Ryzen 5 2400G APU And Memory Speed Analysis

rgMekanic

[H]ard|News
Joined
May 13, 2013
Messages
6,943
PC Perspective have run a battery of benchmarks on the new Ryzen 5 2400G APU to see what effect memory speeds have on performance. We know that Ryzen is more sensitive to memory speeds than what we were used to, but what about when you throw a GPU into the mix that is utilizing the same memory pool. The difference in both synthetic and game benchmarks was nearly 10-15% for DDR4-2400 vs DDR4-3200.

That is quite a bit of a performance impact. And the scaling seems to go up almost perfectly with the higher speed memory. As well it looks like if you are interested in a new Ryzen APU like I am, going dual channel memory is a must. Now if only memory prices weren't drunk right now. Thanks to cageymaru for the story.

For our testing, we are running the Ryzen 5 2400G at three different memory speeds, 2400 MHz, 2933 MHz, and 3200 MHz. While the maximum supported JEDEC memory standard for the R5 2400G is 2933, the memory provided by AMD for our processor review will support overclocking to 3200MHz just fine.
 
Last edited:
How ironic that you need to pony up for more expensive ram to better utilize a budget class APU. How much could 512mb or 1gb of ram on die actually cost anyway? $5? $10? That would be $10 I would gladly spend.
 
This is great info! I'll be building a 2400G system for my wife in the near future. I will definitely be going for some 3200 MHz memory.
 
you do realise the difference of 10% is the difference between 60 and 66 FPS?

Honestly, I don't think the more expensive RAM is worth it. Just grab the best bang-for-buck you can afford.
 
Interesting article and thanks for posting a link about it. I'd really like to see a comparison of those Intel CPU/AMD APU w/ HBM2 onboard partnered chips to these.
 
Love all the narrow-sighted posts complaining about using these APUs with faster RAM. Guess they are not seeing the [within reason] futureproofing of a box with the faster RAM. One less thing to deal with if you plan on moving up to a beefier Ryzen down the line, don't have to swap out RAM from slower to faster stuff.
 
Love all the narrow-sighted posts complaining about using these APUs with faster RAM. Guess they are not seeing the [within reason] futureproofing of a box with the faster RAM. One less thing to deal with if you plan on moving up to a beefier Ryzen down the line, don't have to swap out RAM from slower to faster stuff.

There's nothing short-sighted about RAM prices.
 
The speed increase look almost linear even at 3200 - so going faster like 3466 would be interesting - 3600 maybe even better but that seems to be the limit and few Ryzens can achieve. Not sure if the memory controller has been tweaked with Raven Ridge. So for those that can OC those 3200 modules, they will probably see even better performance.
 
AMD needs to figure out how to get HBM on-package and into an AM4 socket.

That's one of the main reasons AMD's APUs aren't worth it except in the most desperate situations: you have to spend more to get faster RAM, and those funds are better spend on discrete options and better CPUs.
4Gb of L-3 HBM?
That would be a start.
 
you do realise the difference of 10% is the difference between 60 and 66 FPS?

Honestly, I don't think the more expensive RAM is worth it. Just grab the best bang-for-buck you can afford.

yeah i'd say 3000 is probably the best bang for the buck.. some of the 3000mhz kits randomly go on sale for really cheap compared to the 3200 samsung B die kits. what i wish they had done in the tests is seeing what effect timing has on performance too see if there's a difference between say CL16 hynix and CL14 samsung B at 3000mhz.
 
yeah i'd say 3000 is probably the best bang for the buck.. some of the 3000mhz kits randomly go on sale for really cheap compared to the 3200 samsung B die kits. what i wish they had done in the tests is seeing what effect timing has on performance too see if there's a difference between say CL16 hynix and CL14 samsung B at 3000mhz.

From what I understand, graphics performance is not affected much by RAM Latency.
 
From what I understand, graphics performance is not affected much by RAM Latency.
In APU you have two processors needing access to the ram, I would think for an APU Latency could have a very significant impact. How much? No clue. Need some more indebt reviews looking at that.
 
AMD needs to figure out how to get HBM on-package and into an AM4 socket.

That's one of the main reasons AMD's APUs aren't worth it except in the most desperate situations: you have to spend more to get faster RAM, and those funds are better spend on discrete options and better CPUs.
the difference between 2x4gb 2400 vs 3200 isnt much more than 20 dollars...and I think most people buying the APU's are doing so because they don't want a discrete solution in the first place.
 
How ironic that you need to pony up for more expensive ram to better utilize a budget class APU. How much could 512mb or 1gb of ram on die actually cost anyway? $5? $10? That would be $10 I would gladly spend.

512-1gb ondie memory would be very expensive, more like 50$-100$ and it'd be slow...
I advice you to look at DIE shots and see how much space 8mb of memory takes of space.

Adding 8MB of L3 for the vega is core is the most feasable but it'd still be 20 bucks at least and it'd be almost half of the size of the Vega core alone...
Ondie is a dead idea, HBM chip next to it is the only possibility and 512mb-1gb would suffice if bandwidth is 150gb/sec

GTX1060 has 200gb/sec and Ryzen's L3 cache has 170gb/sec, increasing chip size is not an linear cost increase per increase in MM^2, the bigger the chip the lower the yield equals higher cost as more chips are scrapped and the utilization of the wafer is lower as well.
Yield:
There are 15 dots on a 300mm wafer, each dot represent a defect.
Each Die is 210mm^2 and since they're rectangles\squares you can see how many you can fit inside a circle and how much is lost because the wafer is a circle.
In addition how many chips will those Dots\defects hit?

Increase the Die size to 260mm^2 to add 8mb of fast vram to the vega, how many chips can the circle fit, and how many chips does those dots hit?
The amount of defective chips are usually the same, but there are fewer chips thus the percentage of good vs bad is worse and the unusable wafer area is increased because squares in a circle.

Another approach, HBM on AM4.
One HBM stack is 1gb/30mm^2, HBM2 is probably same for 22 MM^2, and integrating it would be a lot more than the cost of an HBM chip.
Not to mention memory controller ondie and suddenly you've added an 50$ memory chip, you've added 50$ in memory controller ondie, interposer for a Vega 11 adds cost too and all that just to compete with a GTX1050TI....

The fact is that AMD has TWO cpu dies for sale, TWO.
Epyc 32 core, -> Epyc 16 core.
Threadripper 8->16 Core.
Ryzen 4->8 Core
= One chip.

Ryzen 5&7 Mobile
Ryzen 3 2200G
Ryzen 5 2400G
= One chip.

I wonder where this new 300-500$ chip would fit in, how big the market is, how much benefit it is cause 100 people on the internet thinks giving the vega core more memory bandwidth resolves all issues in the world.

If it was so cheap, we'd see it, we'd see hilariously big cache sizes in cpu's and what not if Nvidia, AMD and Intel knew what the Internet knew :)

In the end: They have a Vega Mobile 24 CU which has HBM, it's small factor, it's Low power, It is efficient and it's not too pricey.
Nvidia has MX150 and a bunch of mobile chips which is good too.
 
Another explanation for the absence of hbm2 apuS, is they dont have any hbm2.

It may even be the converse. The only reason we already have apuS, is they were needed sooner to shift a glut of unsalable vega gpu cores, for want of ram.

APUs dont need vram.
 
AMD needs to figure out how to get HBM on-package and into an AM4 socket.

That's one of the main reasons AMD's APUs aren't worth it except in the most desperate situations: you have to spend more to get faster RAM, and those funds are better spend on discrete options and better CPUs.

everyone has known for a while that linked to DDR is what holds APU's back but also would you pay 300 bucks for an APU with a stack of HBM to get GTX 1050ti performance when for the same price you can get a 2200G with a 1060. I am sure there will be a time when AMD has its own EMIB style APU for desktop capable of delivering mainstream performance, but right now the evolution from say Kavari to now is massive, what Raven Ridge is a good indicator of is how badly Bulldozer was holding the GCN core back.
 
Its good that they did the testing, but lets be real: these computers are not going to speed-obsessed computer users.
I did some searches of DDR4 memory, and it felt a bit like an 'adventure'.
Also if you are going to modify the platform so much, as to make it the equivalent of having and add-in video card, might as well go for the add-on video card... but looking at video card prices... shiiit I can see why you would want this to perform like at least a lower tier video card.
 
Last edited:
everyone has known for a while that linked to DDR is what holds APU's back but also would you pay 300 bucks for an APU with a stack of HBM to get GTX 1050ti performance when for the same price you can get a 2200G with a 1060. I am sure there will be a time when AMD has its own EMIB style APU for desktop capable of delivering mainstream performance, but right now the evolution from say Kavari to now is massive, what Raven Ridge is a good indicator of is how badly Bulldozer was holding the GCN core back.

We knew Bulldozer was holding GCN back because it was a failed attempt at a top-tier architecture, and it had significantly lower bandwidth available.

RR is still bandwidth limited of course, and it's still a very limited niche offering.
 
We knew Bulldozer was holding GCN back because it was a failed attempt at a top-tier architecture, and it had significantly lower bandwidth available.

RR is still bandwidth limited of course, and it's still a very limited niche offering.
I think a billion people could be tempted by an apu (facebook is 2Bn btw), and thats not counting embedded/industrial apps.
 
Back
Top