Decoding Trinity

[IntelUser2000]

http://translate.google.ca/translate?hl=en&sl=zh-CN&u=http://news.mydrivers.com/1/208/208495_all.htm&ei=Xdi3Tt-cLqGOiAKlsaylAg&sa=X&oi=translate&ct=result&resnum=1&ved=0CBwQ7gEwAA&prev=/search%3Fq%3Dhttp://news.mydrivers.com/1/208/208495_all.htm%26hl%3Den%26biw%3D1280%26bih%3D822%26prmd%3Dimvns

CineBench R11.5 preliminary test, high frequency Trinity as reflected in the actual basically almost the same speed and Llano

According to the people who can read Chinese, it really says: "high frequency version of Trinity is almost the same speed as Llano"

So AMD themselves claim up to 30% improvement over Llano. How can that be true? Let's decipher.

First the benchmarks:

-Llano Desktop: http://www.anandtech.com/show/4476/amd-a83850-review

2.9GHz A8-3850 is equal to little surpassing the 3.1GHz Core i3 2100 in Cinebench and X264 HD. Ok, not surprising so far.

-Llano Mobile: http://www.anandtech.com/show/4444/amd-llano-notebook-review-a-series-fusion-apu-a8-3500m

Core i5 2520M(2.5GHz Base/3.0GHz dual core Turbo) is outperforming A8-3800M(1.5GHz Base/2.4GHz Turbo) by 45-55% in Cinebench and X264 HD.

So, in Cinebench R11.5, A8-3850 with clock speed disadvantage can outperform the Core i3 2100 by little over 10%. Therefore, A8-3800M runs at 1.8GHz all core Turbo to be 49% slower than Core i5 2520M.

Is Piledriver an enhancement over Llano/Phenom II cores?

Cinebench R11.5 64-bit

A8-3850 2.9GHz: 3.45-3.50(http://www.legitreviews.com/article/1649/9/)
FX-4170 4.2GHz: 3.17(http://www.techspot.com/review/452-amd-bulldozer-fx-cpus/page7.html)

It seems YES. Otherwise, the guys leaking the info wouldn't be able to claim the fastest Trinity can perform on par with Llano(since it can't with Bulldozer). The IPC gain seems to be about 10% in Cinebench and "3D Vision" application they mention.

But couldn't they have compared mobile versions and frequency disparity could be confused for IPC increases?

Certainly its possible. But early samples tested are mostly desktop samples. How many mobile samples have been leaked and tested?

Repeating the first question, how does AMD achieve 30% faster performance on Trinity compared to Llano on the CPU side?

Mostly clock speeds on just the mobile side. Decent IPC gain of about 10%, but mostly clock speeds.

What about the GPU? AMD claims 30% gains as well.

Possible that it has architectural changes. However,

Llano mobile GPU: 444MHz for the fastest
Trinity's slowest mobile GPU: one 563MHz model, with 600MHz-ish clocks being common. 600MHz/444MHz = 35%

Coincidence? Maybe. I don't think so though.

 

[Olikan]

there were already some leaks like that at phoronix, it just didn't had turbo speeds. ( or i didn't see )

still, trinity working samples at A0 is nice, they propably will reach the desired clock speeds they want, and fix the power leaks from bulldozer.

 

[Idontcare]

Trinity is on 32nm or is it on 28nm?

 

[Borealis7]

so Trinity is a 2nd generation BD core with IGP? and they're going to market it to the same segment as Llano? that doesn't sound like a good plan to me. Llano had better video performance than the competition. DB failed badly against its competition.

 

[cotak13]

Trinity's 32nm.

 

[LOL_Wut_Axel]

Trinity is 32nm. But if the Trinity IGP has a 600MHz clock speed it definitely has the potential to be 30% faster than the HD 6620G. It'd probably have a VLIW4 architecture and 480SPs.

 

[LOL_Wut_Axel]

Delete.

 

[grimpr]

IntelUser2000, Trinity will ridicule Intels Ivy Bridge in GPU performance, image quality, characteristics and drivers.

 

[IntelUser2000]

IntelUser2000, Trinity will ridicule Intels Ivy Bridge in GPU performance, image quality, characteristics and drivers.

Tell me more of something I didn't know, oh so great one.

 

[Soulkeeper]

hmmm 1 pin less, there is still hope ...
I'd kinda like to see them add more cache, but I guess that's not possible on 32nm with the transistor heavy bulldozer ie: 1.5MB * 4
I noticed one of the slides in the link mentions a power delivery version number increase "svi 2.0", hopefully that won't make all fm1 boards incompatible
also noticed hudson "d4" maybe that'll be part of a triple channel ddr3 ... or better yet ddr4 platform refresh ?

what would be really interesting is if all trinity cpu's were factory unlocked

I just purchased my second fm1 motherboard, would have gotten an A8-3870 to go with it if they had one ...

 

[IntelUser2000]

Chipsets no longer determine number of memory channels in modern CPUs that integrate memory controllers. It would help the graphics decent having an extra channel, but the extra wire routing and increased PCB layers wouldn't be worth it in a part like this, which is for low cost markets.

It does say on one of the slides L2 cache would be 1MB or 2MB, so it looks like it doesn't have L3 cache, like how Llano works.

 

[Soulkeeper]

Chipsets no longer determine number of memory channels in modern CPUs that integrate memory controllers. It would help the graphics decent having an extra channel, but the extra wire routing and increased PCB layers wouldn't be worth it in a part like this, which is for low cost markets.

It does say on one of the slides L2 cache would be 1MB or 2MB, so it looks like it doesn't have L3 cache, like how Llano works.

yes, it'll have the same amount of L2 as llano.
I guess I should have avoided confusion by saying "platform" when talking about a refresh
I am excited about ddr4 the wiki on it is an interesting read, but it's still a ways off

 

[Riek]

Is Piledriver an enhancement over Llano/Phenom II cores?

Cinebench R11.5 64-bit

A8-3850 2.9GHz: 3.45-3.50(http://www.legitreviews.com/article/1649/9/)
FX-4170 4.2GHz: 3.17(http://www.techspot.com/review/452-amd-bulldozer-fx-cpus/page7.html)

Wasn't the top bin 3.8Ghz for trinity that leaked so far? Wouldn't that make it closer to 20%?
As a 3.8Ghz trinity = 4.2GHz FX4170+8-9% in Cinebench.

 

[IntelUser2000]

Wasn't the top bin 3.8Ghz for trinity that leaked so far? Wouldn't that make it closer to 20%?
As a 3.8Ghz trinity = 4.2GHz FX4170+8-9% in Cinebench.

Earlier on I calculated 15% but I guess along the way forgot I was accounting for TurboCore.

I think even on that part its likely multi-core Turbo can be close to max Turbo, but you are right, that would make it better than 10%. 4GHz Turbo would make it 15%.

I would take a guess 3300/3600 chip would be the top mobile SKU.

I am excited about ddr4 the wiki on it is an interesting read, but it's still a ways off

DDR4 as I read isn't even gonna happen in the Haswell generation. So that makes it a 2015 mainstream memory standard. Being late is supposedly due to difficulty implementing the 16-bit prefetch required for DDR4.

 

[wahdangun]

hmm isn't in semi-accurate leak, they said trinity will have stacked memory?

 

[Riek]

Earlier on I calculated 15% but I guess along the way forgot I was accounting for TurboCore.

I think even on that part its likely multi-core Turbo can be close to max Turbo, but you are right, that would make it better than 10%. 4GHz Turbo would make it 15%.

I would take a guess 3300/3600 chip would be the top mobile SKU.

I forgot about the all core turboboost also..

I don't expect to see 3300-3600 on the mobile front, unless you talk about turboboost for half the modules. But given the GPU TDP needs to be accounted for also, i expect the base frequency to be around 2.6-2.8Ghz. (2.5Ghz would be to A8 3530MX what 3.8Ghz is to A8 3850)

 

[IntelUser2000]

I forgot about the all core turboboost also..

I don't expect to see 3300-3600 on the mobile front, unless you talk about turboboost for half the modules. But given the GPU TDP needs to be accounted for also, i expect the base frequency to be around 2.6-2.8Ghz. (2.5Ghz would be to A8 3530MX what 3.8Ghz is to A8 3850)

Well, I thought of that possibility also. The list has options we can choose from.

2500/3200 could work, 2.8-2.9GHz being the multi-core Turbo. The different from my 3300/3600 is due to differences we simply don't know. I also assumed that basing that mobile GPU might be at 6xxMHz rather than 7xxMHz for desktop, then 2500/3200 would be a perfect fit. But it seems too low if, they need 3800/4100 to equal 2.9GHz desktop Llano.

hmm isn't in semi-accurate leak, they said trinity will have stacked memory?

The headline actually says "far future GPU".

http://semiaccurate.com/2011/10/27/amd-far-future-prototype-gpu-pictured/

GCN successor?

 

[cbn]

First the benchmarks:

-Llano Desktop: http://www.anandtech.com/show/4476/amd-a83850-review

2.9GHz A8-3850 is equal to little surpassing the 3.1GHz Core i3 2100 in Cinebench and X264 HD. Ok, not surprising so far.

-Llano Mobile: http://www.anandtech.com/show/4444/amd-llano-notebook-review-a-series-fusion-apu-a8-3500m

Core i5 2520M(2.5GHz Base/3.0GHz dual core Turbo) is outperforming A8-3800M(1.5GHz Base/2.4GHz Turbo) by 45-55% in Cinebench and X264 HD.

So, in Cinebench R11.5, A8-3850 with clock speed disadvantage can outperform the Core i3 2100 by little over 10%. Therefore, A8-3800M runs at 1.8GHz all core Turbo to be 49% slower than Core i5 2520M.

Is Piledriver an enhancement over Llano/Phenom II cores?

Cinebench R11.5 64-bit

A8-3850 2.9GHz: 3.45-3.50(http://www.legitreviews.com/article/1649/9/)
FX-4170 4.2GHz: 3.17(http://www.techspot.com/review/452-amd-bulldozer-fx-cpus/page7.html)

It seems YES. Otherwise, the guys leaking the info wouldn't be able to claim the fastest Trinity can perform on par with Llano(since it can't with Bulldozer). The IPC gain seems to be about 10% in Cinebench and "3D Vision" application they mention.

A 4.2 Ghz "Quad core" Piledriver APU is getting a lower score than a 2.9 Ghz A8-3850 APU?

I fully admit that I haven't followed Bulldozer much, but isn't this due to the fact Cinebench is a rendering benchmark and the Bulldozer design (err, Piledriver) is light on Floating point?

 

[cbn]

With Bulldozer being a server design:

Does anyone have integer benchmarks comparing:

1. Intel processors
2. Bulldozer
3. Llano
4. Phenom II

 

[IntelUser2000]

A 4.2 Ghz "Quad core" Piledriver APU is getting a lower score than a 2.9 Ghz A8-3850 APU?

I fully admit that I haven't followed Bulldozer much, but isn't this due to the fact Cinebench is a rendering benchmark and the Bulldozer design (err, Piledriver) is light on Floating point?

That's not Piledriver. The 4.2GHz chip is Bulldozer.

I see. I looked over at Kuma chip vs Athlon X2. Cinebench had 20% gain. Of course FP can't be isolated because there are architectural improvements in Kuma.