AMD post-Bulldozer x86 CPU architecture

[Fjodor2001]

Is ARM A15 high performance? What about Jaguar, is it high performance or low performance? What about Atom? What about desktop Core?

In all of these cases you could say that these chips are high performance or low performance, depending on what you use as a baseline for comparison. Without AMD stating what *they* consider to be the baseline for "high performance", we have just marketing at work here.

AMD said "FX Series of high-performance [...] the so-called new x86 core really is a high performance", so I think it's pretty clear we're talking about top end desktop CPUs. They will not be using the FX brand for CPUs with cat core performance.

So they have to be quite a lot faster than AMD's current top end desktop CPUs, which ought to put them in the same range as current Intel desktop CPUs. The question is how they can achieve that if they will not have access to a high clockspeed process node?

 

[SlowSpyder]

Is ARM A15 high performance? What about Jaguar, is it high performance or low performance? What about Atom? What about desktop Core?

In all of these cases you could say that these chips are high performance or low performance, depending on what you use as a baseline for comparison. Without AMD stating what *they* consider to be the baseline for "high performance", we have just marketing at work here.

Right, who knows what they mean by that. It could just be two APU's with the graphics disabled stuck under on IHS for an eight core 'high performance' part within AMD's range of products. It could really be anything AMD considers a step up from the standard APU's they sell.

 

[mrmt]

AMD said "FX Series of high-performance [...] the so-called new x86 core really is a high performance", so I think it's pretty clear we're talking about top end desktop CPUs. They will not be using the FX brand for CPUs with cat core performance.

No, it's not clear, especially with them calling FX chips crippled Kaveri APU, or the laughable FX4100...

As I said, until AMD defines what they define as "high performance", AMD's "high performance" quote is meaningless. It could mean "higher than A57", "higher than Cat", or "Power8 performance levels".~

 

[Fjodor2001]

Right, who knows what they mean by that. It could just be two APU's with the graphics disabled stuck under on IHS for an eight core 'high performance' part within AMD's range of products. It could really be anything AMD considers a step up from the standard APU's they sell.

Well not two of AMD's current APUs though, since AMD said it will be using a new non-Bulldozer based x86 uArch. So if they have 8 such next-gen cores, it can be fast. Who cares whether they build an FX CPU using two APUs with the iGPUs disabled? The Haswell-E is also more or less just a normal Haswell with additional CPU cores but without iGPU (yes, I know there are some differences like cache size and so on, but still).

 

[NTMBK]

No, it's not clear, especially with them calling FX chips crippled Kaveri APU, or the laughable FX4100...

As I said, until AMD defines what they define as "high performance", AMD's "high performance" quote is meaningless. It could mean "higher than A57", "higher than Cat", or "Power8 performance levels".~

I would expect performance equivalent to the K12 core, given that they are supposedly closely related designs in uncore and platform, and the K12 is probably targeted to be significantly faster than A57 if the R&D costs are to be justified.

 

[pTmdfx]

AMD said "FX Series of high-performance [...] the so-called new x86 core really is a high performance", so I think it's pretty clear we're talking about top end desktop CPUs. They will not be using the FX brand for CPUs with cat core performance.

So they have to be quite a lot faster than AMD's current top end desktop CPUs, which ought to put them in the same range as current Intel desktop CPUs. The question is how they can achieve that if they will not have access to a high clockspeed process node?

Just as a note: The interviewee i.e. the AMDer just meant there is certainly an update of FX in the next two years. The bold-faced part is the commentary from the editor of that report, in the form of a link to the "w00t! AMD gets a new x86 core in 2016" article.

 

[pTmdfx]

we have just marketing at work here.

AMD calls, IIRC, both Jaguar and K12 "high-performance, low-power core", while Bulldozer got just a single high-performance tag. Yep.

 

[AtenRa]

The major question is, what are they targeting for with the new mArch.

 

[pTmdfx]

I would expect performance equivalent to the K12 core, given that they are supposedly closely related designs in uncore and platform, and the K12 is probably targeted to be significantly faster than A57 if the R&D costs are to be justified.

If the bus doesn't care what your core is speaking, it probably doesn't care how smart, handsome and fast the core is either. But anyway, while you may expect they are performance-wise equivalent, it is more likely that we will see the x86 part's presence in chips with higher TDPs. Say PCs, gaming rigs, workstations and of course general purpose servers (if they want steal more gold from Intel with APUs and ... CPUs). There are friends that love ARM from AMD to present in these spaces. But honestly speaking, if you can choose, would you prefer a thing that has no presence to what these spaces are reluctant to?

 

[mrmt]

I would expect performance equivalent to the K12 core, given that they are supposedly closely related designs in uncore and platform, and the K12 is probably targeted to be significantly faster than A57 if the R&D costs are to be justified.

Maybe not faster, but more efficient or even more scalable than vanilla A57.

 

[Ajay]

That article is far to vague to figure out what is really going on. There are allot of comments from Xbitlabs and editorial content that is not distinguished from the rumor that AMD is developing a new x86 core. While it seems obvious that a replacement for the BD line is necessary, and there are additional leaks about a new desktop x86 core being developed - I think we need a more definitive statement from AMD itself.

 

[Blitzvogel]

No, it's not clear, especially with them calling FX chips crippled Kaveri APU, or the laughable FX4100...

As I said, until AMD defines what they define as "high performance", AMD's "high performance" quote is meaningless. It could mean "higher than A57", "higher than Cat", or "Power8 performance levels".~

I think designing a core that can match one of desktop Kaveri's modules at 3/4 the clock rate and TDW would be a good initial goal.

 

[mrmt]

I think designing a core that can match one of desktop Kaveri's modules at 3/4 the clock rate and TDW would be a good initial goal.

Is it? Without knowing what kind of markets AMD wants to target, we cannot really say that.

 

[NostaSeronx]

I think AMD will be going for THz. AMD will definitely be first to THz. 1 THz K12, yep seems plausible.

 

[Blitzvogel]

Is it? Without knowing what kind of markets AMD wants to target, we cannot really say that.

Well I was speaking specifically of the desktop market, but I completely understand the necessity of your question as I was thinking about it before I even posted that. I was unwilling to even fathom it, but I'll try and make an attempt at it. I'm no EE or business major.........

On one hand you can go for the general notebook market first because it's still a highly viable market that represents plenty of share, even after it shrinks to the natural level that allows it to coexist with tablets. This means targeting good performance per watt. Like Intel is doing now, your technology can "trickle up" into the desktop and server market where mobile cores can take advantage of processes that allow for higher clocks, more cores, and perhaps more cache. AMD still has some market there and I think represents AMD's best chance at gaining any kind of relevance back because the products are in good consumer demand.

Originally I was thinking that since this is a "high performance" core, maybe AMD should target the server market first because they might get away with higher TDWs if they can nail the performance and hardware price aspects. Tech could be trickled down into the desktop market without too much issue, but they might hit a block with mobile where a difference in process might be necessary to mitigate TDW and might increase cost.

Perhaps like with Kaveri, the desktop market may make the ideal first target because of enthusiasts and custom builders, not to mention the plethora of cheap OEMs that attract low spending consumers. Even if this market is shrinking in most regions, there are lower income world regions where desktops are easily marketable. Moving a desktop core into servers wouldn't be out of the question. Moving into mobile may be a problem for the usual reasons.

Too many questions to ask. Many ways for AMD to go about this. In my opinion, mobile should be AMD's first target because of the "trickle up" benefit. It's worked well for Intel even if to the disdain of desktop Intel enthusiasts. Disregarding not being an EE or programmer, I would think AMD could push HSA more and perhaps more tightly integrate GPU CUs into the x86 cores themselves. Pushing for HSA adoption isn't a bad strategy either. Maybe each core could get a CU instead of an FPU?

TL;DR version: Go notebook/mobile first. Same cpu performance of mobile Kaveri at 3/4 the Hz and 1/2 the TDW would be commendable.

On a side note, I think AMD really missed an opportunity in not releasing Krishna.

 

[moonbogg]

Keller knows how to make a CPU and he has the balls to go through with it, based on what he did for AMD in the past. I wouldn't be surprised if AMD had some good performing chips in a couple years. That would be sick if everyone wanted an FX rig in 2 or 3 years.

 

[inf64]

It seems as if we will have more of a "traditional" new x86 core and that AMD will distance itself from CMT approach (my feeling). The chip will probably get the best parts of BD(EX) and K10 mixed in for a mid-3GHz freq. target and high(ish) IPC (I expect ~IB level, so close to Skylake but still some ~15-20% lower). Of course AVX3 should be there with full 512bit execution units(SIMD), just like Skylake.

Of course it can go the other way and end up like a turd, who knows. With new people on board and "brand new" stitched next to x86 I'd rather think they will make something mean and lean, like K8 was.

 

[Sweepr]

I expect ~IB level, so close to Skylake but still some ~15-20% lower). Of course AVX3 should be there with full 512bit execution units(SIMD), just like Skylake.

So somehow you expect them to jump from slightly lower than Phenom II IPC levels to IB IPC in 2 years while Skylake barely improves Haswell's perf/clock by 5-10%. Sounds a little bit too optimistic on AMD's side. If history is any indication, I'd keep the expectations/predictions low this time.

 

[cbn]

It seems as if we will have more of a "traditional" new x86 core and that AMD will distance itself from CMT approach (my feeling). The chip will probably get the best parts of BD(EX) and K10 mixed in for a mid-3GHz freq. target and high(ish) IPC (I expect ~IB level, so close to Skylake but still some ~15-20% lower). Of course AVX3 should be there with full 512bit execution units(SIMD), just like Skylake.

Of course it can go the other way and end up like a turd, who knows. With new people on board and "brand new" stitched next to x86 I'd rather think they will make something mean and lean, like K8 was.

I hope AMD decides to target the new x86 squarely into the high performance niche, rather than the "jack of all trades, master of none" approach that was "construction cores". (So maybe AMD doesn't get into mobile with the new x86 big core......but at least they would have the chance to solidly win desktop and maybe HPC.)

......besides for mobile, there is always "cat cores", so really no reason (at least how I see it) to make another compromise x86 big core product IMO.

 

[Homeles]

Keller knows how to make a CPU and he has the balls to go through with it, based on what he did for AMD in the past. I wouldn't be surprised if AMD had some good performing chips in a couple years. That would be sick if everyone wanted an FX rig in 2 or 3 years.

I've always found it a little offensive to suggest that one guy can make an entire engineering team -- well educated guys -- look like a bunch of fools. The whole Jim Keller superstar thing just doesn't pass the common sense test.

I don't think it's beyond AMD to turn things around, but it's ridiculous to give the credit to one guy, and undoubtedly that's what the AMD-enthused will do.

 

[mrmt]

Originally I was thinking that since this is a "high performance" core, maybe AMD should target the server market first because they might get away with higher TDWs if they can nail the performance and hardware price aspects. Tech could be trickled down into the desktop market without too much issue, but they might hit a block with mobile where a difference in process might be necessary to mitigate TDW and might increase cost.

TDP is not the issue on the server market, if it were IBM wouldn't make millions selling 250W chips. Performance per watt is the issue because it kills TCO, and this is where things go down hill for AMD. Unless this new chip of them can offer significantly better performance/watt, you can pretty much write AMD off on both mobile *and* x86 server market.

 

[frozentundra123456]

I've always found it a little offensive to suggest that one guy can make an entire engineering team -- well educated guys -- look like a bunch of fools. The whole Jim Keller superstar thing just doesn't pass the common sense test.

I don't think it's beyond AMD to turn things around, but it's ridiculous to give the credit to one guy, and undoubtedly that's what the AMD-enthused will do.

I agree with you to a certain extent. I think it was perhaps also easier in the earlier days of cpu design for an especially sharp engineer to make a quantum leap difference than it is now with the increasingly complex cpus and more complex designs, requiring I would assume hundreds of engineers and huge resources to design and test a cpu, especially a totally new architecture. OTOH, there are certain critical decision points in the design process where an exceptionally sharp engineer could be instrumental in following the right path. But I do think the idea that one man is going to suddenly magically design a killer chip to be a bit unrealistic, especially considering AMD will be working with far, far less resources and inferior process nodes.

Edit: Really though, this entire thread just speculation. Even AMD has not said what the design will be like or what applications it will be used for. TBH, if they are trying to design a new big core x86 cpu to compete with intel in the high end, it doesnt make sense to me to spend their limited resources trying to compete in a shrinking market (desktop). The only way it makes sense to me is if they feel they can get a strong foothold in the server market. Not impossible, but seems like a daunting task with intel and IBM entrenched in the high end and ARM attacking the low end.

 

[Idontcare]

I've always found it a little offensive to suggest that one guy can make an entire engineering team -- well educated guys -- look like a bunch of fools. The whole Jim Keller superstar thing just doesn't pass the common sense test.

I don't think it's beyond AMD to turn things around, but it's ridiculous to give the credit to one guy, and undoubtedly that's what the AMD-enthused will do.

Yep. And at his level in the org-chart one would hope he isn't tasking himself with getting bogged down in the design trenches either.

At his level, which is all project management (risk, goals, timeline), the only "true genius" he should be bringing to the table is in terms of keeping his legions of teams all on target to tape-out on schedule.

At no point should any aspect of the actual architecture or circuitry be handled personally by Keller. If he is doing that then he shouldn't be in the position he is in; rather, if that is the case then he should be in a lower position where he gets to touch the design while reporting in to the head guy who would be "merely" managing the project.

 

[Blitzvogel]

TDP is not the issue on the server market, if it were IBM wouldn't make millions selling 250W chips. Performance per watt is the issue because it kills TCO, and this is where things go down hill for AMD. Unless this new chip of them can offer significantly better performance/watt, you can pretty much write AMD off on both mobile *and* x86 server market.

That's why I mentioned getting away with high TDW/TDP (not sure if there is a real difference here, AFAIK W = P) if the performance is high enough. Sorry for not clarifying.

 

[Arachnotronic]

AMD is operating at less than 1/10th of Intel's R&D budget, is starting from a categorically worse base than Intel is starting from for its future CPU architectures, and is at a process node disadvantage.

By 2016 (probably late 2016) when K12 and x86 counterpart hit, Intel will be shipping Cannonlake (10nm Skylake). K12 will be on Samsung/TSMC's 14/16nm process putting it at a pretty significant disadvantage to whatever Intel is shipping if they are targeting the same space.

I would seriously temper my expectations for AMD's future big core. It seems to me right now AMD is hyping 2016 because it will not have a particularly competitive 2015 and it needs to get investors focused on the future.

Think about it. This is the NVIDIA Project Denver strategy...high performance ARM core...bestest thing ever...etc. you know the score