Report: AMD Launching Trinity APUs Soon; Desktop in August

Page 2 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
Status
Not open for further replies.


I've seen a slide (purportedly) from AMD

It actually makes sense (in my convoluted thinking). Think of it as the AMD Teeck/Toke :lol:

It's a logical progression -- Llano is Stars 32nm cores with 40nm GPU. Trinity is 32nm Piledriver cores with 40nm GPU. Kaveri is 28nm GPU with 32nm CPU cores.

You can even take it a step further. Do you think some of the masks for Trinity Piledriver cores could be used for FX-Piledriver, too?

The thing with APU Dual-Graphics (new AMD term -- no more 'hybrid' graphics) is best performance is seen with synchronized cores, clocks and memory -- not to mention VLIW5/4/GCN. The 40nm GPU in Trinity is Turks-based ("HD 7660D") at 800MHz (384 shader units) up from 400 cores running at 600MHz. Generally, an up-clocked Llano Turks.

The cards that closest fit that profile are the HD6570 or HD6670. The HD6670 has GDDR5, so presumably the HD6570 with GDDR3 will be easiest to sync-up for dual-graphics -- into a "HD7790D2" (?-who knows at this point)

With the Kaveri APU (and dual-graphics), the next progression is 28nm Southern Islands. Best guesses will probably come from examining the Southern Islands-based Mobility Radeon 7XXXM chips for clues as to cores and clock speeds -- most likely the HD 7770M -- for the integrated Kaveri GPU.

Which (gasp!) in dual-graphics would seem to sync-up nicely from a spec standpoint with the (drum roll, please ...)

Radeon HD 7750 Cape Verde with GCN cores.




 
[citation][nom]Wisecracker[/nom]I've seen a slide (purportedly) from AMD It actually makes sense (in my convoluted thinking). Think of it as the AMD Teeck/Toke It's a logical progression -- Llano is Stars 32nm cores with 40nm GPU. Trinity is 32nm Piledriver cores with 40nm GPU. Kaveri is 28nm GPU with 32nm CPU cores.You can even take it a step further. Do you think some of the masks for Trinity Piledriver cores could be used for FX-Piledriver, too?The thing with APU Dual-Graphics (new AMD term -- no more 'hybrid' graphics) is best performance is seen with synchronized cores, clocks and memory -- not to mention VLIW5/4/GCN. The 40nm GPU in Trinity is Turks-based ("HD 7660D") at 800MHz (384 shader units) up from 400 cores running at 600MHz. Generally, an up-clocked Llano Turks.The cards that closest fit that profile are the HD6570 or HD6670. The HD6670 has GDDR5, so presumably the HD6570 with GDDR3 will be easiest to sync-up for dual-graphics -- into a "HD7790D2" (?-who knows at this point)With the Kaveri APU (and dual-graphics), the next progression is 28nm Southern Islands. Best guesses will probably come from examining the Southern Islands-based Mobility Radeon 7XXXM chips for clues as to cores and clock speeds -- most likely the HD 7770M -- for the integrated Kaveri GPU.Which (gasp!) in dual-graphics would seem to sync-up nicely from a spec standpoint with the (drum roll, please ...)Radeon HD 7750 Cape Verde with GCN cores.[/citation]

AMD should have just thrown in a modified 7750 Cape Verde GPU into the top Trinity APUs (or a new GCN GPU that is about as good as the 7750's cut down Cape Verde so that Trinity doesn't rely on supply of a second tier GPU), at least in my opinion. It has low enough power usage and it's performance is good too. However, I have to wonder... What is AMD going to do about the ever increasing RAM bandwidth bottleneck? Unless FM2 has extra pins for memory channels or for on-board GDDR5, Trinity could be even more bottle-necked than Llano (which is already pretty badly bottle-necked).

It only makes sense if AMD didn't want to use their 28nm GCN GPUs in the Trinity APUs and just leave them in the discrete cards for now. However, unless supply is too low, they should have, at least in my opinion. Better performance and far higher power efficiency.
 
It's simply too much for 'em, Blaze. (more than 'they can Chew*') Big swings easily end with big misses.

Their hope in advancing APUs and OpenCL compute are simple progressive steps. A lot of singles with an occasional home run.

First-run die shrinks have been foul balls

Next pitch they step back up to the plate and drop voltage 10+%, while upping performance and features. FX-Bulldozer has no where to go but up, right? 😀 FX-Piledriver is Trinity with L3 instead of the GPU, so step-by-step, AMD will bring them together on the same die (if the shared L3 simply adds value and performance in exchange for die space).

The GPU 40- to 28nm die-shrink integration with Piledriver cores looks to be a winner . Next up is Steamroller CPU cores with shrink 32- to 28nm.

Maybe we won't have to beta-test those for AMD :lol:



 
Just in time to grab from Intel all those business class customers that want to spend around $150 or less on decent CPU+IGP. Intel has good product in HD4000 but CPUs with it cost $250+. People who need IGP as cheaper solution than discrete card will find this price utterly and ridiculously high, while people who are looking for entry level gaming will find combos like Pentium G840/850/860 + Radeon HD 5750/6750/7750 outperforming it by quite a bit.
 
Call me mad, but I've never heard of Llano using a different process technology for its GPU to the Stars cores. I don't understand why it, nor Trinity, would be that way.
 
[citation][nom]vaughn2k[/nom]I hope AMD will remove itself from too much false hope. I really want to see an improve CPU/APU from them and not just another Bulldozer fiasco. We need them to compete with over-priced over-hyped CPUs from Intel.[/citation]

I'd like for them to compete with Intel but it seems to me that they have given up on doing such thing. So unless they go back tot he drawing boards and bring something that can seriously rival IB in all areas, I'm sticking with Intel.
 
[citation][nom]killerclick[/nom]AMD is so much behind Intel, it's ridiculous. And don't give me that AMD is 10 times smaller than Intel, ARM is 5 times smaller than AMD, yet they manage to dominate a huge market segment. Also, as big as Intel is, they aren't able to compete with nVidia and ATI/AMD in the GPU market, so AMD doesn't have any excuses. They really messed up with Bulldozer (and to a lesser extent with Stars), while Intel did everything right.[/citation]

AMD is behind because of Intel's anti-competitive practices back in the first Athlon era. If AMD had been allowed to compete, they would have sold twice to three times as many chips, and made enough profit to hire engineers to keep up the pace. Instead, they lost out on major sales because Intel threatened motherboard and system makers with reducing their supplies of 440BX chipsets if they made Athlon based systems. This would have been death for any motherboard maker. (Remember how the first Athlon motherboards came in unmarked, unbranded packaging and no support?) AMD was kept behind, and they're still feeling the effects. If they would have had enough engineers, they would have had the Athlon 64 out a year faster, at least, the Athlon II out two years faster, and the Bulldozer would have been out to compete with the Core 2, easily beating it.
 
[citation][nom]dgingeri[/nom]AMD is behind because of Intel's anti-competitive practices back in the first Athlon era. If AMD had been allowed to compete, they would have sold twice to three times as many chips, and made enough profit to hire engineers to keep up the pace. Instead, they lost out on major sales because Intel threatened motherboard and system makers with reducing their supplies of 440BX chipsets if they made Athlon based systems. This would have been death for any motherboard maker. (Remember how the first Athlon motherboards came in unmarked, unbranded packaging and no support?) AMD was kept behind, and they're still feeling the effects. If they would have had enough engineers, they would have had the Athlon 64 out a year faster, at least, the Athlon II out two years faster, and the Bulldozer would have been out to compete with the Core 2, easily beating it.[/citation]

Well, the problem with Bulldozer chips is too many engineers, so having more would not help. Yes, Intel's anti-competitive tactics hurt AMD. It hurt AMD badly. However, more engineers is not the solution. It rarely is. The team that built Bulldozer CPUs is several times larger than the older teams and look how that went. Bulldozer is a good architecture, but chips based on it have a very poor implementation of it because the higher ups in AMD decided instead of an expert team who can design transistor-by-transistor and optimize a chip for top performance, they would have a huge team of unskilled engineers who can't do anything other than use crappy auto-designing tools that hinder performance, performance per watt even more, and increase cost of manufacturing (decreasing cost of design probably more than it increases cost of manufacturing, the reason they did this).

Just be making a new stepping that was designed properly, Bulldozer could have 20% smaller dies that use 20% less power and give 20% more performance, all at the same time. Then we have cache that has between two and four times higher latency than Intel's cache and that might hurt performance even more than the poor design did. Then there's the crap memory controller that only gets about 75% of the bandwidth that the Sandy bridge controller gets at the same frequency (that's why AMD CPUs are more reliant on memory performance than Intel CPUs since Nehalem).

Bulldozer sucks because of stupid decisions. That memory controller is the same one that is used in Llano. Instead of improving the old design, AMD just used the same memory controller. The cache is no better than Phenom II and Athlon II either (worse, in some cases). The FPUs are half as wide as they should be (two 128 bit FPUs per module? Really, AMD, it should either be four such 128 bit FPUs or two better 256 bit FPUs) and that's why the eight core FXs don't fly past Intel in FP performance like they should.

Basically, right now, all of BD's problems are because AMD decided to not improve on a lot of older hardware and just leave it more or less the same, even fi they added more of it. The only improvement was the BD architecture (which really is a good architecture), but the crap design ruined the gains that it offered over Phenom II. AMD's laziness is what stopped them from catching up to Intel with BD. They had almost ten years of developing time with BD, yet they didn't even do half of the work that needed to be done and ruined a lot of the work that was done by implementing it in such a crap way.

A lot of engineers quit AMD because of this fiasco. In fact, they were in the news on a lot of sites. BD could have been so much better if AMD didn't get lazy with it.

Just a new stepping that features a redesign (done the right way) could improve Bulldozer performance per watt by more than 40%... Let alone improving the cache and memory controller. Furthermore, architectural improvements could be made, but it would already be a highly competitive product with Intel.

http://hardforum.com/showpost.php?p=1037482638&postcount=88

http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2

http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/6
 
[citation][nom]tourist[/nom]Trinity had my hopes up with gcn but an up clocked turks does not sound so appealing seeing as a a8 6550d will hit 900 MHz with ease.[/citation]

6550D @ 900MHz doesn't churn out much more performance than stock. Trinity has to have some sort of improvements that solve that in some way, or at least alleviate it. Perhaps the Turks is less memory bandwidth sensitive (might have more and/or better cache) than Llano's Redwood GPUs.
 
[citation][nom]tourist[/nom]Looks Like about a 15% increase across the board from the stock 600 ant that is only using 1600 mem[/citation]

15% increase from a 50% increase in clock frequency, meaning what... A 60% to 100% increase in power used by the IGP? I don't think that it's worth it. Increase the GPU's performance by almost 50%, yet it only manages to churn out 15% higher FPS. That's probably from the memory bottle-neck. Regardless of why, it's hardly an improvement, especially considering what it takes to get it.

Trinity has to solve that problem somehow, or else it would not be much faster than Llano and would most certainly be even less energy efficient.
 
[citation][nom]tourist[/nom]toms benchmarks showed roughly a 15% increase in fps at 960 over the the stock 600and yes i believe it is due to only using 1600 mem Heck just using 1866 gives a big boost to the stock speeds without using more power.[/citation]

Point is that Trinity needs to have a significant improvement on it's memory bandwidth reliance, or else any performance enhancements from a faster GPU are almost null, as evidenced by a 6550D being overclocked by more than 50%, yet not even showing a 20% gain. GPU improvements only go so far. Without memory improvements, GPU improvements are far less important.

When going from 1333MHz to 1600MHz gives a similar (possibly greater) jump in performance to the 60% overclock from 600MHz to 960MHz on the GPU clock did, then increasing the GPU clock is next to useless without more memory bandwidth to eat through.
 
[citation][nom]silverblue[/nom]Call me mad, but I've never heard of Llano using a different process technology for its GPU to the Stars cores. I don't understand why it, nor Trinity, would be that way.[/citation]
Ok, you're mad. Stars is a 40nm process while Llano is 32nm. I have high hopes for Trinity because if my Llano A8 laptop is any indication, it's going to be great. Llano was a huge success for AMD and I can't see Trinity being anything less than that.
 
[citation][nom]silverblue[/nom]Call me mad, but I've never heard of Llano using a different process technology for its GPU to the Stars cores. I don't understand why it, nor Trinity, would be that way.[/citation]

AMD uses 32 Husky cores in Llano, not Stars. Also, the modified Redwood GPU (called Sumo) in Llano is 40nm, so Llano already has different processes used by the CPU and GPU portions of the chip.

[citation][nom]Avro Arrow[/nom]Ok, you're mad. Stars is a 40nm process while Llano is 32nm. I have high hopes for Trinity because if my Llano A8 laptop is any indication, it's going to be great. Llano was a huge success for AMD and I can't see Trinity being anything less than that.[/citation]

Stars cores are on a 45nm process, not 40nm. 40nm is not used by many x86 CPUs, although it is the process used by a lot of GPUs for the last three graphics card generations and some ARMs.
 
[citation][nom]blazorthon[/nom]Well, if AMD gets Steamroller out before Haswell and it at least matches Sandy in performance per watt (not an unreasonable gain if Trinity can match or at least come close to Nehalem), then the next release could match Haswell, or at least come close. AMD could be back in the high end for CPUs in the next few years. Bulldozer is really a good architecture, it's just hindered by a poor implementation in the FX designs (150-250 poor engineers do not match 20-30 great engineers in anything except payroll, but AMD corporate guys seem to disagree), huge cache latency, and poor memory controllers. It could be very fast if the non-architectural problems were solved, maybe keep pace with Sandy in performance per watt (Ivy would probably still be out of reach for power efficiency, it's just too efficient to expect AMD to catch up in even two huge improvement releases).Intel could match AMD in GPU performance if they wanted to, but that would mean that Intel would need to allocate more die area to their IGP and such. It's already pretty big and the majority of Intel's customers would not see a benefit from going from HD 200 to 4000 already since they don't do graphically heavy things, so Intel simply doesn't want to go farther because it would mean increased cost of manufacturing without many people using the improved hardware to the full extent of it's capabilities. Honestly, HD 4000 is kinda overpowered for even HTPC work. The only people that would benefit from it are people who use graphics cards such as a 6570 and above, mostly the entry level gamers. AMD's APUs are focused on that group and although they are making AMD money, are they making enough money for Intel to care and does Intel want to hurt their rival in such a time? Let's not forget how bad things would be for Intel if AMD were to tank under.Anti-trust lawsuits galore. Intel needs AMD to be successful. So, Intel will leave AMD the lower profit markets while taking the higher profit markets. It works out for Intel who are rich and AMD gets enough money to sty afloat. If AMD can manage to catch Intel or at least close the gap in the next few years, then Intel might need to make some serious leaps to stay far ahead of AMD. The next few years could get much more interesting than the last few.[/citation]

^

LOL! Been trough a rough brainwashing I see. Nice..When the next Intel HD will catch up with Llano, let me know and don't tell me it's because "Intel doesn't want to" thats the most idiotic thing to say.
 


It's interesting that by going 28nm half-node CPU with Steamroller, it aligns with 28nm Southern Islands and, presumably, Sea Islands on the GPU side.

Rumors have Macau 28nm HkMG dumping gate first. The next logical step in the Son of Macau/Steamroller progression is 22nm HkMG - probably ET-SOI.


 


Well, 28nm should be better than 32nm, even if not by much.
 
[citation][nom]blazorthon[/nom]AMD uses 32 Husky cores in Llano, not Stars. Also, the modified Redwood GPU (called Sumo) in Llano is 40nm, so Llano already has different processes used by the CPU and GPU portions of the chip.
Stars cores are on a 45nm process, not 40nm. 40nm is not used by any x86 CPUs, although it is the process used by a lot of GPUs for the last three graphics card generations and some ARMs.[/citation]

Husky is a variant of the Stars line, and it is considered a direct descendant of the Stars architecture. Essentially it is correct to say the Llano are Stars based CPUs, but it is more correct to call them Husky cores. Stars started with Agena(Phenom I 65nm) and ends with Husky(A-series 32nm), so it actually spans 3 die sizes. Also Bobcat and Nano X2 are both 40nm x86 cores, so 40nm x86 CPUs do exist, albeit in niche markets.
 
[citation][nom]blazorthon[/nom]AMD uses 32 Husky cores in Llano, not Stars. Also, the modified Redwood GPU (called Sumo) in Llano is 40nm, so Llano already has different processes used by the CPU and GPU portions of the chip.
Stars cores are on a 45nm process, not 40nm. 40nm is not used by any x86 CPUs, although it is the process used by a lot of GPUs for the last three graphics card generations and some ARMs.[/citation]

Also, the Llano die is monolithic, everything in it is 32nm. The Gfx portion is based on a 40nm core, but it was reworked to be done in 32nm.
 
[citation][nom]e36_Jeff[/nom]Husky is a variant of the Stars line, and it is considered a direct descendant of the Stars architecture. Essentially it is correct to say the Llano are Stars based CPUs, but it is more correct to call them Husky cores. Stars started with Agena(Phenom I 65nm) and ends with Husky(A-series 32nm), so it actually spans 3 die sizes. Also Bobcat and Nano X2 are both 40nm x86 cores, so 40nm x86 CPUs do exist, albeit in niche markets.[/citation]

There's a distinct difference between being something and being based off of it and that was the point that I was making. On the no 40nm X86s, that was a typo (supposed to say many, not any). I thought that I fixed it already, but I see that I did not. It is corrected now.

[citation][nom]e36_Jeff[/nom]Also, the Llano die is monolithic, everything in it is 32nm. The Gfx portion is based on a 40nm core, but it was reworked to be done in 32nm.[/citation]

All right then, my bad there.
 
[citation][nom]e36_Jeff[/nom]Also, the Llano die is monolithic, everything in it is 32nm. The Gfx portion is based on a 40nm core, but it was reworked to be done in 32nm.[/citation]
THANK YOU. Sucks to be downvoted for something as petty as being correct.
 
Status
Not open for further replies.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom