.

With that price difference, I cant see any reason at all to buy the 250 over the 250x, unless one has a power supply with no six pin connector. That 250x at 80.00 and a 70.00 Athlon X4 is a killer price for an entry level gaming rig, pretty much invalidating any APU except in very power constrained scenarios.