AMD CEO talks of long-term turnaround

[erunion]

this hbm stuff just introduced in amd graphics cards makes me wonder if we might one day see hbm married to a cpu.

http://www.extremetech.com/computin...es-between-wide-io-hbm-and-hybrid-memory-cube

 

[MongGrel]

long term turn around = were trying to stay viable atm.

 

[Fjodor2001]

Zen needs to be nothing short than the biggest miracle at AMD ever to forfill your expectations. And we all know how such dreams turns out.

Well if Zen is 8 core and Intel mainstream CPUs are still on 4 cores, it's quite obvious Intel will be demolished in MT performance.

And if Zen cores are at least around SB performance level, it will be close to Intel in ST performance.

So in that case what I said will be fulfilled. What exactly are your objections to that?

 

[Fjodor2001]

When did you say Skylake would launch again? Was it summer 2017?

No, I said 2016-2017; a wide window because the leaked release plans kept shifting, and quite late because 14 nm was seriously delayed.

In the end it turns out that Skylake will be released in 2015Q3. And not the complete lineup until 2015Q4 or 2016Q1.

So the prediction turned out to be quite accurate, and close to the beginning of the 2016-2017 window I estimated. I'm actually surprised it turned out so close to the estimate, given the uncertainty of the info available at the time.

 

[erunion]

Well if Zen is 8 core and Intel mainstream CPUs are still on 4 cores, it's quite obvious Intel will be demolished in MT performance.

And if Zen cores are at least around SB performance level, it will be close to Intel in ST performance.

So in that case what I said will be fulfilled. What exactly are your objections to that?

My objection is that your scenario wont reflect how it actually plays out.

Zen will likely suffer a large hit to clock speed, and will be in a similar competitive position versus Intel as the current FXs are.

You've convinced yourself that the best case scenario is the most likely scenario, it's not.

 

[Fjodor2001]

My objection is that your scenario wont reflect how it actually plays out.

Zen will likely suffer a large hit to clock speed, and will be in a similar competitive position versus Intel as the current FXs are.

You've convinced yourself that the best case scenario is the most likely scenario, it's not.

Ok, and what data do we have that leads to this conclusion for sure?

And could you clarify what you mean by that Zen will be in an FX like position versus Intels mainstream CPUs? Are you saying Zen performance will be around the same number of percent behind Intel's CPUs in both MT and ST performance as FX CPUs are today?

Finally, I've not convinced myself of anything, but primarily gathered opinions from others on this forum and collected info that is known so far. None of us knows how this will turn out in the end.

 

[MongGrel]

The slide from analyst day:

Fiji was a completely flop for the pro segment. So I guess the 3 vectors are server, embedded and semicustom.

Nice fantasy...

I'm not believing it myself.

 

[Erenhardt]

The slide from analyst day:

Fiji was a completely flop for the pro segment. So I guess the 3 vectors are server, embedded and semicustom.

GTX 970 is even bigger flop for pro segment as well, and its hardly on topic too!

 

[Phynaz]

this hbm stuff just introduced in amd graphics cards makes me wonder if we might one day see hbm married to a cpu.

Would you use a system that didn't have upgradable memory?

Memory bandwidth isn't an issue with CPUs anyway.

Assuming you mean x86. If not, then here ya go, http://forums.anandtech.com/showthread.php?t=2426221&highlight=

 

[Fjodor2001]

Would you use a system that didn't have upgradable memory?

How many percent of computer owners upgrade the memory on a computer today anyway? 1-5% perhaps?

Most just buy a pre-built computer and stick with it until they upgrade to a completely new computer.

 

[Dresdenboy]

I think if they could in the first place, they would have considered doing it in Vishera to begin with.

This would have led to a 3-4 years delay ...

 

[Dresdenboy]

My objection is that your scenario wont reflect how it actually plays out.

Zen will likely suffer a large hit to clock speed, and will be in a similar competitive position versus Intel as the current FXs are.

You've convinced yourself that the best case scenario is the most likely scenario, it's not.

If Zen wouldn't be an improvement over a shrunk XV, they'd already scrapped it.

And what are their goals? Operating range, perf/W, $/perf/W, etc.

 

[ShintaiDK]

If Zen wouldn't be an improvement over a shrunk XV, they'd already scrapped it.

We wouldnt have bulldozer in the first place then would we?

 

[dave_the_nerd]

Would you use a system that didn't have upgradable memory?

Memory bandwidth isn't an issue with CPUs anyway.

Assuming you mean x86. If not, then here ya go, http://forums.anandtech.com/showthread.php?t=2426221&highlight=

I am typing this on a laptop without upgradable memory.

It also has a full 4c/8t i7 CPU that will be, at least, competitive with anything AMD sells for desktop use.

*shrug*

There's a product out there for everybody, and somebody for every product.

I used to know a hard-core Mac zealot that bought a Zune and was head over heels in love with it.

 

[ShintaiDK]

Memory you cant upgrade is the future, nomatter if we like it or not. The only question is when its going to work cost wise.

 

[Fjodor2001]

Memory you cant upgrade is the future, nomatter if we like it or not. The only question is when its going to work cost wise.

I think so too, except for perhaps workstation or server computers.

For desktops there will probably be AIO computers where the motherboards contain soldered flash memory chips for storage instead separate SSDs, and HBM/HMC on the CPU die being used both as iGPU RAM and system RAM. Hopefully there will still be enthusiast class motherboards with separate components too though, even if sold in smaller quantities.

 

[DrMrLordX]

that's...not really saying it's going to be fast

Maybe, maybe not. Construction cores are odd, at least from an Intel-centric perspective. Stuff that runs well on a Sandy/Ivy/Haswell might make an AMD chip choke and die, or vice-versa. For those of us that were watching the 3DPM "fixer-upper" thread I started, I managed to produce Java code that did essentially the same thing as Dr. Cutress' 3DPM Stage 1 and did it much faster on a Kaveri. But for an Ivy Bridge, my Java code was slower. Sadly the C++ efforts in that thread haven't borne out yet. I've gotta get back on the ball one of these days and put up my own entry.

In general, Intel's faster cache implementation (and other factors) seem to make their processors more flexible when it comes to code that has not been tightly optimized for any cutting-edge uarch. AMD . . . not so much. Zen is going to have a much more "Intel-like" cache hierarchy, and it is also moving away from CMT and towards SMT, so, overall, we may see it behave a bit more like Intel's processors when handling sub-optimal code. AMD is also supporting FMA3 and AVX2 in Zen, which means the whole FMA4+xOP and split AVX instruction thing goes away.

Regardless, "40% faster IPC than Excavator" would be pretty impressive if, say, it applied to integer performance AND Zen had relatively balanced integer and fp performance profiles. It would be less impressive if Zen has 40% higher fp IPC and still has relatively-balanced integer and fp performance profiles.

 

[Tuna-Fish]

Regardless, "40% faster IPC than Excavator" would be pretty impressive if, say, it applied to integer performance AND Zen had relatively balanced integer and fp performance profiles. It would be less impressive if Zen has 40% higher fp IPC and still has relatively-balanced integer and fp performance profiles.

I've said this a few times already, and I'll say it again: IPC alone tells you absolutely nothing about performance, and AMD absolutely can hit any IPC target they want, assuming they are willing to relax clocks enough.

What matters is performance. Performance is IPC * clock speed. How difficult it is to get high IPC depends entirely on how high you set your clock targets, which determines the maximum length of the critical paths in your CPU, and thus how complex things you can do per each clock cycle.

Saying "our next chip will have 40% more ipc than the last" is kind of bullshit marketing speech. It sounds good, "40% better", but it tells you practically nothing except that they decided to aim for a more brainiac, probably wider design. Zen could have 40% higher ipc, and 25% lower clock targets, for a net gain of 5%.

 

[Phynaz]

We wouldnt have bulldozer in the first place then would we?

This. AMD needs a product to sell.

 

[Dresdenboy]

We wouldnt have bulldozer in the first place then would we?

As I heard of ppl who were not that good in school, if there is a test and time is over, one just has to deliver, whether it's a C or D.

If bonuses are paid for execution and not best CB single thread performance, then you'll see things happen the way as they did. And real silicon still behaves a bit different than in the best simulations. Of course one can't run a full Orochi chip in sth like PSPICE.

 

[erunion]

If Zen wouldn't be an improvement over a shrunk XV, they'd already scrapped it.

I didn't say that. I said it would be in the same competitive position, meaning it won't beat Intel's 2016 chips.

 

[mrmt]

If bonuses are paid for execution and not best CB single thread performance, then you'll see things happen the way as they did.

I doubt that a company with execution so broken as AMD is competent enough to judge whether something is worth a bonus or not.

 

[Phynaz]

I didn't say that. I said it would be in the same competitive position, meaning it won't beat Intel's 2016 chips.

I don't think anyone reasonable is expecting it to.

 

[DrMrLordX]

I've said this a few times already, and I'll say it again: IPC alone tells you absolutely nothing about performance, and AMD absolutely can hit any IPC target they want, assuming they are willing to relax clocks enough.

AMD is constantly hammered over the low IPC of Bulldozer. Constantly. Their marketing as of late has been shaped (or warped, depending on your point of view) by the criticism heaped upon them. What do you expect them to say about their new microarchitecture?

And what do you think Intel had to say about Broadwell or Skylake with respect to Haswell? They don't talk clockspeeds, they talk IPC.

Saying "our next chip will have 40% more ipc than the last" is kind of bullshit marketing speech.

It's par for the course these days. Nobody is advertising clockspeed during the development cycle anymore. Nobody is bragging about how they'll "break the gigahertz barrier" or anything like that. It's about efficiency. And AMD has something to prove: that they can still produce a desktop/server processor that isn't a clockspeed whore like Vishera.

It sounds good, "40% better", but it tells you practically nothing except that they decided to aim for a more brainiac, probably wider design. Zen could have 40% higher ipc, and 25% lower clock targets, for a net gain of 5%.

25% lower clockspeed targets than Excavator? Seriously? The fastest Excavator on the market right now has a base clock of 2.1 GHz, and is a 35W mobile part. Come on now.

 

[IlllI]

Would you use a system that didn't have upgradable memory?

Memory bandwidth isn't an issue with CPUs anyway.

Assuming you mean x86. If not, then here ya go, http://forums.anandtech.com/showthread.php?t=2426221&highlight=

i was thinking hbm on the cpu for the integrated graphics component. i don't really pay attention to integrated gpus but doesn't the integrated gpu rely on ddr memory? I'm wondering if instead, using hbm would give a performance improvement

http://www.extremetech.com/computin...es-between-wide-io-hbm-and-hybrid-memory-cube

thanks for the link. interesting stuff