Trinity review

[Nemesis 1]

No, you're right. AMD's Trinity is currently priced way below the i7 IBs with HD4000. The 3720QM costs close to $400 for the chip alone so that's 2-3x more than AMD is charging for the Trinity A10 and close to a brand new A6 Llano laptop... yes, the cost of the entire laptop and not just the APU

If we get i5 and i3 Ivy's with HD4000 then it's going to be interesting, but if Intel only equip the lower end with HD2500 then Trinity will stomp on anything in its price range. I hope they offer some i3's with HD4000, as that would make me debate whether to go Intel or AMD this round.

As far as bang for your buck goes, Trinity seems to do very very well against Sandy-based i5's and i3's even with discrete graphics. It's definitely impressive.

all lowend laptops will have G3 graphics. Highend desktop only gets g2 graphics

 

[Nemesis 1]

It's going to be pretty cool when the desktop version of this comes out.

If the mobile part's GPU is clocked around 450 mhz, the desktop I've heard will be 800. Probably can hit 1ghz with good water cooling.

So your adding water to a system that people buy based on its cheapness. a discrete graphics card would be cheaper than watercooling/ If your an AMD fanbois knock yourself out on a third rate system . Intel SB-e is number one Intel IVB is number 2 which leaves AMD at 3rd rate

 

[pelov]

all lowend laptops will have G3 graphics. Highend desktop only gets g2 graphics

http://www.anandtech.com/show/5192/ivy-bridge-mobile-lineup-overview

We still don't know anything about the i3's, including release date. They've been pushed back like three times now

Notice the clocks are roughly equal whereas AMD's chips differ quite a bit as far as graphics goes, both shader count and clock speeds. The Intel chips decrease in cache quite a bit, though. Considering it's shared in the Intel design it should impact performance but by how much, iunno.

 

[Nemesis 1]

Well an i3 needs alot lessll3 cache than i5 as its only 2 cores and that is what this apu from amd will be comparred to . as its away to make it look closer than it actually is . cpu wise that is. intel only lowered the bottom alot in mgz. the top is only 100mgz slower.or 150 not going to look and see.AMds 17watt apu will be smoked by intels i3 17 watt cpu in all things every metric except cost. Than there is the i5 17 watt units that isn't to be compared to amd 17 watt units

 

[pelov]

I posted the following on [H] but I think it bears repeating here. The two single-threaded benchmarks that we do have show quite a bit of improvement, Cinebench and x264 first pass:

Sure it has, granted it looks to be mostly done via clock speed gains. I don't think they gained 10% IPC in most workloads. It's probably in the ~5% range? Either way the performance increased via clock speeds and the perf-per-watt increased dramatically. So much so that Trinity actually consumes less power than Ivy Bridge does at 22nm despite those crazy clock speeds... It's almost on-par with mobile Sandy Bridge as far as power consumption goes. Those are great signs, imo. Granted, ~10% IPC bump would have been ideal but Trinity is already at a disadvantage because it lacks L3 cache so in gaming that would roughly equal about 10% performance gain.

here
http://forums.anandtech.com/showthread.php?t=2245809

At first glance it can look somewhat unimpressive, but it's actually very good. I'd expect the games where it loses to the HD4000 to turn around in AMD's favor with drivers and the games where it does lead, which is all but 2, to look even better.

I think you're expecting SB-levels of single-threaded performance which was never going to happen. At least now the single-threaded workloads are respectable and actually improve upon Llano/Stars which BD failed to do. They're still off, by 20-30%? But considering the gap was 40-50% I think that's quite the achievement. Vishera won't have an on-die GPU and will likely clock very high, but how much that resonant clock mesh tech has helped Trinity and how much it will help Vishera is another matter. The clock mesh tech decreases power consumption by 10% or increases clock speeds by 10% at equal TDP. Those gains diminish as the clocks go past 4ghz.

err, in some places it's actually very good IPC gains. First pass x264 is a single-threaded benchmark so let's take that as an example:
49 for Trinity
53 for the i5 Sandy

The Trinity APU clocks up to 3.2ghz and the i5 clocks up to 2.9ghz, both include turbo.
Trinity gets roughly 1.53FPS per-mhz
the i5 Sandy gets roughly 1.83FPS per-mhz

That's only a 16% IPC deficit in that workload.

Here are the numbers from BD vs. Sandy in the same workload:

i5 2500K at 3.7ghz w/ turbo
8150 at 4.2ghz w/ turbo

2.7FPS per-mhz for the 2500K
1.8FPS per-mhz for BD

That's roughly a third of the IPC deficit for that workload when comparing the 2500K to the BD, meaning AMD gained ~half of that in IPC and not just clock speeds.


Cinebench is only ~30% now from ~50% between an i5 Sandy and BD with roughly the same clock speed gap (2600K vs. 8150 Cinebench 11.5 single-threaded). It was never meant to catch up to Sandy with IPC but it looks like they may have gained a good bit of that 10% back (and then some in certain workloads), but it does bode well for Vishera as far as power consumption goes and at least showing some respectable performance, granted a year too late.

It won't be a sandy/IB killer but if they price it well nobody will complain about the power consumption, the heat and the performance won't be too bad either

Note that I'm not buying Vishera anyway because I don't want anything CMT-related and I'm looking to dump my desktop entirely, but for those that are looking forward to AMD producing something that's actually worth a damn on the desktop again it looks as if they've made some decent improvements in IPC. Given a higher TDP, 50% extra die space being a straight CPU (not an APU) and much MUCH better perf-per-watt and power consumption, Vishera might not be all that bad. Granted, this was derived from only 2 benchmarks that we have so it's not much. I'd love to see somebody take a deeper look into the IPC gains over a greater spread of workloads.

 

[Phynaz]

It's going to be pretty cool when the desktop version of this comes out.

If the mobile part's GPU is clocked around 450 mhz, the desktop I've heard will be 800. Probably can hit 1ghz with good water cooling.

Water cool an IGP? For what purpose??

 

[nyker96]

OP, please add the trinity review links for anadtech to your post plz:

http://www.anandtech.com/show/5831/amd-trinity-review-a10-4600m-a-new-hope/1

 

[Asterox]

not likely.....

bulldozer is very good when kept the voltages lower,
once you need to more volts to achieve highter clocks, things go downhill

Piledriver Core is not the same as Bulldozer Core you are very poorly informed how I see it. Pilerdiver Core(Trinity APU/Vishera Desktop Piledriver Core) use a Cyclos Resonant Clock Mash Tehnology and this brings higher CPU frequency lower power consumption and less heat.

http://www.cyclos-semi.com/technology/

http://www.cyclos-semi.com/pdfs/time_to_change_the_clocks.pdf

http://www.techpowerup.com/160927/C...al-Implementation-of-Resonant-Clock-Mesh.html

 

[pelov]

Here's cinebench 11.5 single-threaded:

BD at 4.2 w/ Turbo gets ~.024 per-mhz
2600K at 3.7 w/ Turbo gets ~.411 per-mhz
That's absolutely massive...

A8-3520m clocks up to 2.4ghz turbo
i5 2410m clocks up to 2.9ghz turbo
Trinity APU clocks up to 3.2ghz turbo
All 3 are 35W TDP mobiles

Trinity gets .024 per-mhz
Llano gets .027 per-mhz
i5 Sandy gets .0396 per-mhz

How this is going to translate to desktop chips I have no idea, but it seems the improvements are pretty significant.

 

[Khato]

BD at 4.2 w/ Turbo gets ~.024 per-mhz
2600K at 3.7 w/ Turbo gets ~.411 per-mhz

Trinity gets .024 per-mhz
i5 Sandy gets .0396 per-mhz

How this is going to translate to desktop chips I have no idea, but it seems the improvements are pretty significant.

So according to that, Trinity shows no improvement in single threaded performance. The x264 results are not as easy to directly compare as it's not a single threaded task, it's 'lightly threaded' - there's more than one thread running, but it's a static number that I believe is greater than 2 (looking at bench, a core 2 duo E8600 at 3.33 GHz scores 46.7 fps while a core 2 quad Q8400 at 2.66 GHz scores 59.9 fps.) Going over 2 threads explains why the i7-2820QM sees such a marked gain over the i5-2410M (both more cores and markedly higher turbo frequencies while more than one core is in use.)

 

[sandorski]

Trinity is an improvement over Lano across the board. It will Sell very well.

 

[dma0991]

Intel haswell does not have an L4 cache . What it has is crystelwell which is not L4 cache google is your friend

Whether it is a L4 cache or not, the main concept still remains, an extended memory for the IGP that is placed side by side with the main die. This news leads back to Charlie, I'll give it a 50/50 chance.

 

[pelov]

So according to that, Trinity shows no improvement in single threaded performance.

It is a marked improvement. The IPC for that benchmark is roughly equal between Llano and Trinity so the performance gains are derived from clock speeds, which means a gain roughly equal to the difference in clock speeds between the two architectures. Trinity is devoting 50% of its die to the on-die GPU as well as stuck at 3.2ghz clock speeds so don't compare the per-mhz to BD per-mhz and assume it hasn't changed. Comparing BD desktop to Trinity laptop isn't a fair comparison, just like comparing a mobile Sandy to a 2500K. What you should get from those 2 examples (unfortunately the only 2 I can seem to find ) is that they improved IPC significantly across both workloads and that single-threaded performance has increased drastically in both workloads. The BD vs 2500K and Llano vs Trinity vs i5 sandy was just to see by how much relative to equal TDP.

Mind you, it's IPC in only those 2 specific workloads. Unfortunately the reviews are pretty lackluster as far as analyzing the Piledriver module's performance so at the moment that's all we have. I'm sure IPC across a far more broad set of workloads will vary significantly (like it does with any architecture) but what AMD needed to make up was roughly ~10% they lost from going Thuban > Bulldozer (the 10% being an average across various workloads). The 2 benchmarks above seem to indicate they've at least worked on those 2 I guess we'll see how much info we get in the days prior to people getting their hands on them.

 

[grimpr]

Trinity is an improvement over Lano across the board. It will Sell very well.

I Agree, its a good chip, personally i cant wait for an 100$ desktop unlocked Quad Triny that will clock very good, paired with 60$ FM2 motherboards will make a potent entry allaround, gaming and home video editing machine.

 

[frozentundra123456]

I'll be nabbing a Trinity laptop. My IBM T42 just isn't cutting it anymore. Would be really nice to be able to play games on a laptop...

Yea, agreed, it would be nice to game on a laptop without a discrete card. However, the problem is that Trinity only improves graphics by about 20%. I was hoping for a bigger improvement. Still have to use pretty low resolutions and settings. Doesnt quite make the "up to 50%" improvement that AMD was touting in their slides.

So all in all, not bad, but not really improved enough over Llano to make a big difference either. I was hoping for something like HD5670 level performance.

 

[pelov]

If it's anything like the Llano release, it'll improve with driver updates. For some reason the guys at AMD are complete dumbasses when it comes to releasing products. Instead of looking to make a very good first impression they just screw it up with really crappy beta drivers. By the time you get a proper driver, generally 2-3 months down the line, everyone's forgotten about your product or has already formed their opinion.

I'm pretty sure that 20% is going to look more like 30%+ as soon as they get up off their ass.

We saw 35-45% higher scores in 3DMark 11 and Vantage, which tend to remove the CPU from the equation more than actual games, so our guess would be that if AMD continues with their APU plan they’re going to need to work more on the CPU side of the equation. We also see the same thing looking at the VAIO SE scores in the earlier gaming charts: the HD 6630M scores are 20% faster on average, but much of that appears to come from the faster CPU rather than the GPU.

Considering AMD has also improved CPU performance by a decent margin I highly doubt there's any sort of CPU bottleneck going on here. The Starcraft result seems to point to the opposite, actually. Considering Starcraft and BF3 saw the largest gains it's almost certainly not CPU related but rather another driver issue... again.

 

[Khato]

It is a marked improvement.

Guess I have a different definition as to what constitutes a 'marked improvement'. I will agree that Trinity does appear to have slight improvements, but around or below the 10% mark.

For Cinebench 11.5 Single-Threaded:
Desktop A8-3850 - 0.89/2.5GHz = 0.3069 per GHz
Mobile A8-3500M - 0.64/2.4GHz = 0.2667 per GHz

Desktop FX-8150 - 1.02/4.2GHz = 0.2429 per GHz
Mobile A10-4600M - 0.77/3.2GHz = 0.2406 per GHz (note that there's a discrepency in that the Bench has it with a score of 0.7 instead, which would be 0.2187 per GHz)

Assuming that you should see the same scaling between desktop vs mobile, then you either have a ~10% performance increase over Bulldozer IPC if the 0.77 score is correct or a flat performance... Since I can't find anyone else that does single-threaded cinebench R11.5, let's drop back to R10. There the FX-8150 does 3938 for 0.9376 per MHz while according to notebookcheck.net the A10-4600M gets 2820 for 0.88125 per MHz, which isn't quite so close.

 

[pelov]

Assuming that you should see the same scaling between desktop vs mobile,

That's never the case. In fact, both you and I showed that above

For Cinebench 11.5 Single-Threaded:
Desktop A8-3850 - 0.89/2.5GHz = 0.3069 per GHz
Mobile A8-3500M - 0.64/2.4GHz = 0.2667 per GHz

Don't compare Bulldozer IPC or benchmarks to a mobile benchmark. It's a completely different animal.

i5 Sandy gets .0396 per-mhz

2600K at 3.7 w/ Turbo gets ~.411 per-mhz

That's a substantial difference.

The point wasn't to compare the desktop to mobile parts, doing so would be idiotic, but rather to compare mobile versus mobile and desktop versus desktop with respect to architectures. At TDP constrained segments the Trinity chips do very well and in those two benchmarks have shown to make up ground with their SB mobile counterparts. That's great news considering they're single-threaded workloads. On the desktop side we don't have a Vishera chip to compare yet but when we look at Bulldozer versus Sandy we saw a far more massive difference between the two. The key takeaway being the differences between desktop versus desktop and mobile versus mobile.

We can't put Trinity into big boy shoes because there are a lot of factors that would go into coming up with its theoretical performance. The list is quite long: the VLIW4 shaders in Trinity take up 50% of the die space that Vishera will have freed up, that single Trinity chip is at 35W TDP compared to a much higher TDP Vishera, clock mesh tech has diminishing returns past 4ghz, we don't know the TDP of Vishera, we don't know what else AMD has changed in the Piledriver cores for Vishera (they said there would be some small improvements on top of Trinity), etc. etc.

So according to that, Trinity shows no improvement in single threaded performance.

You're mistake is in comparing desktop to mobile. The two are staggeringly different. Single-threaded performance certainly looks to have improved, though.

~16% and ~13% improvement on those two.

 

[Olikan]

Piledriver Core is not the same as Bulldozer Core you are very poorly informed how I see it. Pilerdiver Core(Trinity APU/Vishera Desktop Piledriver Core) use a Cyclos Resonant Clock Mash Tehnology and this brings higher CPU frequency lower power consumption and less heat.

i know all these stuffs, but bulldozer is not much diferent than piledriver...

bulldozer clock/watt is very manufacturing sensitive, everytime that it needs some voltage bump, it consumes alot more...

that's bad when looking at desktops, when performance is king!
But at lower voltage/clocks, like trinity, bulldozer shines in power consumption...

resonant clock mesh just helped amd to stay competitive against 22nm from intel...

 

[mrcmtl]

Guess I have a different definition as to what constitutes a 'marked improvement'. I will agree that Trinity does appear to have slight improvements, but around or below the 10% mark.

For Cinebench 11.5 Single-Threaded:
Desktop A8-3850 - 0.89/2.5GHz = 0.3069 per GHz
Mobile A8-3500M - 0.64/2.4GHz = 0.2667 per GHz

Desktop FX-8150 - 1.02/4.2GHz = 0.2429 per GHz
Mobile A10-4600M - 0.77/3.2GHz = 0.2406 per GHz (note that there's a discrepency in that the Bench has it with a score of 0.7 instead, which would be 0.2187 per GHz)

You cannot use boost clocks to get IPC. Nor can you use single-threaded performance to get the actual IPC just because turbo is "screwing" your results. The only way to get a clear picture of Trinity's IPC improvements is to use a multithreaded application that loads all cores and compare it to a FX-4100 at the same clocks.

Aside from the IPC talk, according to the THG review, AMD actually made tremendous overall CPU performance gains compared to Llano: http://www.tomshardware.com/reviews/a10-4600m-trinity-piledriver,3202-18.html

 

[IntelUser2000]

Pelov:

i5 Sandy gets .0396 per-mhz

2600K at 3.7 w/ Turbo gets ~.411 per-mhz

There's no such gap. 2600K has a single core Turbo of 3.8GHz, that puts it at 0.4, which is practically no different from the mobile Sandy Bridge.

~16% and ~13% improvement on those two.

That's due to clock speeds. The First Pass test of X264 takes advantage of more than 2 cores. Llano's Turbo Core is borderline useless, while Trinity is much better there. Actually even assuming Trinity doesn't have Turbo Core kick in, the base clock of 2.3GHz is far higher than Llano's 1.6GHz.

It's same with 1 thread Cinebench as well. Llano's max Turbo is 2.5GHz while A10-4600M is at 3.2GHz. Though I wouldn't deny there's an IPC gain over Bulldozer.

To determine IPC improvements, you can see how the Phenom II 980 compares with FX-4100 here: http://www.legitreviews.com/article/1766/9/

Phenom II is 19.4% faster with Bulldozer having 2.7% advantage in clocks. Trinity is 28% faster than A8-3520M and needs 14.9% clock advantage to do it. That means Trinity is 10% faster in Cinebench compared to Bulldozer.

resonant clock mesh just helped amd to stay competitive against 22nm from intel...

The power savings improvement is due to better power management. In idle when CPUs are using C-states, clocks are turned off and neither process tech nor RCM matters.

RCM would help in allowing AMD to make a 17W chip, where previously it couldn't have done it before.

 

[georgec84]

http://www.notebookcheck.net/Trinity-in-Review-AMD-A10-4600M-APU.74852.0.html

 

[mrcmtl]

Llano's Turbo Core is borderline useless, while Trinity is much better there.

Llano's Turbo Core is far from useless. It is not the Turbo that existed in Thuban but the improved version similar to the one in BD. In purely single threaded apps, Llano is in turbo state more than 50% of the time.

 

[Khato]

You cannot use boost clocks to get IPC. Nor can you use single-threaded performance to get the actual IPC just because turbo is "screwing" your results. The only way to get a clear picture of Trinity's IPC improvements is to use a multithreaded application that loads all cores and compare it to a FX-4100 at the same clocks.

I agree completely - I was simply playing the same game Pelov was, but to show that there haven't been single-threaded performance gains instead. Is it faster than Llano? Yes, definitely... just like Bulldozer was faster than the desktop Llano.

 

[Vesku]

The A8s have the same cache hierarchy, Trinity A10 lacks the L3 of FX.

For Cinebench 11.5 Single-Threaded:
Desktop A8-3850 - 0.89/2.5GHz = 0.3069 per GHz
Mobile A8-3500M - 0.64/2.4GHz = 0.2667 per GHz

Desktop FX-8150 - 1.02/4.2GHz = 0.2429 per GHz
Mobile A10-4600M - 0.77/3.2GHz = 0.2406 per GHz (note that there's a discrepency in that the Bench has it with a score of 0.7 instead, which would be 0.2187 per GHz)