Upcoming AMD Richland A10-6700 - Radeon HD 8670D tested!

[monstercameron]

Richland = Piledriver 2.0 with enhanced IMC. Its an improvement on the current architecture. AMD is slowly optimizing the once very slow and power hungry dulldozer. Nice to see them make good improvements!

is kaveri/steamroller a stepping stone from bulldozer or a major/minor redesign? if the latter then does tweaking bulldozer really matter if it will be thrown away?

 

[Rvenger]

is kaveri/steamroller a stepping stone from bulldozer or a major/minor redesign? if the latter then does tweaking bulldozer really matter if it will be thrown away?

Where are you IDC?? I need some expertise here.


Steamroller will be launched on 28nm and its a vast improvement on the piledriver arch. They probably have some errata to work out etc but it was promised that it would be launched on Glofo SOI bulk 28nm. Correct me if I am wrong IDC.

 

[frozentundra123456]

The clock increase can only account for a small amount of the performance increase (as it is roughly 5% higher on both the CPU [with turbo] and GPU side)--the rest is probably due to switching over to Steamroller from Bulldozer. It's nice to see AMD pushing hard in the budget market. This brings it to roughly GT640 levels, assuming the benchmark scores will scale with the framerate, which is pretty acceptable for budget gaming (assuming the price is around $130).

If correct, a nice improvement, but still pretty marginal for 1080p in newer titles. If equal to gt640 that looks like about half the speed of a hd7770. I still would want a discrete card.

 

[Sleepingforest]

Oh yeah, I would totally recommend a discrete GPU for everyone that can afford it (like $500 baseline, I think), but it is nice for mobile and really cheap builds.

 

[Idontcare]

Where are you IDC?? I need some expertise here.


Steamroller will be launched on 28nm and its a vast improvement on the piledriver arch. They probably have some errata to work out etc but it was promised that it would be launched on Glofo SOI bulk 28nm. Correct me if I am wrong IDC.

28nm should really increase the xtor budget AMD had for the design engineers.

The clockspeeds should be improved, 28nm being higher Idrive than 32nm, and I'd expect dynamic power to be lowered over that of 32nm as well. But I'd expect static leakage to be higher (or at best, no worse) than 32nm because they won't have SOI to assist them in lowering substrate leakage at 28nm.

So depending on what they do with their power-consumption budget and xtor budget, we'll either get something on the order of 20% higher IPC but no faster clockspeed (they throw all the xtors towards bolstering IPC), or they'll keep it lean and clean and push for 5GHz+ clockspeeds (small die, IPC not much better than PD, but higher clocks from the power-budget).

The problem is they are coming out late even for 28nm with Steamroller. 28nm has been out now for what, 18 months? AMD really needs steamroller to be on TSMC 20nm IMO with a roadmap to porting it to 16nm w/FinFet. At GloFo they simply have no future.

 

[frozentundra123456]

Oh yeah, I would totally recommend a discrete GPU for everyone that can afford it (like $500 baseline, I think), but it is nice for mobile and really cheap builds.

Yea, I have never been a fan of apus on the desktop, but I see a place for them in mobile, especially since many laptops are still 768p. It will be interesting to see how lower power versions compare to haswell.

 

[Hitman928]

What is enhanced?

Some minor tweaks as have been mentioned. The biggest tweak from what I've heard though is a reworking of the turbo system allowing the chips to use their turbo headroom much more effectively and stay at turbo frequencies longer which of course increases performance.

 

[OS]

Thanks, this is looking promising for my next laptop. Both haswell and richland have good projected IGP performance.

 

[mrmt]

The problem is they are coming out late even for 28nm with Steamroller. 28nm has been out now for what, 18 months? AMD really needs steamroller to be on TSMC 20nm IMO with a roadmap to porting it to 16nm w/FinFet. At GloFo they simply have no future.

Speaking of the WSA, I've been through the section in their latest annual report, they do mention the last amendment, the 2024 time frame and they do mention that there are purchase commitments until that date.

What caught my attention is that they don't mention anything about the exclusivity clause, but at least they do mention that once AMD acquires a new company, they have two years to transition the production for GLF.

 

[inf64]

RIchland is no faster (IPC wise) than Trinity. Ok maybe there can be 1 or 2 % deviation ,it's not impossible, but this fall right into the margin of error of testing. Richland does support higher memory clocks out of the box so GPU performance can scale much better.
On the other hand most of the Richland's improved performance comes from much improved and better controlled Turbo core. Charlie had a good article explaining it so you can look it up. Essentially it is boosting the component that needs the boost the most and is not boosting the component that is likely bottlenecked .

 

[KompuKare]

Interesting results but the first thing I looked for were details of memory speeds and configuration.

I've recently been playing around with an i5 mobile part (i5-2520M in a Thinkpad T420). The difference in iGPU performance between single channel PC3-10600 and dual channel PC3-8500 (hey, that's all I had to test with - hopefully will get some PC3-12800 soon to try) is huge. The difference was +20% on average but more in specific settings (Oblivion saw +45% for instance), PC3-12800 should gain even more if the results which Computerbase got in their HD4000 review are any indication.

Point being, for CPU performance RAM speed and single vs dual channel barely matters but for iGPU it's hugely important so while these results are interesting without listing the RAM speeds they leave a lot of question unanswered.

EDIT My bad, the CPUZ ram speeds are there in the third screenshot. Dual-channel PC3-17000. Which is about the fastest possible. While impressive I do wonder what RAM HD4000 had in those benches?

 

[inf64]

Interesting results but the first thing I looked for were details of memory speeds and configuration.

I've recently been playing around with an i5 mobile part (i5-2520M in a Thinkpad T420). The difference in iGPU performance between single channel PC3-10600 and dual channel PC3-8500 (hey, that's all I had to test with - hopefully will get some PC3-12800 soon to try) is huge. The difference was +20% on average but more in specific settings (Oblivion saw +45% for instance), PC3-12800 should gain even more if the results which Computerbase got in their HD4000 review are any indication.

Point being, for CPU performance RAM speed and single vs dual channel barely matters but for iGPU it's hugely important so while these results are interesting without listing the RAM speeds they leave a lot of question unanswered.

All of the non-richland results in the 1st post of this thread are with the dual channel configuration. So whatever the memory configuration for ES Richland 6700 was, the results listed cannot be lower than what they are(they can only be higher). I think the results are with dual channel configuration so this is how 6700 will likely perform when it's released.

 

[NTMBK]

Where are you IDC?? I need some expertise here.

[...some expertise...]

Like magic :awe: We need an IDC-signal installed on the roof or something.

 

[KompuKare]

Yes, I hadn't zoomed in on the third screenshot. It was dual-channel PC3-17000 (aka DDR-2133). But I tried to look up those other scores (to see what they were tested with) but couldn't find them. I though they were from SweClockers but I wasn't able to find anything there. But I found them now, the 3dmark is from this:
http://www.sweclockers.com/artikel/16499-futuremark-lanserar-nya-3dmark-for-windows/2
and they link to a test system:
http://www.sweclockers.com/recension/16383-asus-rog-ares-ii/3#pagehead
which has "16 GB Corsair Vengeance, 1600 MHz, 9-9-9-24".

So the other results are with PC3-10600 and this is PC3-PC3-17000. Taking those Comptuerbase results again
http://www.computerbase.de/artikel/grafikkarten/2012/test-intel-graphics-hd-4000-und-2500/15/
they got 11% between PC3-10600 and PC3-14900. But on the other hand this A10-6700 is possibly an engineering sample.

 

[Idontcare]

Speaking of the WSA, I've been through the section in their latest annual report, they do mention the last amendment, the 2024 time frame and they do mention that there are purchase commitments until that date.

What caught my attention is that they don't mention anything about the exclusivity clause, but at least they do mention that once AMD acquires a new company, they have two years to transition the production for GLF.

2 years!? That is ridiculous. Basically requires a complete redesign of the product to transfer between foundries.

What that clause means in reality is that for any company that AMD buys, AMD basically has 2yrs to sell the existing products that the company was already selling before AMD bought them.

The economics of redesigning and ramping to production an already-in-production is simply cost-prohibitive. Even Apple can't make it happen at 28nm and they deplore Samsung. How the heck would an AMD pull that off?

(not to mention if they bought anything that had a leading-edge design, meaning foundered at TSMC, what the hell would they be porting it to at GloFo anyways? A slower, bigger-die, power hungry process node?)

Hector and Dirk really setup AMD to fail. I just can't believe the shareholders weren't made aware of all this at the time, huge implied liability and asset dissolving contract and not a peep was mentioned at the time of the spin-off.

 

[FlanK3r]

I benchmarked A10-5800K+HD7660D OCed at the same clocks as A10-6800K...It is in Coolaler forum too. Difference are very small, maybe 1-2% clock to clock. Only my lowe preset 3DMark showed me some bug score over 120 000 :-D...

 

[spaceman]

looks nice

 

[csbin]

I benchmarked A10-5800K+HD7660D OCed at the same clocks as A10-6800K...It is in Coolaler forum too. Difference are very small, maybe 1-2% clock to clock. Only my lowe preset 3DMark showed me some bug score over 120 000 :-D...

120 000 :biggrin:

 

[FlanK3r]

yes, WR bug

 

[mikk]

Looking at that chart what would account for the increased performance? Max clocks are less that 10% higher. Hasn't AMD said there are no architecture changes?

If the driver aren't the same then it can give us inaccurate results. I guess AMD optimized the drivers for the new 3dmark in recent drivers and if Richland used a newer driver...

 

[sm625]

I benchmarked A10-5800K+HD7660D OCed at the same clocks as A10-6800K...It is in Coolaler forum too. Difference are very small, maybe 1-2% clock to clock. Only my lowe preset 3DMark showed me some bug score over 120 000 :-D...

When you say difference is small, do you mean the difference between the 5800K overclocked and 5800K stock clocked scores is very small, or the difference between the 5800k overclocked and the OP's 6700 scores is very small?

 

[The Alias]

you guys are getting it wrong that's not just nice that's effing amazing he put his 5800k to 6800k clocks and scored less than the 6700 that's REALLY good for basically the same chip

 

[Centauri]

When you say difference is small, do you mean the difference between the 5800K overclocked and 5800K stock clocked scores is very small, or the difference between the 5800k overclocked and the OP's 6700 scores is very small?

5800 and 6700

 

[inf64]

6700 supports 2133Mhz RAM so its GPU ,even at 5800K's GPU clocks, will perform a lot better. Couple that with more Mhz on GPU and better CPU Turbo when needed and you have much better perf./watt chip (65W vs 100W,better performance).

 

[mrmt]

2 years!? That is ridiculous. Basically requires a complete redesign of the product to transfer between foundries.

(...)

The economics of redesigning and ramping to production an already-in-production is simply cost-prohibitive. Even Apple can't make it happen at 28nm and they deplore Samsung. How the heck would an AMD pull that off?

It is there, write in black in their annual report:

(...) If we acquire a third-party business that manufactures MPU products, we will have up to two years to transition the manufacture of such MPU products to GF.

In typical AMD shaddy style, the WSA information is spread around a lot of topics, there isn't a section with the entire relevant information, you need to read the entire report to get a picture of the agreement. Other relevant information, AMD had to pay 700 million to get Kabini to TSMC, and only during a limited time frame.

Hector and Dirk really setup AMD to fail. I just can't believe the shareholders weren't made aware of all this at the time, huge implied liability and asset dissolving contract and not a peep was mentioned at the time of the spin-off.

They weren't. The WSA wasn't disclosed until sometime later and with a lot of redacted clauses, and there were the amendments that changed a lot of those redacted clauses... The only surprise here is that nobody ever filled a lawsuit against the BoD for breach of fiduciary duty.