Could AMD's Jaguar pull a Dothan?

[Chiropteran]

It's like history is repeating itself.

In 2000, Pentium 4 was released and very lackluster, much like Bulldozer. Vishera fixed a few issues and was more competitive, just like the Northwood generation of Pentium 4. After Northwood, Prescott was almost a step back... hopefully Steamroller doesn't follow in that pattern.

But all that time, Intel had a backup plan, which it eventually implemented when they build the Intel Core Duo and Core Solo CPU, based on the mobile CPU Dothan. This led to Core 2, and Nehalem, and basically gave Intel a huge lead which AMD still hasn't caught up with.

Could AMD do the same thing, and release new high-end desktop CPU based on the next generation of Jaguar? Could such a CPU compete with Intel on both performance and power usage?

 

[mrmt]

Brazos was everything AMD needed in the past: Efficient, small, balanced. Jaguar is a good improvement. It would be nice to see AMD scaling up Jaguar to Bulldozer performance levels.

But I'm not sure if this is doable when AMD is going full steam ahead with ARM. There they need to compete against Qualcomm, Calxeda, Samsung and others. I don't think they could muster all the resources to stay on par with these companies on ARM, take on Intel on x86 and stay on top in graphics.

 

[Arachnotronic]

Could AMD do the same thing, and release new high-end desktop CPU based on the next generation of Jaguar? Could such a CPU compete with Intel on both performance and power usage?

Not likely; Intel has a large (and growing) process lead. In 1H 2014, you will see Broadwell based Ultrabooks at crazy low TDPs with even better power management at 14nm while AMD will still be just launching its 28nm part. Today's all-in-one oriented 4770R blows away anything AMD will likely have on the CPU side of things, rails Richland/Trinity in GPU, and still does it with half the max power consumption.

This will NOT be pretty, and I think given AMD's anemic R&D budget, it would be foolish to think that as CPUs become harder to design that AMD will be able to "catch up". Do not mistake me for an AMD hater...I am merely a realist.

 

[NTMBK]

The issue is that in the process of scaling up Dothan to Conroe, Nehalem and Sandy Bridge, Intel lost that low-power edge in thin laptops- and have been desperately trying to claw it back since, with IB and HW. And this is with Intel's superior processes! I suspect that trying to scale Jaguar up in performance will kill its usefulness in its intended market- tablets, and embedded.

 

[zir_blazer]

I thinked the same some months ago. However, we still don't know what can happen if you put a Jaguar on a Desktop envelop. One of the things that hyped people up about Banias and Dothan, were some Desktop Socket 479 Motherboards and adapters to use Pentium M in Desktop glory, something that didn't happened with Jaguar.

 

[frozentundra123456]

The issue is that in the process of scaling up Dothan to Conroe, Nehalem and Sandy Bridge, Intel lost that low-power edge in thin laptops- and have been desperately trying to claw it back since, with IB and HW. And this is with Intel's superior processes! I suspect that trying to scale Jaguar up in performance will kill its usefulness in its intended market- tablets, and embedded.

Doesnt jaguar scale only to about 2ghz? I dont really see it as a competitor in the high end desktop segment.

Not really sure I would agree that Intel has lost the low power edge in laptops either, at least to AMD. AMDs mobile lead over intel is rather small in graphics, and they trail in cpu performance. Jaguar may change this in low end laptops, but really it seems to be more of a tablet chip than a laptop one. If intel has lost the low power segment, it is really to ARM, not AMD.

 

[ShintaiDK]

Jaguar is not even in the same class if you wish to draw on history. Its only a 2 issue wide uarch while Pentium-M was 3 issue wide.

And AMD already killed of the cats cores for servers.

 

[Enigmoid]

I don't think so. Jaguar is really not that good at higher power levels.

A6-5200 (kabini): 4 cores at 2.0 ghz, 2 MB cache, 128 GCN shaders at 600 mhz, single channel RAM.

A10-5745m (richland): 4 cores at 2.1-2.9 GHz, 4 MB cache, 384 VLIW4 shaders at 533-626 mhz, dual channel RAM, requires additional chipset.

Both are 25 watts tdp. Richland will require a few more watts because its not as integrated as kabini.

Richland looks that same as if not better than kabini; higher clocks, more cache, better gpu (dual channel ram) for a little more power.

 

[Ancalagon44]

I don't think so. Jaguar is really not that good at higher power levels.

A6-5200 (kabini): 4 cores at 2.0 ghz, 2 MB cache, 128 GCN shaders at 600 mhz, single channel RAM.

A10-5745m (richland): 4 cores at 2.1-2.9 GHz, 4 MB cache, 384 VLIW4 shaders at 533-626 mhz, dual channel RAM, requires additional chipset.

Both are 25 watts tdp. Richland will require a few more watts because its not as integrated as kabini.

Richland looks that same as if not better than kabini; higher clocks, more cache, better gpu (dual channel ram) for a little more power.

My guess is that if Jaguar were built on a different process, its power consumption characteristics would be different. Jaguar is built on a bulk process to save costs, I dont think its the best process as far as power efficiency goes.

I think they use Jaguar as a base for a high performance core, but it would practically have to be redesigned again, since it would require substantial changes.

They would have to dramatically widen its architecture and probably re-implement it on a different process in order to achieve decent clockspeeds. Maybe even have to add a pipeline stage or two.

 

[mrmt]

Both are 25 watts tdp. Richland will require a few more watts because its not as integrated as kabini.

Richland looks that same as if not better than kabini; higher clocks, more cache, better gpu (dual channel ram) for a little more power.

Is there a review comparing the two (or Trinity)?

I only saw reviews comparing it with IVB/SNB and Clovertrail.

 

[sefsefsefsef]

AMD's netbook cores will have a hard time scaling to high frequencies because they are laid out and routed largely using automatic tools (rather than by hand). It would take a lot of work to transform Jaguar into a suitable halo processor from that perspective, plus you'd have to widen it a lot to make its performance really competitive.

I think continuing to evolve the bulldozer core is the right move on the high end. It's only a few changes away from being great, IMO.

 

[Idontcare]

It's like history is repeating itself.

I agree. I've felt the parallels were striking and on multiple levels.

After Northwood, Prescott was almost a step back... hopefully Steamroller doesn't follow in that pattern.

I fear it will, and again for reasons that are like history repeating.

Intel's 90nm had issues with static leakage. It was their first node to include copper wiring in the BEOL and the transistors were aggressively scaled (physical-dimension wise) 130nm xtors which leaked like hell when they got hot and had the volts pumping through them (as needed to hit 3.8GHz).

What set Intel's 90nm apart from AMD's 90nm was AMD had SOI in their node and that made the K8's perform in a whole other league for performance/W.

Now look at AMD and what they are using for 28nm...it is their first node since that same era of the Prescott for which they won't have SOI to save their power-consumption bacon.

Combine that with the fact they won't have Finfet, it will be 32nm planar HKMG xtors on steroids (physically shrunk to get drive currents higher) and they will leak all the more than the current 32nm xtors do (once volted as needed to hit the clockspeeds steamroller will need to hit).

Unfortunately for AMD I see the parallels and history repetition continuing at 28nm because of the process node situation. It is going exactly against them, not in their favor at all, which is exactly what happened to Intel at 90nm.

(and don't be confused with Dothan at 90nm vs. Prescott, Dothan was intentionally designed to operate within a clockspeed envelope that prevented the operating temperatures and requisite voltages that led to the cascading effect that Prescott operated under...Dothan being a good product was because of its design, not because of 90nm being healthy)

But all that time, Intel had a backup plan, which it eventually implemented when they build the Intel Core Duo and Core Solo CPU, based on the mobile CPU Dothan. This led to Core 2, and Nehalem, and basically gave Intel a huge lead which AMD still hasn't caught up with.

Intel's backup plan was a two-pronged combination that entailed advancing the Dothan to 65nm (which they did with Prescott in the creation of Cedar Mill as well), but what made it work was they got their 65nm process under reasonable control in terms of leakage and voltages.

Compared to the clockspeeds of Dothan laptops, Intel took the clockspeeds of Core up quite a bit with 65nm without destroying the TDP bank, and they overclocked rather well too.

I only bring this part of history up because I also think it is relevant for AMD...they can't just rely on design and design alone to get them out of their current funk. It is going to take a technology partner who is capable of delivering a process-competitive node in terms of both electrical parametrics as well as release timeline.

Could AMD do the same thing, and release new high-end desktop CPU based on the next generation of Jaguar? Could such a CPU compete with Intel on both performance and power usage?

AMD could...if they weren't tied to mast of S.S. GloFo. And unfortunately for everyone (competition is always good), GloFo is a ship that seems to be piloted by Captain Peachfuzz.

According to the narration of Jet Fuel Formula, Peachfuzz was, from his youngest days, an incompetent sailor. As a child, even his toy boats sank. At the age of 18 he joined the navy. He was awarded numerous medals, all of which were donated by the enemy. Sailing the wrong way through the Panama Canal and becoming the only captain of an icebreaker in the South Seas earned him the nickname "Wrong Way" (an allusion to the American pilot Douglas "Wrong Way" Corrigan). After receiving a large inheritance from an aunt he purchased and took command of the S.S. Andalusia (called Athabaska in some episodes). His crew considered mutiny but decided rather to install a dummy control room, so that Peachfuzz would think he was in command, while the crew actually controlled the ship from another location. Unfortunately, Peachfuzz takes a wrong turn and winds up in the real control room.

If only AMD was free as a fabless company to negotiate and partner with leading edge foundries of their choosing then their fate would be in their hands. As it stands now, AMD is a subdivision of GloFo and they have no future other than whatever GloFo is willing to accord them.

 

[Arachnotronic]

IDC,

TSMC's process nodes are still behind what Intel offers...not sure if "freedom" in the process space would help 'em any.

 

[insertcarehere]

IDC,

TSMC's process nodes are still behind what Intel offers...not sure if "freedom" in the process space would help 'em any.

At the very least TSMC's process nodes seem to arrive in a reasonable schedule, it's Q2 2013 now and Glo-Fo still haven't released a 28nm process node yet.

 

[Chiropteran]

Jaguar is not even in the same class if you wish to draw on history. Its only a 2 issue wide uarch while Pentium-M was 3 issue wide.

And AMD already killed of the cats cores for servers.

Pentium M was a single core, and didn't include a GPU. A lot of things have changed in the last ten years and priorities aren't the same. That is, during the time of the Pentium M it *was* the low end mobile CPU, because Atom and such didn't exist.

 

[Arachnotronic]

At the very least TSMC's process nodes seem to arrive in a reasonable schedule, it's Q2 2013 now and Glo-Fo still haven't released a 28nm process node yet.

Heh, TSMC isn't that much better. Their 28nm HPm still isn't out, according to their latest conference call, 20nm doesn't even begin to show up for revenue until Q2 2014 (which really means you MIGHT see some 20nm Snapdragons/Tegra in 2H 2014).

They're making a lot of GloFo-like BS claims about 16nm, but even at their own CCs they say that the volumes won't be anything remotely resembling high volume until 2016.

 

[SlowSpyder]

With Dothan Intel had a chip that already had performance in the ballpark of AMD and their own desktop processors, sometimes even faster depending on the task. AMD would have a much more uphill battle with Jaguar, I think. Jaguar would probably take a lot of reworking to get the clocks and/or IPC up to where it would need to be to compete.

 

[mrmt]

Heh, TSMC isn't that much better. Their 28nm HPm still isn't out, according to their latest conference call, 20nm doesn't even begin to show up for revenue until Q2 2014 (which really means you MIGHT see some 20nm Snapdragons/Tegra in 2H 2014).

TSMC has a functional 28nm node to talk about, which is more than what we can say about GLF.

 

[Arachnotronic]

TSMC has a functional 28nm node to talk about, which is more than what we can say about GLF.

Sure. TSMC was fine until they started BSing about 16nm.

 

[Exophase]

Although Intel shifted the P6 line to mobile after releasing Netburst it was at its heart always a good high performance uarch and stayed that way at every iteration. You could literally take a Pentium M and put it in an adapter for us on a desktop board and have respectable performance. Some people did this.

Jaguar is no where close to that. It's clock limited by design (and it's not something like lack of SOI holding it back), it simply won't scale that much higher than they're currently scaling it. And its typical IPC still tends to be behind K10. It's good if you need to shove a lot of cores into a relatively tight power budget (see consoles) but the single threaded performance potential is just not there.

The real parallel would have been if AMD kept working on K10 beyond what they did for Llano, but they hadn't.

On the side-topic of GF 28nm, everyone keeps ignoring Rockchip 3188 which has been in real products for months now. Just because AMD hasn't released something using it doesn't mean it's a bust, and it's not like a bleeding low margin operation like Rockchip can afford to waste a lot of money on awful yields (they probably went with GF because they priced lower than TSMC, but who knows). It may only be the low power variants of the process that work, or it may be there are just no takers ready to use the high power one. Nonetheless, the low power one qualifies as having released 28nm.

 

[USER8000]

Are we not jumping the gun regarding whether AMD is going to have issues with the GF 28NM process??

How about we wait and see how the CPUs pan out first??

 

[sm625]

Jaguar cant even outperform a celeron 887 found in bajillions of bargain bin notebooks. We're talking under $200 here, just due to the sheer volume of product. You will never find jaguar based products that cheap because no vendor is going to build that many. Sure, jaguar might use a little less power than a celeron 887, but its not enough to put it into a new class of product.

I dont know why this isnt available in america yet, but look at this thing:

http://www.techhypermart.com/notebo...-131-21292g06-11-6-notebook-laptop-touch.html We're going to be seeing systems like this for well under $500 by christmas. That's an ivy bridge ulv pentium, and that's what AMD is going to be competing against.

 

[AtenRa]

I don't think so. Jaguar is really not that good at higher power levels.

A6-5200 (kabini): 4 cores at 2.0 ghz, 2 MB cache, 128 GCN shaders at 600 mhz, single channel RAM.

A10-5745m (richland): 4 cores at 2.1-2.9 GHz, 4 MB cache, 384 VLIW4 shaders at 533-626 mhz, dual channel RAM, requires additional chipset.

Both are 25 watts tdp. Richland will require a few more watts because its not as integrated as kabini.

Richland looks that same as if not better than kabini; higher clocks, more cache, better gpu (dual channel ram) for a little more power.

Believe me that Quad Core Kabini at 2GHz is faster than Quad core Richland at 2GHz. First, Jaguar IPC is higher than Richland, secondly Jaguar Core scaling(MultiTread) is higher because Richland uses CMT. The only part that Richland wins is in iGPU due to the bigger size of it.

 

[Ventanni]

Is Jaguar's IPC really higher than that of Piledriver? That's a fairly large core size difference.

 

[Idontcare]

Are we not jumping the gun regarding whether AMD is going to have issues with the GF 28NM process??

How about we wait and see how the CPUs pan out first??

Being 2yrs or so behind TSMC, the issues sure had better be ironed out by now. Leaving nothing but smooth sailing ahead for AMD at 28nm.

The issue here is that even if we assume 28nm is trouble free (hitting clockspeeds and parametric yields) and is essentially mature at time-zero (delivering entitlement functional yields) the near 2yr delay relative to TSMC is a profit and marketshare killer for AMD.

There is a reason TSMC spends billions to beat their foundry competition to the market first, and there is a reason GloFo is spending billions to try and keep up with TSMC. Time is money for their customers, and money is the reason the customers are interested in doing business with the foundries.

IDC,

TSMC's process nodes are still behind what Intel offers...not sure if "freedom" in the process space would help 'em any.

I'm not worried about the TSMC to Intel gap, I'm more concerned with the fact that whether it is GPUs, APUs, x86 CPUs or ARM-based CPUS...everywhere AMD is competing they are competing with IDM's like Intel or fabless companies like Qualcomm or Nvidia who are free to choose a foundry that offers the best chance of maximizing their own market timing and product feature capabilities.

AMD is tied to the (at best) second-best foundry solution, which hamstrings their ability to compete with basically anyone who is able to get a foundry deal with TSMC or an IDM situation like Intel and Samsung.

Even with GPUs where they have access to TSMC, AMD is not getting to operate at the same margins as their competition because they pay TSMC for the wafers (as does Nvidia) and they pay GloFo for an exclusivity waiver so they can buy those wafers from TSMC (which Nvidia doesn't have to do).

There is a leach attached to AMD, sucking its lifeblood away, that no other company which AMD is competing against has to deal with. That is what is going to prevent Jaguar being AMD's Dothan because when Dothan was Intel's Dothan they (Intel) had no such leach in the equation.