Just for fun (come join in!) - Estimating Trinity performance

[randomaeiou]

I know BD is on everyone's mind at the moment, but frankly, I've never been one to spend exponential amounts for linear gains at the halo regions of hardware. Trinity interests me far more...

Anyway, here's a combining of a number of rumours to try and estimate Trinity performance.

CPU - most rumours/reports/leaks/conjecture believe that BD will perform at least on par with Nehalem, with the sky being the limit of predicted performance. Considering that an i3-2100 is considered acceptable for all but the most demanding gamers, a 2 module BD part (i.e. Trinity, which, if Nehalem is used as a yardstick/estimate, will pan out around i5-750 performance) should have little to worry about in the CPU department, even using the lower estimates of BD performance. Keep in mind that Trinity is not meant to be used with quad-fire 9990s or GTX990s - that is for the 300-core 0.1nm Timbuktu chip coming out in 2050. or the i9-9990x based on the OMFGPWNZORs architecture.

GPU - now for the interesting bit. It seems it is likely to be VLIW4, although some consider GCN a possibility. Let's work with VLIW4, because we can then extrapolate off Cayman. First off, AMD has promised +50% FLOPs compared to Llano's ~480 GFLOPs. The assumption is that most of this will come from the GPU rather than CPU. For ease, let's call it 100% GPU gains (likely an overestimate).

~720 GFLOPs, around HD6670. Next level is HD5750/6750 at ~1000 GLOPs. (keep in mind locked Llanos can still overclock about 15%).

We know: Llano die size = 228mm2, BD module size = 31mm2, Cayman die size = 389mm2.

Consider rumours of HD7570/7670 having 768 shaders. Compare HD6550d and HD6570 which both have 400 shaders. *If* HD7570 = iGPU for Trinity, then 768 VLIW4 shaders seems possible/likely.

Consider die size: Cayman = 389mm2 @ 40nm. Half Cayman (i.e. 768 VLIW4 shaders + bits) = ~125mm2 @ 32nm.

Now add 2 BD modules = ~187mm2, which leaves ~40mm2 for various bits and pieces before Trinity matches Llano die size.

These 768 shaders are likely to be clocked low to maintain decent TDPs. However, correctly clocked/overclocked, 768 VLIW4 shaders are likely to be able to hit the ballpark of the 720/800 VLIW5 shaders of the HD5750/5770 ESPECIALLY if AMD release an unlocked Black Edition Trinity like they plan to with Llano (3870BE). In the same place leaking 768 shaders for HD7570/7670, they give TDPs in the 50-60w range - make what you will of that information.

The only snag - memory bandwidth. Llano needs 1866-2000 RAM to feed its 400 VLIW5's properly. Realistically, by Trinity's release next year, 2133 RAM is about what we can expect to be affordable (unaffordable stuff should not be used with Trinity - go for the halo Zambezi/i7 chips instead if you can afford 2400+ RAM). Feeding 768 VLIW4s on 2133 might cause asphyxiation - think about the bandwidth on a HD5750.

(BTW, I think NVIDIA is in big doo-doo if this actually pans out - how many games CANNOT run well/medium (or even high!) @1080p, which is a common resolution in general, using a HD5750/5770?? Increasing compute power is all well and fine, if there is software to match.) And there's still the option of hybrid crossfire for even more grunt - pre DX10 games which do not benefit from hybrid crossfire are likely to be completely playable with a HD5750, and post DX9 games can harness hybrid crossfire... (in comparison to Llano, where preDX10 can still overwhelm its IGP but cannot benefit from hybrid crossfire)

 

[Dulak]

Im quite excited about trinity also ... and wish I didnt have to upgrade to a 990fx first and I could just go for the fm2.

Id be almost tempted atm to go for new unlocked llano - but its too bad you cant pair it with a faster gpu.

 

[Michaelius]

What's the point of giving 768 shares when you will starve them with non adequate bandwidth ?

768 shaders in VLIW4 is more than 800 shaders VLIW5 radeon 5770 and that card used fast 128 bit GDDR3.

 

[sirmonkey1985]

its more then likely going to be VLIW4 for the pure fact that its going to be used in the new xbox which means theres no way for them to completely rework it to support GCN within the time period microsofts going to want.

as far as memory bandwidth goes i think its going to change with the new socket, i'm sure they will have re-worked the memory controller to allow much higher bandwidth between the gpu and ram. heck it would be nice to see if they implemented something close to sideporting into it and gave it dedicated ram on the board its self(maybe something they will do with the new xbox, i don't know).

trinity is definitely something i'm interested in so i'll probably be keeping a eye on it until its released.

as far as the whole AM3+ FM2 deal its not really a big deal to me given the fact that trinity isn't going out until late 2012, so i'll already have a year into an AM3+ board before its released. which means if trinity is worth the money am3+ will still have great resell value left in it.

 

[randomaeiou]

oooh!! I'd totally forgotten about the new XBox rumours - that adds some perspective.

Consider that the maximum, and increasingly common resolution on TVs are either 720p or 1080p. I doubt MS are likely to settle for 720p given the average lifespan of a console generation. So, if MS are going to consume large amounts of Trinity chips, it is very likely that AMD will (or rather, they damn well better should) optimize performance for 1080p... Problem is, console software is VERY different from x86 software, so there really is no simple way to predict hardware performance from extrapolating console power/performance (or is there? I lack knowledge of console hardware).

Question now is: how to solve that memory bottleneck?! Considering the newly announced revised Komodo platform is dual-channel RAM, it is VERY unlikely that AMD's mainstream/budget platform will be triple or quad-channel.

Call me stupid, but I don't see how even dual channel 2400+++ RAM is going to keep 768 VLIW4 shaders busy, let alone "mainstream" speeds which are likely to be around 2133 by the time Trinity comes to market. Sideporting kind of goes against the spirit of Fusion, doesn't it? Unified address and all that jazz... Also, no matter how AMD tweak the memory controller, wouldn't the ultimate limiting factor be the TOTAL bandwidth provided by the RAM (which is already insufficient in itself for 5750 level GPU performance), even if the CPU takes up negligible bandwidth?

I don't know enough about RAM, but would having large AMOUNTS of RAM be able to compensate for limited bandwidth? (Imagine 4x4GB and using excess RAM space as a buffer during peak periods, which can then be whittled away when demand drops - workable?? Fast(ish) 4GB sticks are getting mainstream-cheap now) Or would increasing die size somewhat and sticking in a decent amount of unified L3 cache help?


Another random thought - HD6970 has ~2700 GFLOPs with 1536 shaders at 880MHz. For 768 shaders to hit ~720 GFLOPs (what AMD promised) we're looking at ~470MHz clocks. If we look at current Fusion products, all the 32nm GPUs seem to run at between 400-600MHz, which is right on the money! According to the same rumours, HD7670 will run at 900MHz and HD7570 will run at 750MHz. HD7670 rumoured to have 80GB/s of bandwidth, which means that the ~34GB/s of dual-channel 2133 DDR3 RAM might be in the correct ballpark to feed 768 shaders clocked at 450ish MHz. The numbers really are starting to look like they fit... Overclocking RAM looks like its going to become fashionable all over again.

Oh, I just realized that the rumoured 50-60W TDP of the HD7570/7670 are for 28nm chips. But still... half-clocked 32nm GPU can't run *that* hot, can it?

 

[ecmaster76]

You guys are forgetting that if the next XBox uses an AMD cpu that it will be a custom variant and not the actual Trinity product for PCs. Most likely they would equip it with a GDDR5 memory controller and solder 2-4GB of RAM for the whole system to share. No more memory bottleneck

 

[Michaelius]

Isn't next xbox using ibm cpu for backwards compatibility and amd gpu ?

 

[randomaeiou]

You guys are forgetting that if the next XBox uses an AMD cpu that it will be a custom variant and not the actual Trinity product for PCs. Most likely they would equip it with a GDDR5 memory controller and solder 2-4GB of RAM for the whole system to share. No more memory bottleneck

Oh, i'm fully aware that XBox Next is unlikely to have a full-fat x86 BD module/core on it. Don't see any reason they'll mess around with the GPU portion though... Would be fun to make the XBox Win8/HTPC compatible though - would open up quite a few markets I imagine, although OEMs might not like that much.

As for GDDR5 - I haven't come across any good explanation of the difference between GDDR5 and the DDR3 it is based upon, as in how that will affect the use of GDDR5 to feed a CPU rather than a GPU workload. So, I have no idea, basically. Also, rumour has it that AMD are considering/using XDR2 for 79xx series GCN GPUs - again not sure if that rumour has any bearing on topic at hand.

 

[sirmonkey1985]

Isn't next xbox using ibm cpu for backwards compatibility and amd gpu ?

theres still rumors floating around that they will just use the entire trinity APU which makes sense for them as a whole for manufacturing, cost, power usage, and heat. i don't think the IBM cpu really makes a difference for backwards compatibility, even going from the xbox to xbox 360 they had to put out updates for each xbox game to be supported on the 360, so it shouldn't be to hard to do the same thing for the 360 games.

Oh, i'm fully aware that XBox Next is unlikely to have a full-fat x86 BD module/core on it. Don't see any reason they'll mess around with the GPU portion though... Would be fun to make the XBox Win8/HTPC compatible though - would open up quite a few markets I imagine, although OEMs might not like that much.

As for GDDR5 - I haven't come across any good explanation of the difference between GDDR5 and the DDR3 it is based upon, as in how that will affect the use of GDDR5 to feed a CPU rather than a GPU workload. So, I have no idea, basically. Also, rumour has it that AMD are considering/using XDR2 for 79xx series GCN GPUs - again not sure if that rumour has any bearing on topic at hand.

the way G/DDR5 works its incapable of being used as a system memory without redesigning the architecture of the ram. the XDR2 rumor is probably legit. i wouldn't put it past AMD to not try something new, i mean hell they did it with the 3k series using GDDR4.. the 4k series using GDDR5.. after 3 generations of GDDR5 its about time they try something new. if it succeeds expect nvidia to copy it with the 700 series. i think the whole point of them going to XDR2 is the power usage. i highly doubt XDR2 will see mass market as a system memory, rambus doesn't have that good of a track record as for as system memory goes. if anyone remembers the RDRAM that Dell tried to mass market with the p4 OEM systems.. yeah that was a total fluke. but if you were smart enough to buy one of those systems for dirt cheap and hold on to the memory, you could make bank off them about 2 years ago, sold 4 sets for 200 a pop and that was considerably low for what they were usually going for.

 

[griff30]

I could see it using GDDR5 for the CPU and quad channel at that. Bandwidth will be everything.

 

[randomaeiou]

the way G/DDR5 works its incapable of being used as a system memory without redesigning the architecture of the ram. the XDR2 rumor is probably legit....... i think the whole point of them going to XDR2 is the power usage. i highly doubt XDR2 will see mass market as a system memory, rambus doesn't have that good of a track record as for as system memory goes.....

Care to offer a clueless person an understandable explanation why GDDR5 cannot be used as system memory although it is based upon DDR3? Also, what's the difference between GDDR5 and XDR2?

Was RDRAM's failure at mass market due to marketing/economic reasons, or performance reasons - having had a previous failure, one would hope rambus would learn from the lesson the second time round with XDR2!

 

[fobis]

theres still rumors floating around that they will just use the entire trinity APU which makes sense for them as a whole for manufacturing, cost, power usage, and heat. i don't think the IBM cpu really makes a difference for backwards compatibility, even going from the xbox to xbox 360 they had to put out updates for each xbox game to be supported on the 360, so it shouldn't be to hard to do the same thing for the 360 games.

Wait, what? The 360 uses a tri-core PowerPC (i.e., RISC) processor. It emulated original Xbox titles. The original Xbox used an x86 PIII class processor. There is no fathomable way that they will be able to achieve backwards compatibility if the next Xbox is based on an x86 platform. And frankly, I don't see how it would make sense for a gaming console. The architecture of a general-use PC is quite different from the modern-day gaming console.