Haswell vs Kaveri

Itanium's code density is unusually bad.
The extension-fest x86 has gone through over the years has significantly eaten into the code density advantages it had over RISC ISAs, so other factors must intrude.
As noted, Intel has a vastly superior cache subsystem outside of the Icache, while AMD's is in some ways cringeworthy and ARM implementations almost as a rule are as cut-rate as can be done.
The biggest thing the leading ARM architectural licensees do besides customizing the cores is putting in a competent memory subsystem.

Being able to compensate for one architectural point by updating everything else around it has its benefits, and it's not like an 8-way non-shared L1 can't do better at times than whatever was attached to Bulldozer.
 
Just read hardware.fr review and it made clear that owners of phenom II have no reason to upgrade.
I'm really disappointed, without gddr5m kaveri is a pointless product. It would not surprise me in the slightest if HIV favors jaguar ant its upcoming rev in laptops.
The bothering part is that the gain in IPC are below what AMD announced and at the time they had to make compromise (vs piledriver) to get there.
Something I simply don't understand is why AMD went for such a big chip, they made it tough for themselves to fight on price, extra transistors don't pay for them selves through an increase of performances.
I listened to more knowledgeable members a couples of times when I stated that that CMT was a dead end but now that is going to be a hard sell, it looks more and more to me like AMD lost 4 years and we are still counting
 
Was GDDR5 ever truly considered for Kaveri?
Using GDDR5 would be alien to the motherboard industry. Using quad-channel DDR3 would not.


It seems to me that the low-power Kaveris will be the real threat to Intel's ultrabook line.
That A8 in 45W mode is the only inspiring factor I've seen so far.

Can we hope for AMD to release FM3 boards with quad-channel memory?
 
AMD should simply rework the Jaguar architecture for more scalability upwards (laptop, desktop, etc.) with relaxed pipeline to allow for higher clockrates and full-speed L2. I would rather see eight pumped Jaguar cores in Kaveri than yet another botched attempt to iterate the poor Bulldozer. This is not suitable for an APU, just leave it for the high-end workstation and server systems.
 
So far I've seen no serious hint of quad channel, only marketing produced die shot.
As for being alien, their position is alien now, they've close to double the transistor budget to push a chip that perform barely better than their previous product.
Apu with such an investment on the GPU are aliens on their own they need matching alien memory.
I would think that hardware.fr is on to something when they stated that epilda bankeroute threw a wench in AMD plans,either way they are scared of hurting so their GPU sales. In both case spending that much silicon on chips that sells for cheap doesn't sound clever to me.
The issue is that AMD is as passive as it gets, looking at what nvidia does on the mobile market should make their management react, it does not.
 
AMD should simply rework the Jaguar architecture for more scalability upwards (laptop, desktop, etc.) with relaxed pipeline to allow for higher clockrates and full-speed L2. I would rather see eight pumped Jaguar cores in Kaveri than yet another botched attempt to iterate the poor Bulldozer. This is not suitable for an APU, just leave it for the high-end workstation and server systems.

I'm not sure improved jaguar would cut it either, they might not work in that direction anyway as power consumption has to be their primary concern.
As for BD in the high end, it is even less of a fit, a core I 3 kicks Kaveri's butt, they can't compete with core i5 let alone high end Intel chips :(
 
AMD planned the chips to work both with ddr and gddr?
Anyway quad channel is pretty alien and comes with overhead on its own (/ which would translate in Dollars on the mobo price and Dollars again in the form of extra memory so-dimm).
Anyway the whole point is that Kaveri is as we know it and that is a 240mm^2 chip, two third of th7e PS4 chip, that performs poorly in comparison.
I see no saving grace for the chip, sales are going to be ugly that is my bet.
EDIT
One thing is that I think AMD does not really want to byte into its HDX7xx (not sure of the new naming) still they invest dearly in the GPU of their APU only for them to underperform /death by bandwidth starvation. If they think it can hurt them as a whole why even try to pack beefy GPU in their APU? Really they no longer make sense to me.
 
I find it curious they specifically mention GDDR5 not being supported - could it be that the 2 disabled channels are in fact for GDDR5?
Or what other explanations are there for specifically mentioning it?
 
I find it curious they specifically mention GDDR5 not being supported - could it be that the 2 disabled channels are in fact for GDDR5?
Or what other explanations are there for specifically mentioning it?
prevent bugs? No matter the reason why those controller are disabled.
 
I'm not sure improved jaguar would cut it either, they might not work in that direction anyway as power consumption has to be their primary concern.
As for BD in the high end, it is even less of a fit, a core I 3 kicks Kaveri's butt, they can't compete with core i5 let alone high end Intel chips :(
Looking at the situation, it would be much easier for AMD to scale up their "native" low-power architecture, than to try and patch over Bulldozer that was really never meant for such tight envelope. They made too many compromises with the Steamroller variation in Kaveri, to fit the poor thing on a tight diet. It is evident that AMD doesn't have suitable singular architecture to span a wide market range, like what Intel have developed since Nehalem.
 
The bothering part is that the gain in IPC are below what AMD announced and at the time they had to make compromise (vs piledriver) to get there.

Actually the IPC is improved (my guess is that you meant per-thread overall performance, not IPC). AMD mostly failed to deliver the clockspeeds needed to make this revision competitive.

And I think this lower-than-expected frequencies (or higher power consumption, if you want) had been the mantra of the BD architecture all along.

.
I listened to more knowledgeable members a couples of times when I stated that that CMT was a dead end but now that is going to be a hard sell, it looks more and more to me like AMD lost 4 years and we are still counting

Why you blame the failure of this CMT implementation on CMT itself?
 
Anyway the whole point is that Kaveri is as we know it and that is a 240mm^2 chip, two third of th7e PS4 chip, that performs poorly in comparison.

Not on the CPU side which is arguably more important in a PC APU where graphics are "good enough".

On the graphics side *if* it weren't bandwidth limited it'd be a little over half the performance of the XBO APU and probably around 2/5 of the PS4. But on the CPU side it's arguably faster than both.
 
I find it curious they specifically mention GDDR5 not being supported - could it be that the 2 disabled channels are in fact for GDDR5?
Or what other explanations are there for specifically mentioning it?

Considering that the PHYs look identical, and that AMD's low-end GPUs that support both DDR3 and GDDR5 do so using the same PHYs (to my knowledge, and I could be wrong) I'd say that's unlikely.

Perhaps AMD originally wanted 4 memory channels that could all support either DDR3 or GDDR5m, allowing for a pretty broad range of possible configurations, but somewhere along the line, it all went tits up.
 
Actually the IPC is improved (my guess is that you meant per-thread overall performance, not IPC). AMD mostly failed to deliver the clockspeeds needed to make this revision competitive.
Hardware.fr has a clock per clock comparison, IPC indeed improves but results are below their promises, actually those old phenom II might still be better.
And I think this lower-than-expected frequencies (or higher power consumption, if you want) had been the mantra of the BD architecture all along.
It would take insane clock speed to compete with Intel, actually oit would have taken crazy speed for BD 1 to compete in IPC with Phenom. Goin


Why you blame the failure of this CMT implementation on CMT itself?[/QUOTE]
Well after years and three revisions I think it might be crystal for those in the industry, it is not working as intended /could put AMD out of business. I would bet that nobody will ever try again after such a debacle.
Not on the CPU side which is arguably more important in a PC APU where graphics are "good enough".
Too bad it is up against INtel offering, gaming should be the bright point /selling point as it is with the ps4, it is not. overall is it a waste of silicon, 2 CU 4 Rops is enough indeed (kabini style), they would be better competing on price harder because now the threat is not Core i3 and lesser graphic performance but Pentium +whatever cheap discrete GPU (more than often Nvidia).
On the graphics side *if* it weren't bandwidth limited it'd be a little over half the performance of the XBO APU and probably around 2/5 of the PS4. But on the CPU side it's arguably faster than both.
Hopefully the CPU is faster as I guess it burns way more power. The issue remain the selling point is not delivered, still not enough for comfortable gaming / death by bandwidth starvation.

Back to Entity279 point, another thing with CMT is that it scales poorly from a production pov, you move by increment of 2 cores /1 module with the matching amount of L2.
There is no rule stating that the number of cores has to be an even number, those Phenom X3 were nice. AMD can't compete on perfs, though if they have std cores (single threaded) with good AVX 1 support I think that say 3 cores +2 CUs would be interesting option vs INtel pentium, and they would would not be 240mm^2 set up to sell ultimately in (forced) salvaged form below 100$.
 
Looking at the situation, it would be much easier for AMD to scale up their "native" low-power architecture, than to try and patch over Bulldozer that was really never meant for such tight envelope. They made too many compromises with the Steamroller variation in Kaveri, to fit the poor thing on a tight diet. It is evident that AMD doesn't have suitable singular architecture to span a wide market range, like what Intel have developed since Nehalem.
I don't know if scaling up is the way to go but indeed they should focus on "traditional" core.

An issue I see is that I don't think that AMD has still what it takes to support and develop multiple architecture. I would wish that they stick to one, though as they are moving ARM the minimal number of architecture seems to be 2 till unforeseen future.
I'm not sure it is their best bet they would be better focusing on either ARM or X86.
Which one is the most promising that is OT, more a fit to the business thread.
X86 may look safer but actually I think that they should focus on ARM. THey can no longer compete with Intel. I start to seriously wonder about how much life is left in the X86/windows combo in the personal realm. Overall putting my self in the driver seat I would do a bold move cut the X86 line and focus on GPU and ARM SoC. Ultimately I could see X86 ending like PPC aka making lot of money but as a professional product. On that segment AMD can't compete with Intel, its ISA, Its financial power, its foundries, etc.
In the realm world the situation is more open, with a proper business plan they could have quite some success while making money in the discrete GPU market as long as it exists.
Say AMD starts now how till it have a compliant solution in the X86 market? On the other hand Google is broadening the reach of Android now we will soon know if "non mobile" Android device are doomed to suffer the same fate as PC in their many form or if there is actually "something" in the PC that most constumers are "done for".
 
I haven't followed the whole Kaveri story, but it seems to me that AMD was making a big deal of this whole HSA thing, which I understand as a CPU and GPU living harmoniously under the same roof, sharing the bread etc. Intuitively, that sounds like something that can have a lot of benefits, but are there any tests right now that confirm this?

Or do we have to wait for developers to write software that specifically uses this feature? (Which, for a low end offering like this, doesn't seem all the likely?)
 
Back
Top