Memory Scaling, AMD's Trinity-Based APUs, And Game Performance

Page 2 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Status
Not open for further replies.

disolitude

Honorable
Aug 7, 2012
40
0
10,540
3
[citation][nom]disolitude[/nom]This is very neat article. I do wish you included the 6670 APU crossfire in these benchmarks. I've never seen the effect of memory speed on gaming performance with a 6670 crossfire.[/citation]

Actually it's an epic win when you want to build a 250W m-ITX build and would like some gaming performance in a small low power package.
 
G

Guest

Guest
In my testing, dual-rank (4 sticks) is another 50% performance bump up to at least 1866.
If this is an exploration of memory bandwidth scaling, not just frequency, you may want to run those numbers as well.
 

matthelm

Distinguished
Mar 26, 2009
40
0
18,530
0
Anyone know if the A10 memory interface speed increases if you go from 2 slots of RAM to 4? I've never seen anyone remark on that.
 

oxiide

Distinguished
Nov 8, 2011
850
0
19,360
108
[citation][nom]Mansfield[/nom]is ares better than ripjaws x? i have 16GB of 1600 ripjaws[/citation]
I think they're essentially the same chips. Ares is lower-profile though, whereas Ripjaws have unnecessarily tall heatsinks that can get in the way of certain CPU coolers.
 
[citation][nom]Memnarchon[/nom]Actually the growth (in power) of the APUs make them need faster memory bandwith. They are screaming for the DDR4 to become available at the market.[/citation]

Exactly. When the GCN APUs come out with something like GPU performance comparable to a Radeon 7750's cut-down Cape Verde GPU, DDR4 may be the only way for regular system memory to feed them without a greater bottle-neck than the Radeon 7750's own memory. DDR3 certainly can't do the job except in a very wide bus and that'd probably be expensive.
 
[citation][nom]blibba[/nom]Hasn't this topic been done to death already? I still don't understand Trinity for gaming, you're generally better with an older dual-core and discrete card.That's what Trinity is.[/citation]

You generally aren't better with an older dual-core and discrete card. Trinity is often cheap enough that even if you don't include its IGP and use it purely as a CPU, it still is worth the price and the IGP is just the icing on the cake for value. For example, as a CPU, the A10-5800K is almost as fast as the FX-4300 and FX-4170, yet it's about the same price as them.

It includes an IGP that's easily worth another $40-50 as it's comparable to the GPU of the Radeon 6670 and with decent memory, has performance comparable to the Radeon 6670 DDR3 which goes for around those prices. Trinity is a great value for the money and only loses to a cheaper dual core when the Radeon 6670 GDDR5 is on sale and even then, that's only a lose if you don't mind the trade-off in generally lower CPU to greater graphics performance.
 
G

Guest

Guest
Thanks for doing this article! I had no idea that you were already planning on doing this when I (anonymously, as I am now) suggested it a week-or-so ago. For informative! Special thanks to the poster who mentioned the Von Neumann bottleneck. That just explained everything.

Love this site! Thanks guys.
 

mohit9206

Distinguished
Aug 8, 2012
1,913
0
20,160
120
[citation][nom]blazorthon[/nom].It includes an IGP that's easily worth another $40-50 as it's comparable to the GPU of the Radeon 6670 and with decent memory, has performance comparable to the Radeon 6670 DDR3 which goes for around those prices. Trinity is a great value for the money and only loses to a cheaper dual core when the Radeon 6670 GDDR5 is on sale and even then, that's only a lose if you don't mind the trade-off in generally lower CPU to greater graphics performance.[/citation]
the i gpu of A10-5800k is at best comparable to a 6570 not 6670. so a intel dual core with 6670 will be better than A10 in itself unless the game takes full advantage of all 4 cores
 

tmk221

Distinguished
Jul 27, 2008
173
0
18,690
4
[citation][nom]esrever[/nom]Can't wait to see what DDR4 can do for APUs.[/citation]

I was kind of doubting APU, but once we will have ddr4 who knows?
 
G

Guest

Guest
How about GDDR5 ram and Trinity???

I was thinking you could create a hybrid motherboard that would have 2 or 4 DDR slots and 1 GDDR slot in it.

You can use the DDR ram if the GDDR slot is empty, but if you wanted even more performance put in one stick if GDDR ram. I do not know how much the GDDR ram would cost though.
 
[citation][nom]mohit9206[/nom]the i gpu of A10-5800k is at best comparable to a 6570 not 6670. so a intel dual core with 6670 will be better than A10 in itself unless the game takes full advantage of all 4 cores[/citation]

What makes you say that the GPU is weaker than the 6670's GPU?

Also, many modern games can work with four threads effectively. Tom's went as far as adding in a recommendation for an AMD quad core and stating this as the reason for the current best gaming CPUs for the money article.

[citation][nom]richardginn[/nom]How about GDDR5 ram and Trinity???I was thinking you could create a hybrid motherboard that would have 2 or 4 DDR slots and 1 GDDR slot in it.You can use the DDR ram if the GDDR slot is empty, but if you wanted even more performance put in one stick if GDDR ram. I do not know how much the GDDR ram would cost though.[/citation]

There aren't any GDDR5 memory modules and the APU market probably isn't big enough for enough incentive to make them, so its probably better to wait for DDR4.
 

RedJaron

Splendid
Moderator
That is an interesting question: how much is someone willing to spend on RAM to pair with a $130 CPU? I wonder if there's a sweet spot around $300 for an A10, 7770, and RAM. Something that's affordable for anyone, but still gives decent performance in some of the less demanding games. I'm guessing the hybrid crossfire would yield better results than an i3 + 7770 and cost just a little less.
 

artk2219

Distinguished
Jun 30, 2010
480
42
18,890
30
[citation][nom]richardginn[/nom]How about GDDR5 ram and Trinity???I was thinking you could create a hybrid motherboard that would have 2 or 4 DDR slots and 1 GDDR slot in it.You can use the DDR ram if the GDDR slot is empty, but if you wanted even more performance put in one stick if GDDR ram. I do not know how much the GDDR ram would cost though.[/citation]


Motherboard makers used to integrate small amounts of GDDR onto the motherboard, it was called sideport memory, Toms did an article on it several years ago.

http://www.tomshardware.com/reviews/790gx-graphics-sideport,2088.html

edit: It was DDR2 or 3 that was available as sideport memory, thanks for the correction blazorthon!
 


As I recall, the integrated memory was just regular DDR2 or DDR3 memory that was allocated to the IGP so that it woudn't need to use the main system memory. I don't think that it was a GDDR memory. For example, my 890GX motherboard has 512MiB of DDR3 side port memory.

I think that it may be possible to use GDDR5 in such a way and it's been suggested before (by myself and I've read several other people mention it here at Tom's), but I don't know if it's been done before except with regular system memory chips.
 
Don't think that will happen, Blaze. The GCN cores should ultimately use a common memory controller and would not need the 'sideport'

With the HSA, UNB and unified memory the *SIMD Engine Array* cores will have direct access to it's own mapped system memory (with the high-end processors gaining direct access to CPU L3 cache,?).

Yah never know with AMD's ultimate "execution" but it appears to me when combined with DDR4 the overall package may be LOL fast ...



 

Bananaman8910

Distinguished
Aug 10, 2011
10
0
18,520
1
[citation][nom]RedJaron[/nom]That is an interesting question: how much is someone willing to spend on RAM to pair with a $130 CPU? I wonder if there's a sweet spot around $300 for an A10, 7770, and RAM. Something that's affordable for anyone, but still gives decent performance in some of the less demanding games. I'm guessing the hybrid crossfire would yield better results than an i3 + 7770 and cost just a little less.[/citation]

Unfortunately the current generation of APU's can't be crossfired with anything larger than a 6670. A 7770 could be used as a discrete card, but that kinda defeats the entire point of an APU. You'd be better off with an i3 and a discrete 7770 in most cases.

There are reports that the next generation of APU will be able to crossfire with more modern cards, including the 7770.

Toms recently did an article on this:
 


I agree that DDR4 should negate the usefulness of GDDR5 sideport memory. It was merely something that we talked about back around the launch of Llano when DDR4 seemed too far in the future to consider in the short term as a solution to the IGP's hunger for memory bandwidth.
 

merikafyeah

Honorable
Jun 20, 2012
264
0
10,790
2



"The biggest bottleneck in any architecture is shared communication between all components"

I know, but there isn't just one way data travels between components. A "collective bottleneck" is hardly a bottleneck so much as a limitation of the entire system. It's like saying "your system is limited by your system's limitations". PCI-Express for example, is just one specific component, the interface which connects a CPU to its internal caches is another, and so is the component which connects the CPU to the RAM, which was primarily the FSB before CPUs integrated that function on-dye.
What you're saying is that the collective limitations of PCI-Express+internal CPU interface+FSB+SATA+everything-else-I-missed = ONE "bottleneck".

I was referring to more specific bottlenecks, rather than one giant generalization. Other benchmarks have proven that PCI-Express is by far the least limiting factor of overall system performance than say, SATA. So in terms of priority in increasing performance SATA>PCI-Express. In the same way, in prioritizing the increase in performance of a consumer gaming PC with a discrete GPU the order is generally, GPU>CPU>RAM because of the simple fact that games are more GPU bound than CPU-bound, and likewise more limited by CPU frequency than RAM frequency. Discrete GPUs are not really affected by RAM speeds since they have their own large store of VRAM unlike the CPU which has very tiny caches. APUs are different since they're basically GPUs+CPUs in one, but without the large VRAM. Thus they are more reliant on faster system RAM which now has to do double-duty as the VRAM and the system RAM at the same time, so naturally higher RAM speed = higher performance, but as we've seen, only up to a certain point, after which further increases result in declining returns.

"Hypothetically, if you were to continue to increase the data rate of the system's memory you would see performance gains up until the point where the GPU's instructions units can no longer make use of the available interface."

Again, I didn't say there wouldn't be any increase. I said there wouldn't be any significant increase, or does my understanding of "diminishing returns" differ from yours? I draw my assertions from the numbers presented in the article, not hypothetical conjecture or theoretical maximums. Did you not notice how increasing the rate from 2133 to 2400 yields a difference that's practically margin of error? Not only that, there's a "trend" of decreasing performance gains in each successive bump in frequency. Going from 1600 to 2133 is a difference of 533 and yields a significant increase. But does going from 2133 to 2400 yield precisely half the increase of going from 1600 to 2133? Nope. There is a massive drop in performance gained per increase in MHz.
The numbers don't lie.

"Having said all that, until DDR-4 is out we can't say for certain that it will not have a huge impact on both AMD and Intel systems. This is because if DDR-4 manages to lower latency or greatly increase bandwidth you will see gains, especially if DDR-4 is able to achieve both lower latency and higher bandwidth at the same time."

There is fundamentally no difference between DDR, DDR2, and DDR3. Each successive generation increases the bandwidth respective of latency, and perhaps lowers power consumption a bit. That's it. If you're expecting DDR4 to radically change this scheme, prepare to be disappointed. DDR4-2800 won't be any different from DDR3-2800.

Hypothetically speaking, if you were to snatch some DDR4-4266 memory from the future and plopped it into today's APU benchmark system with a magical BIOS update that gets everything working as it should, will you see double the performance of DDR3-2133? If real-world results mean anything, the answer is a resounding NO. THIS is what I mean by RAM speed stops being the primary bottleneck in today's APU systems at around DDR3-2133. If DDR4 suddenly arrived tomorrow, will it make any difference with current APUs? Based on what we've seen, (wishful thinking aside) definitely not. Just like if PCI-Express 4.0 arrived tomorrow, will we see any difference with current discrete GPUs? Nope. How do we know this?
Because going from PCI-E 3.0 x8 to PCI-E 3.0 x16 yields no appreciable gain whatsoever, and that's doubling the bandwidth.

Of course APUs of the future will most likely benefit from faster DDR4 RAM. That's just stating the SUPER OBVIOUS. Just like saying "the GPUs of the future will benefit from PCI-Express 4.0".
I'm not denying the SUPER OBVIOUS.

"Oh and, to correct your first inaccuracy, DDR-4 will be lower power than what is currently available so it will use less electricity than DDR3-2400 therefore providing more performance per Watt of energy used."

I don't think you read what I said correctly. Read it again. Slowly this time. I said DDR4 won't provide any significant savings in energy consumption since RAM in general consumes negligible amounts of electricity.
What proof do I have? Just the readings of something called a "Kill-A-Watt" meter. Going from regular voltage DDR3 RAM to Low-voltage DDR3 RAM lowers the annual cost of your PC's electricity consumption by how much? Zilch. Nada. Nothing.
Even if RAM used ZERO watts, you STILL wouldn't save anything off your electric bill. It's that negligible.
Do you understand what I'm saying now?
 

RedJaron

Splendid
Moderator

Read the article, couldn't remember what the limit was on discrete cards. Thought it could go up to a 67xx/77xx card. So never mind on that experiment.
 
Uhh, dude?



I don't have a dog in this fight, but 2400MHz DDR3 effectively maxes-out the bandwidth available at 128-bit.

If you're expecting DDR4 to radically change this scheme, prepare to be disappointed. DDR4-2800 won't be any different from DDR3-2800.
SWING! and a miss.

... DDR3-2400 is only a tiny fraction better than DDR3-2133
Strike THREE!








 

A Bad Day

Distinguished
Nov 25, 2011
2,256
0
19,790
2
Throw in OCing and RAM bandwidth will become even more important.

[citation][nom]merikafyeah[/nom]"There is fundamentally no difference between DDR, DDR2, and DDR3. Each successive generation increases the bandwidth respective of latency, and perhaps lowers power consumption a bit. That's it. If you're expecting DDR4 to radically change this scheme, prepare to be disappointed. DDR4-2800 won't be any different from DDR3-2800.Hypothetically speaking, if you were to snatch some DDR4-4266 memory from the future and plopped it into today's APU benchmark system with a magical BIOS update that gets everything working as it should, will you see double the performance of DDR3-2133? If real-world results mean anything, the answer is a resounding NO.[/citation]

DDR2-1066 has 8533⅓ MB/s peak transfer.

DDR3-1066 has 8533⅓ MB/s peak transfer.

DDR3-2133 has 17066⅔ MB/s peak transfer.


Sure, DDR4-2133 might have the same bandwidth as the DDR3 version, but the DDR4 can clock up to around 4000.
 

merikafyeah

Honorable
Jun 20, 2012
264
0
10,790
2



I too, am aware of theoretical maximums. Are you however, aware that theoretical maximums don't often translate into real-world gains? I know DDR4 can go to 4000+, what's your point?
 

A Bad Day

Distinguished
Nov 25, 2011
2,256
0
19,790
2
Given that the Trinity APU (at least the A10) still shows increase in performance even when going from 2133 to 2400, DDR4 would allow even more increase in performance.

My laptop's Radeon 5730m has 25.6 GB/s of DDR3 bandwidth, and it's only somewhat better than an A10 4600m (7660G) GPU. For comparison, Anandtech estimated that the 4600m is only around 33% slower than the 5800k.

When I push the Radeon 5730's memory clock from 800 MHz to 1100 MHz, I could increase anti-aliasing or anisotropic filtering with minimal decrease in FPS.
 
Status
Not open for further replies.

ASK THE COMMUNITY

TRENDING THREADS