How long will 290x memory controller go?

[krumme]

The initial reading of the review of Ryan, led me to beliewe AMD use of a wider 512b bus, and the associated cost of eg. pcb, was more than outweigted by the reduction in diesize. Giving good cost perf. going from 288GB/s at Tahity to 320BG/s Hawai (5GHz mem).
http://www.anandtech.com/show/7457/the-radeon-r9-290x-review/3
I interpreted this as the saved diesize would lead to lower clocked - and cheaper - mem. Still sounds reasonable and sensible to me.

But then the oc results started to come in, and what stroke me was the seemingly inability
to run lower than 6.4Ghz whatever 5Ghz ram brand was on the card giving 20-25% oc out the gate.

Nearly 1½ year ago Charlie at SA predicted NV 680 would beat Tahiti and the reason was a "fixed" memory Interface because it was "broken" earlier. When 680 hit what was striking and still is, the far more efficient memory interface compared to Tahiti. Among fully inaccurate things Charlie got this fully acurate.

I therefore went to Charlie to see his matter on the subject. And his part on the memory on Hawai is imho quite interesting:

"The memory controllers are claimed to be far more efficient and use 20% less area than the 384b controller in Tahiti. 20% faster and 20% smaller seems a bit odd for a 33% wider controller but the reason for it is quite simple, powers of two. With a power of two memory width there was no need to put crossbar in front of the memory controllers so that area was saved. Not having to make a few extra switches also means faster and lower latency throughput as well. The dual DMA engines can saturate a PCIe3 16x bus so getting data on and off the card itself shouldn’t be a problem either.

The memory controller architecture is based on those in Bonaire with only minor changes but Hawaii runs them significantly slower than the two generation older Tahiti. Why did we call a 5.0GHz GDDR5 interface a good thing earlier? Easy, Bonaire had a screamingly fast memory controller, it wasn’t hard to find one that would clock to more than 7.0GHz, way more, if you had the right memory. This wasn’t golden sample or hand polished memory just vanilla chips and random memory would often overclock like mad. So think about a cleaned up version of that, twice as wide and clocked very low. Then think about the special edition 290XXX’s or whatever that will surely come out soon. Then grin."
http://semiaccurate.com/2013/10/23/long-look-amds-hawaii-volcanic-islands-architecture/

If one is able to handle Charlies typical style that what was "broken" and "fixed" for NV is an improvement for AMD 1½ year later - and lead to grin, one might apriciate the valuable information actually presumably given here.

That lead me to look at bonaire (Its labelled 7790, for those of you who like me, dont follow the low end). I have always looked as that gpu as one of the most uninteresting gpu of all time and didnt care for it a second. Well it probably is the forefather/mother of Hawai, all iterations of comming playstation, xbox, kaveri. Probably the most important gpu for AMD ever. A soc without cpu and 3g and lte. lol.

And the overclocking of this cheap gpu is like this:
http://www.techpowerup.com/reviews/XFX/HD_7790_Black_Edition_OC/29.html

XFX HD 7790 Black Ed. core, 1220 MHz, memory 1940 MHz (near 8Ghz)

So is the memory controller of Hawai good perf/mm2 but made for lower speed or was the memory controller in Tahiti "broken" and "fixed" in Hawai leading to 25% higher speed and bandwith for the non ref cards or perhaps upcomming 295?

Whats your take on it?

 

[3DVagabond]

I agree with the thought of wait for the special edition 290X's. There's performance left on the table, IMO.

 

[BrightCandle]

The low reference speed is likely more about power saving than it is performance. I can't remember which review it was (maybe hardocp) but they overclocked the memory interface quite a bit (5.0 to 6.4) and the benchmark results barely moved. The card just isn't memory bandwidth constrained in anything they tested. So its a key place to save power which can go directly into high clock speed on the main core (both power budget and thermal budget).

The cards with better coolers are likely going to clock that memory much much higher and along with removing the downclocking should result in a decent boost in performance even without increasing the maximum clock speed.

 

[krumme]

290x bandwith is only 10% more bandwith than 7970 for what aprox 30% more shaders. And we see on eg. Anand Toxic 7970 review 6.5GHz still is usefull. Clearly there is a huge efficiency improvement besides pure bandwith. Its also shown by the 4k results. No matter eg more rops there is mem perf in spades.

But that also begs the question. Why not keep the 384b and use eg 6.5GHz? Its hardly that more power consuming. My take on it; it would also be cheaper than pcb for a wider 512b bus and the added mem chips. The 512b makes only sense if you want to really use it.

Perhaps eg for professional market??

 

[Leadbox]

290x bandwith is only 10% more bandwith than 7970 for what aprox 30% more shaders. And we see on eg. Anand Toxic 7970 review 6.5GHz still is usefull. Clearly there is a huge efficiency improvement besides pure bandwith. Its also shown by the 4k results. No matter eg more rops there is mem perf in spades.

But that also begs the question. Why not keep the 384b and use eg 6.5GHz? Its hardly that more power consuming. My take on it; it would also be cheaper than pcb for a wider 512b bus and the added mem chips. The 512b makes only sense if you want to really use it.

Perhaps eg for professional market??

Did you miss the part where 512 bit bus is 20% smaller than the 384 bit bus in Tahiti( less complex PHYs) whilst providing an additional 11% bandwidth and maybe even power savings?

 

[Lonyo]

Maybe the PCB isn't the main cost at the moment, and the cost of the die is more significant. Savings on the die from being smaller are more than the additional cost on PCB.

Plus one might assume they would have gone to 6GB of RAM if they were on a 384-bit bus, and now they can get away with only 4GB, which is also cheaper.

Die saving + RAM saving is more than the additional PCB cost, meaning overall cheaper.

 

[Erenhardt]

Did you miss the part where 512 bit bus is 20% smaller than the 384 bit bus in Tahiti( less complex PHYs) whilst providing an additional 11% bandwidth and maybe even power savings?

Power what? /joke

It really is different. There are currently some problems with beta drivers. Memory clocks are jumping up and down when the card is idling, etc. Wonder if this has anything to do with "memory rewrite".

They had to do something. Before release some asked about pixie dust. Die size went up by less than 30%, shaders up by 40%, TMU + 37%, ROPs +100%, DP is now 1/2=1408 (50% disabled in 290x) from 1/4=512?.

 

[raghu78]

the hawaii memory controller is all about perf/mm2.

http://forum.beyond3d.com/showthread.php?p=1785506#post1785506

"Per 128b chunk the Pitcairn PHY is about half the size and Tahiti has 3 of them. "

hawaii's memory controller was basically a Pitcairn's memory controller doubled. with Hawaii at 438 sq mm AMD has effectively provided the same performance as Nvidia did with a 550 sq mm die. Hawaii is massive in terms of perf / sq mm. It helps AMD carve more chips out of a wafer and the yields are also better on a 438 sq mm chip when compared to a 551 sq mm chip.

AMD did not need any more bandwidth than 320 GB/s. so clocking a 512 bit memory controller at 5 Ghz and saving power made more sense. custom R9 290X cards like Toxic , Lightning can easily clock to 5.2 - 5.4 Ghz and bring close to 350 Gb/s bandwidth. btw remember that effective bandwidth is always better on a 512 bit memory controller at lower speeds than a 384 bit memory controller at higher speeds even if theoretically they are supposed to provide same bandwidth. nothing runs at 100% efficiency :biggrin:

 

[GaiaHunter]

Clearly for AMD die size is the main factor.
Otherwise they would have packed Hawaii transistors in a bigger but less dense pack (which would probably improve power consumption).

 

[brandonmatic]

Clearly for AMD die size is the main factor.
Otherwise they would have packed Hawaii transistors in a bigger but less dense pack (which would probably improve power consumption).

Makes sense. Anyone want to hazard a guess about how much more expensive each GPU die would have been if AMD went from 438mm to 550mm? That's 25% larger. Adding the higher defect rate you'd get on a larger die, what would the actual cost difference be? 40%? 50%?

 

[krumme]

Did you miss the part where 512 bit bus is 20% smaller than the 384 bit bus in Tahiti( less complex PHYs) whilst providing an additional 11% bandwidth and maybe even power savings?

Did you miss my first sentence?
"...by the reduction in diesize"

Take this advice: Try to understand before you try to be understood.

Did you even read the first post at all?