AMD Kaveri OC On Planet Neptune

[NaroonGTX]

I think it's possible that Carrizo will be 20nm, but as usual AMD is at the mercy of GloFo and whether or not they screw up again. Not really sure why Kaveri got delayed in the first place (from Steamroller 1.0 to 2.0), but is it plausible that it was due to GloFo having issues with their nodes?

 

[zir_blazer]

What is all the fuzz regarding "Steamroller B"? I saw that in some Kaveri slides, but nothing regarding what changed from original Steamroller to that B version or if Kaveri delays are related to it.

 

[inf64]

Not much changed from A to B and delays are not due to this but due to process node.

 

[NaroonGTX]

Well Steamroller B was known about ever since at least late 2012/early 2013: Source

Even then, no one (outside of AMD themselves) have any idea what makes SR-B any different from the original SR. My guess is that SR-A was canceled -- it was the version that was gonna have up to three modules and GDDR5 support.

Some people have speculated that the leaked die shot was SR-B, however I don't believe this.

 

[mrmt]

It certainly would have been a bigger jump if Kaveri was being produced on a very high performance process (28nm SOI/FD-SOI w/hidden gate). But, it's looking like 28nm Bulk is going to give the stock clocks a considerable haircut - hence the big mystery in terms of actual performance gains. On the face of it, it looks like AMD made some solid improvement to Kaveri B, but their sub par foundry partner may eat up much of AMD's gains.

The haircut has to be expected, no? Kaveri has a bigger iGPU and more units on the CPU, all that with just a half node to offset the extra power consumption. While the sub par foundry can get the blame for a lot of things, I think they aren't to blame on this one, SOI or no SOI.

 

[inf64]

The haircut has to be expected, no? Kaveri has a bigger iGPU and more units on the CPU, all that with just a half node to offset the extra power consumption. While the sub par foundry can get the blame for a lot of things, I think they aren't to blame on this one, SOI or no SOI.

You are wrong... The node is pathetic when compared to 32nm SOI that was used for PD parts. This will be clear when it gets reviewed.

 

[mrmt]

You are wrong... The node is pathetic when compared to 32nm SOI that was used for PD parts. This will be clear when it gets reviewed.

You mean it won't clock as high as the FDSOI node, or leakage will be worse even at idle? What about density, will they get some benefits on this area?

 

[inf64]

Density is good, everything else not so great.

 

[NaroonGTX]

By nature Kaveri on 28nm Bulk won't clock as high as Piledriver via 32nm SOI. Top-end Kaveri is 3.7ghz base with 4.0ghz turbo. Not really a huge downgrade compared to Trinity (3.8ghz base and 4.2ghz turbo), but a decent gap compared to Richland (4.1ghz base, 4.4ghz turbo)

The thing I'm most concerned about is what the typical max OC will be on air/liquid.

 

[schmuckley]

What I'm most concerned with is IPC and GFLOPS.
..just sayin..
Y U no L3 cache?

 

[NaroonGTX]

None of the APU's have had L3 cache, and AMD would need to fix their slow L3 cache performance before wasting more die space with it.

 

[mrmt]

Density is good, everything else not so great.

Makes sense... if 28nm is not good on leakage, then it's a goner in the mobile market. That would explain the shift to desktops. If Steamroller and Carrizo are both on a node that leaks like hell, they have no chance on mobile.

 

[inf64]

Makes sense... if 28nm is not good on leakage, then it's a goner in the mobile market. That would explain the shift to desktops. If Steamroller and Carrizo are both on a node that leaks like hell, they have no chance on mobile.

Good news is they are actively working on the issues for a while now. Maybe if we are lucky we get a better stepping at launch (but when is hard launch anyway?). Carrizo would be better off if it was on 20nm node since this one (28nm bulk) is not a good match for a product like Kaveri or Carrizo. Kaveri , if it was to be done on 32nm SOI , would be extremely potent product with somewhat bigger die than Trinity.

 

[NaroonGTX]

Kaveri initial (or "paper") launch = Jan. 14th, 2014
Retail/e-tail availability = c. Feb. 2014

 

[mrmt]

Good news is they are actively working on the issues for a while now. Maybe if we are lucky we get a better stepping at launch (but when is hard launch anyway?). Carrizo would be better off if it was on 20nm node since this one (28nm bulk) is not a good match for a product like Kaveri or Carrizo. Kaveri , if it was to be done on 32nm SOI , would be extremely potent product with somewhat bigger die than Trinity.

I wouldn't like to be in the shoes of AMD management team. Trinity had already very poor economics, and while 28nm should allow smaller dies, it is also destroying part of the value proposition of the line up, which should be a no go. It's a lose-lose situation, and they seem to have cut losses and moved to bulk, which should have been the lesser evil.

Another reason for AMD to be using 28nm bulk in that state is that they shouldn't have enough strength to push a node only for themselves anymore. 32nm was only viable because AMD purchase commitment was north of 1.5 billion. 20nm should push this number even further, maybe 2 billion worth in orders. As AMD is buying only 1 billion, they have to pick what Globalfoundries has to offer, and 28nm bulk it is.

Ed: If you are correct, and 32nm is indeed better than 28nm in terms of electrical parameters, another scenario would be AMD being forced to move to 28nm because by Globalfoundries' hand. GLF might have better costs with 28nm and thus forced AMD to move to 28nm and compensating AMD with a smaller WSA commitment.

 

[Ajay]

The haircut has to be expected, no? Kaveri has a bigger iGPU and more units on the CPU, all that with just a half node to offset the extra power consumption. While the sub par foundry can get the blame for a lot of things, I think they aren't to blame on this one, SOI or no SOI.

Well, some of us were concerned about this before hand and Homeles sums this up well. As far as AMD's 'blame' is concerned, due to the fact that they are poorly capitalized at the moment, they opted not to pay for extra R&D for their own 28nm node - which could have been PD/FD-SOI - for which they would have likely had higher clocks and less leakage. So, yes they hurt themselves competitively, as the cost of trying to be fiscally responsible - that must have been a tough decision. It must suck being R. Read right now.

 

[AtenRa]

Why im the only one that still believes Kaveri is on TSMC ???
And dont say about the WSA, they will keep producing FX/Opterons and Trinity/Richland at 32nm SOI at Glofo. Also current WSA ends in Q1 2014 and they will re-negotiate for Q2 2014 onwards.

 

[krumme]

Amd is, and will not be in a situation on the x86 market where exotic ibm soi solutions is for this kind of product. Wsa or not. Its a low cost solution for the low end consumer market not a server chip.

We dont know for sure how the process is on the apu. But the arm a7 variants made by gf is, as i remember, leaking like nothing else on this planet. It takes a really bad process to make that 0.45mm2 smal in order arm core to leak. Perhaps it have improved?...naaa

Tsmc hpm process could have gived a different product. But its all wsa related in the end. The problem is known and predictable. If the cost is low thats a huge advantage too. (It probably isnt lol)

 

[NaroonGTX]

Why im the only one that still believes Kaveri is on TSMC ???
And dont say about the WSA, they will keep producing FX/Opterons and Trinity/Richland at 32nm SOI at Glofo. Also current WSA ends in Q1 2014 and they will re-negotiate for Q2 2014 onwards.

From what I recall, the engineering sample screenshot of a Kaveri chip said "diffused in dresden" I believe, which points to GloFo. I'll post a pic later when I get back home.

 

[SiliconWars]

28nm bulk was AMD's decision. Way back when it was first discussed nearly 2 years ago everybody said that it would be a step down from 32nm SOI.

It doesn't seem THAT bad though, if 3.7GHz/4GHz turbo is right. That's only 10% slower than 32nm SOI after 2 years. I don't really understand how it can be so much worse at the same 95W TDP? Maybe temps are off the scale but that would just be the same as what happened to Intel at 22nm.

I've been expecting nothing much from Kaveri from the start so I'm not going to be disappointed either way. There's only so much you can do with a terrible base architecture.

 

[PPB]

Remember guys you are comparing a 28nm that it is in it's early stages vs a 32nm process that has had it's time to mature. From the go we knew the clock subject could be a problem considering it was bulk, but it was the wiser choice considering making it in 32nm would have ment porting GCN to a bigger node with different capabilities.

I would preferably had had Kaveri on TSMC @ 28nm a while ago, and move the production volume to GF when they had their node ready (ala console semi custom chip in the future), but hey, if it was my decision, I would have chosen to make bigger dies to cover the WSA agreement easier and compensate AMD's deficits in IPC and process with more transistors till they could renegotiate WSA to better terms for them (for them, I mean, GF must be having a blast having a customer forced to buy them or pay them anyways).

 

[erunion]

Bigger dies wouldn't help AMD, they are TDP limited. They would have to drop clocks in order to add more transistors.

 

[NTMBK]

Kaveri , if it was to be done on 32nm SOI , would be extremely potent product with somewhat bigger die than Trinity.

Any evidence for that?

 

[krumme]

Bigger dies wouldn't help AMD, they are TDP limited. They would have to drop clocks in order to add more transistors.

They are bandwith limited for performance anyway. The office perf is going to be plenty for the consumer market but gaming is bandwith restrained anyway. In performance perspective the low perf of 28nm bulk is therefore not a huge issue imho.

What is an issue is the potential leaking making this a less attractive products for mobile. But its not compettitive with hw here anyway. So if the 28nm bulk is dirt cheap its the best option anyway. There is a solid room for this product on the b2c market.

What is a problem for amd is if the hugely beefed up excavator dont have the mem bandwith or process improvement to work at low power or low cost. Then they have an arch without use imho. Excavator on 28nm with ddr3 makes no sense from a business perspective. Its the end of x86 low/midrange for amd then.

 

[ShintaiDK]

They are bandwith limited for performance anyway. The office perf is going to be plenty for the consumer market but gaming is bandwith restrained anyway. In performance perspective the low perf of 28nm bulk is therefore not a huge issue imho.

What is an issue is the potential leaking making this a less attractive products for mobile. But its not compettitive with hw here anyway. So if the 28nm bulk is dirt cheap its the best option anyway. There is a solid room for this product on the b2c market.

What is a problem for amd is if the hugely beefed up excavator dont have the mem bandwith or process improvement to work at low power or low cost. Then they have an arch without use imho. Excavator on 28nm with ddr3 makes no sense from a business perspective. Its the end of x86 low/midrange for amd then.

Office wise, AMD is as attractive as they are in the mobile segment. Higher TDP and lower performance.

I would be surprised if there is any big core from AMD after Excavator.