Why did AMD clock the 290X memory so low?

[Actaeon]

I did some research and it seems like lots of folks are over 6+ghz pretty easily compared to 5ghz stock.

I can understand why they were not aggressive with the cores as their terrible reference cooler had a hard time even with stock speeds, but why did AMD spec the memory so low?

Couldn't they have done 5.5ghz or 5.8ghz? With some of the numbers I've seen floating around, even 6ghz seems reasonable. That would bring memory bandwidth close to 400GB/s. Seems like that would help for 4K, where it really gets to stretch it's legs.

 

[raghu78]

For power reasons. Running a 512 bit memory bus does requires a lot of power. The power consumption increase by running at 6 Ghz would not have been worth it.

 

[Gloomy]

A GPU's memory subsystem comprises almost half of the GPU's power consumption. Back in the 6xxx days it was something like 37%. I'm not sure what it is now but it's probably in the same ballpark.

It is therefore a juicy target for efficiency improvements. These can take a lot of forms, from new memory types (HBM) to just plain lowering the reliance on off-chip memory then making the memory subsystem leaner like Nvidia did with Maxwell.

When designing a GPU both companies try to find a balance between performance and bandwidth. They try to hit the sweet spot where the chip isn't held back by memory bandwidth. Giving it more bandwidth won't improve performance significantly but it will increase power consumption. And giving it too little bandwidth will hamper performance.

For Hawaii, AMD decided to go for a wide but slow memory setup. In comparison, the 780Ti from Nvidia went with a slimmer but faster setup. The results are similar, they just went with what fit their need best.

My 290 can't go past 1400MHz memclock before it needs a massive increase in voltage.

 

[Actaeon]

Got it, that makes sense. If they could have found efficiency improvements elsewhere, this seems like it was an area of opportunity to further showcase its performance capabilities at higher resolution gaming where that extra bandwidth will come in handy.

Thanks to you both for clearing that up.

 

[ocre]

I do find this thread interesting. But the suggested reasoning brings to me many more questions.

Here is where it takes me:

If it is all because of power consumption, how does nvidia keep this under control?

The immediate answer might be that nvidia uses smaller buses which gives them the ability to run it faster. Like there is some sort of trade off.

But then we have tonga.....
the r9 285

Its also running at very conservative speeds, especially in the light of Nvidia's designs.

So there is more to this story for sure

 

[Paul98]

R9 290x has a ton of memory bandwidth. It can be clocked slower because it's 512 bit others are 384 or 256.

 

[III-V]

I did some research and it seems like lots of folks are over 6+ghz pretty easily compared to 5ghz stock.

I can understand why they were not aggressive with the cores as their terrible reference cooler had a hard time even with stock speeds, but why did AMD spec the memory so low?

Couldn't they have done 5.5ghz or 5.8ghz? With some of the numbers I've seen floating around, even 6ghz seems reasonable. That would bring memory bandwidth close to 400GB/s. Seems like that would help for 4K, where it really gets to stretch it's legs.

A lot of the answers in this thread are a bit off.

AMD specifically used MCs on Hawaii that ran at lower clocks, but were denser. They were essentially able to get four somewhat slower memory controllers for the price of Tahiti's three fast ones.

AMD also seems to have inferior memory controllers in terms of clocking potential, compared to Nvidia. This holds true, even on Tahiti, where they were built for higher speeds than Pitcairn or Hawaii.

 

[Actaeon]

R9 290x has a ton of memory bandwidth. It can be clocked slower because it's 512 bit others are 384 or 256.

I agree the 290/x has a ton of bandwidth relative to other cards and doesn't need as much clock speed to generate such high levels of memory bandwidth. This is one of the reasons they've been able to do so well in higher resolution gaming even compared to the more modern and higher priced competition.

It just seemed like they could have had even more bandwidth quite easily. Without trying too hard, I just set mine at 6.5Ghz. Thats an extra 1.5ghz and an extra 100GB/s of bandwidth over stock. If if they only got half that, that seems like a lot of extra performance to leave on the table.

A lot of the answers in this thread are a bit off.

AMD specifically used MCs on Hawaii that ran at lower clocks, but were denser. They were essentially able to get four somewhat slower memory controllers for the price of Tahiti's three fast ones.

AMD also seems to have inferior memory controllers in terms of clocking potential, compared to Nvidia. This holds true, even on Tahiti, where they were built for higher speeds than Pitcairn or Hawaii.

While that may be true, I'm skeptical that is the explanation or the answer to my question.

I've seen that review units and in my own personal experience, this card seems to be able to get much higher clocks than what AMD set them at. If the reason is that the Memory Controllers on the cards are poor, then why are owners able to set them so much higher than stock easily? I would expect if they were that inferior and it was close to the limit at stock speeds, we wouldn't be able to get much additional headroom, but its quite the opposite. Given how easy it was to go as high as I did, I have to assume its not a technical limitation.


With a 512bit bus, raising the clock speed is even an bigger benefit. Again, even half of what I got is another 50GB of memory bandwidth for nearly 400 total.

 

[Z15CAM]

Alpida Vs Hynix and Samsung.

Whatever happened to Mosel?

Today I'm forced to run Alpidia within cost at the specified or slightly OC'd voltage/frequency and timing for stability - AMD or nVidia.

Who's buying all the best Ram - NSA, CSIS or NSA maybe NATO or the KGM but I bet it's going to the Highest Bidder - Really what happened to the availability of the Samsung MV-3V4G3D-US_DDR3 sticks ;o)

 

[III-V]

While that may be true, I'm skeptical that is the explanation or the answer to my question.

I've seen that review units and in my own personal experience, this card seems to be able to get much higher clocks than what AMD set them at. If the reason is that the Memory Controllers on the cards are poor, then why are owners able to set them so much higher than stock easily? I would expect if they were that inferior and it was close to the limit at stock speeds, we wouldn't be able to get much additional headroom, but its quite the opposite. Given how easy it was to go as high as I did, I have to assume its not a technical limitation.

With a 512bit bus, raising the clock speed is even an bigger benefit. Again, even half of what I got is another 50GB of memory bandwidth for nearly 400 total.

It's hard to apply overclocking to how much "headroom" there is for a company to raise their factory clocks. There are a number of reasons for this.

Higher clocks kill yields -- you have a lower percentage of chips capable of hitting the higher clock target. This is countered by offering different SKUs that use the same silicon, it seems that it's only really economical to have around 3 for a high end design.

Power consumption is another reason, perhaps the biggest. AMD and Nvidia work within a power budget, and logically, AMD and Nvidia chose a combination of core clock, core configuration, bus width, and memory clock that they deemed most optimal for their architecture. If you were to bump up the memory clock, you'd have to make cuts elsewhere to keep the design under 250W (or whatever the TDP is for the GPU).

There's also reliability issues. AMD and Nvidia have much higher standards for stability, and they need to ensure that their designs last for their targeted lifetimes. RMA rates also need to be kept below a threshold, which higher clock speeds would work against, with RMAs increasing exponentially with increased clock speeds (primarily because of higher heat production).

 

[rtsurfer]

A lot of the answers in this thread are a bit off.

AMD also seems to have inferior memory controllers in terms of clocking potential, compared to Nvidia. This holds true, even on Tahiti, where they were built for higher speeds than Pitcairn or Hawaii.

I take issue with this.
Do you have any facts to support your argument.?

 

[III-V]

I take issue with this.
Do you have any facts to support your argument.?

Go look at techpowerup, and look at their overclocking results. They've reviewed a fairly large number of cards from both companies. The trend is very clear.

 

[rtsurfer]

Go look at techpowerup, and look at their overclocking results. They've reviewed a fairly large number of cards from both companies. The trend is very clear.

So AMD keeps their Memory Clocked lowered than Nvidia, that means that they have a weaker memory controller.?

Can't they just be using a different approach to satisfy their bandwidth needs.

 

[III-V]

So AMD keeps their Memory Clocked lowered than Nvidia, that means that they have a weaker memory controller.?

Can't they just be using a different approach to satisfy their bandwidth needs.

What? I specifically stated overclocking.

It's visible in factory clocks as well, but regardless, bandwidth needs aren't relevant. Nvidia's memory controllers are capable of clocking signicifantly higher than AMD's. That's a fact. Historically, the opposite was true. Nvidia struggled quite a bit with memory clocks on Fermi, for example.

 

[rtsurfer]

What? I specifically stated overclocking.

It's visible in factory clocks as well, but regardless, bandwidth needs aren't relevant. Nvidia's memory controllers are capable of clocking signicifantly higher than AMD's. That's a fact. Historically, the opposite was true. Nvidia struggled quite a bit with memory clocks on Fermi, for example.

Why aren't Bandwidth needs relevant.?
What else do you gain by higher speed memory on GPUs.?

 

[Actaeon]

It's hard to apply overclocking to how much "headroom" there is for a company to raise their factory clocks. There are a number of reasons for this.

Higher clocks kill yields -- you have a lower percentage of chips capable of hitting the higher clock target. This is countered by offering different SKUs that use the same silicon, it seems that it's only really economical to have around 3 for a high end design.

Power consumption is another reason, perhaps the biggest. AMD and Nvidia work within a power budget, and logically, AMD and Nvidia chose a combination of core clock, core configuration, bus width, and memory clock that they deemed most optimal for their architecture. If you were to bump up the memory clock, you'd have to make cuts elsewhere to keep the design under 250W (or whatever the TDP is for the GPU).

There's also reliability issues. AMD and Nvidia have much higher standards for stability, and they need to ensure that their designs last for their targeted lifetimes. RMA rates also need to be kept below a threshold, which higher clock speeds would work against, with RMAs increasing exponentially with increased clock speeds (primarily because of higher heat production).

I understand that higher clock reduces usable yields, but I haven't seen a review where they were unable to hit 5.5GHz memory, with many over 6GHz. Some even higher. 5.5Ghz would have given them 10% more bandwidth and still a great usable yield for 290Xs. If any yield didn't make the cut could have been used on the standard 290 or other cards.

From all the points presented I've seen and given how 'hot' the 290 series runs, power consumption seems to be the most likely answer here and matches what the others have said. I do find it a little strange you bring it up since you seemed to initially disagree with this and stated it was the memory controller. However, to your point, if they have to keep it under 250W, then perhaps reducing the memory clocks had the least repercussions. This seems logical.

I agree that OEMs need to have a higher standard for reliability than a consumer would, as they are on the hook if it fails. Though I do think a modest increase (10% as noted earlier) wouldn't create any meaningful impact to the failure rate while still delivering a performance boost. Heat is a concern and a better reference cooler for 10% more performance seems like a good trade to me.

 

[Z15CAM]

I haven't come across an OEM Elpida 290X AMD card that will not do 1500MHz's vRam frequency at say +0.143Mv's running at l130MHz's core.

Put her under water with Core and VRM at the above frequency temps never exceed 70C and never Thermal Down Clock.

With a 512 Bus Hawaii is not too shabby.

 

[Erenhardt]

yea... new folklore:"amd memory controllers sux!"
http://www.techpowerup.com/reviews/AMD/R7_260X/29.html

GPU clock reaches exactly the average clock for Bonaire. Memory overclocks much better than on HD 7790, thanks to the new high-frequency GDDR5 memory chips used, but I could not overclock the memory beyond 1800 MHz, which looked more like a software issue than a limitation of the memory itself.

512 bit... pfff Give me some of that sweat 4096 bit bus:
http://www.kitguru.net/components/g...ay-feature-4096-sps-extreme-memory-bandwidth/

 

[3DVagabond]

What? I specifically stated overclocking.

It's visible in factory clocks as well, but regardless, bandwidth needs aren't relevant. Nvidia's memory controllers are capable of clocking signicifantly higher than AMD's. That's a fact. Historically, the opposite was true. Nvidia struggled quite a bit with memory clocks on Fermi, for example.

They couldn't be using lower spec RAM? Tighter timings? Face it, you are taking a single metric and applying your own solution to it.

 

[ShintaiDK]

It doesnt really matter. GPUs mainly just want pure bandwidth.

This is also why stacked DRAM is merely an evolution on GPUs rather than a revolution. If there was a performance gain in it, then AMD would have gone with faster GDDR as we see with 256bit. Hell even on Tonga they didnt go full speed on 256bit.

So the short answer is simply that the 290X doesnt need faster memory.

 

[III-V]

Basically, AMD already gets higher Bandwidth than Nvidia because of the wider-bus, what is the point in clocking their memory higher.?

Because the number of units that need that bandwidth, ROPs, are generally tied to the memory bus width in a fixed ratio. They need the wider bus to feed the ROPs. Today's workloads are pretty pixel heavy, from what I've read.

They couldn't be using lower spec RAM? Tighter timings? Face it, you are taking a single metric and applying your own solution to it.

And you're getting needlessly defensive over a fact that was presented to be informative, not pejorative. Relax, please. It is okay for companies to shine in some areas and not in others.

 

[rtsurfer]

Because the number of units that need that bandwidth, ROPs, are generally tied to the memory bus width in a fixed ratio. They need the wider bus to feed the ROPs. Today's workloads are pretty pixel heavy, from what I've read.

Hmmn.
Now that is an interesting point.


Although we would need an evidence that the ROP's in AMD cards have a significant memory bottleneck, that a higher memory clock will help eradicate.

I guess something like this can be tested by increasing the RAM speed while keeping the core clock constant.

 

[III-V]

Hmmn.
Now that is an interesting point.


Although we would need an evidence that the ROP's in AMD cards have a significant memory bottleneck, that a higher memory clock will help eradicate.

I guess something like this can be tested by increasing the RAM speed while keeping the core clock constant.

Yeah, but no site bothers doing things like that, unfortunately. I love that kind of investigative work.

 

[rtsurfer]

Yeah, but no site bothers doing things like that, unfortunately. I love that kind of investigative work.

Interesting excerpt on the ROP issue from Ryan Smith's Hawaii launch article,

On the other hand GPU clockspeeds on 290X are being held consistent versus the recently released 280X, with AMD shipping the card with a maximum boost clock of 1GHz (they’re unfortunately still not telling us the base GPU clockspeed), which means any significant performance gains will come from the larger number of functional units. With that in mind we’re looking at a video card that has 200% of 280X’s geometry/ROP performance and 138% of its shader/texturing performance. In the real world performance will trend closer to the increased shader/texturing performance – ROP/geometry bottlenecks don’t easily scale out like shading bottlenecks – so for most scenarios the upper bound for performance increases is that 38%.

So I guess the doubling of ROP from Tahiti, was maybe properly handled by the Bandwidth increase from 288GB/sec to 320GB/sec.

http://www.anandtech.com/show/7457/the-radeon-r9-290x-review


I can run some tests later on in the week if you want, my 290X goes to 1650-1700Mhz memory.

 

[Wall Street]

With a 512 bit memory bus vs. a 256 bit memory bus on the GTX 980 and a 384 bit memory bus on the R9 280X, the R9 290X has much more memory bandwidth. If the R9 290X had double the shaders and double the ROPs of those other chips, that than having double the memory bandwidth would make sense. However, the R9 290X only has 37.5% more shaders than the GTX 980 and R9 280X, so extra memory bandwidth doesn't contribute much past a certain point. I bet that AMD tested the chip with 6.5 Ghz memory

Here is a video of a guy who tests an R9 290 with 1350 Mhz RAM vs. 1700 Mhz RAM (i.e. 5.2 Ghz vs. 6.8 Ghz). As you can see, the results are less than 10% from a 30% overclock.