AMD Raven Ridge 8GB vs. 16GB Reserved Memory Benchmark & Explanation

[CheekyP]

The RX550 has plenty of bandwith relative to its (for dGPU standards) low shader/texture/compute units. This is why it pulls ahead of the APUs at 1080P but at 720P sees it perform closer to the APUs, as the bandwith requirements are lower at 720P.

Anyhow, good write up Steve, it's good to see that 8GB of RAM won't degrade gaming performance as I feel that is what most of these APUs will end up running, especially with current DDR4 prices.

I just hope people buying the 2200G/2400G realise that you need at minimum, dual channel RAM, and preferably DC DDR4-3200 for optimal performance, but I bet a lot of people will unwittingly try to save a few dollars and get a single DDR4-2400 stick... in which case you are essentially back to Intel levels of graphics performance.

 

[gamoniac]

This is a good read. Thanks, Steve.

 

[msroadkill612]

"Using the AIDA64 GPGPU benchmark tool we can measure read and write performance between the CPU and GPU,

effectively measuring the performance the GPU could move data in and out of its own device memory into the system memory. This is also called Device-to-Host Bandwidth...... "

Well yes, except for a fundamental logic fail here afaik.

In theory, max effective System memory speed for the gpu IS NOT 35GB/s. It is theoretically limited to the GPU's connecting bus speed to system "ram" - 16GB/s for 16 lane, 8GB/s for 8 lane.

The results here belie this. Why is it so?

There may be a good story in it, because further inquiry may reveal hidden effectiveness of Fabric. (yes, similar results are obtained by the discrete gpu here, but that's a distraction from the hot topic of apuS ATM.)

Both apuS here, get a ~33GB/s result, or 4x their theoretical 8 lane max pcie bandwidth?

In short, it would be good & v relevant to know if these gpu<>RAM Fabric links on apu's can outperform their nominal 8 lane pcie link rating?

 

[Dimitrios]

To the writer of this article and being a reader on many tech sites for over 20yrs, I just wanna say I'm impressed. Thank you and this is coming from someone that is too lazy to comment (lol) and even though I would not purchase an APU I appreciate your hard work!

 

[Steve]

"Using the AIDA64 GPGPU benchmark tool we can measure read and write performance between the CPU and GPU,

effectively measuring the performance the GPU could move data in and out of its own device memory into the system memory. This is also called Device-to-Host Bandwidth...... "

Well yes, except for a fundamental logic fail here afaik.

In theory, max effective System memory speed for the gpu IS NOT 35GB/s. It is theoretically limited to the GPU's connecting bus speed to system "ram" - 16GB/s for 16 lane, 8GB/s for 8 lane.

The results here belie this. Why is it so?

There may be a good story in it, because further inquiry may reveal hidden effectiveness of Fabric. (yes, similar results are obtained by the discrete gpu here, but that's a distraction from the hot topic of apuS ATM.)

Both apuS here, get a ~33GB/s result, or 4x their theoretical 8 lane max pcie bandwidth?

In short, it would be good & v relevant to know if these gpu<>RAM Fabric links on apu's can outperform their nominal 8 lane pcie link rating?

I had a bit of a hard time following this comment but I don’t think there is a fundamental logic fail here at all.

“Shifting data in and out of system memory is significantly slower than VRAM. The Raven Ridge APUs for example are limited to a memory bandwidth of around 35GB/s for system memory when using DDR4-3200. So in the case of the RX 550, it has a bandwidth of 112GB/s when accessing data locally using the VRAM, but when accessing data from system memory it's limited to 16GB/s (PCIe 3.0 x16 limit), which is to say that it takes at least seven times longer to process the same data.”

The Vega 8 and 11 GPUs don't use the PCIe bus but I think you realize that based on your comment, I'm just missing whatever the point is sorry.

To the writer of this article and being a reader on many tech sites for over 20yrs, I just wanna say I'm impressed. Thank you and this is coming from someone that is too lazy to comment (lol) and even though I would not purchase an APU I appreciate your hard work!

Thank you for stopping by and leaving a comment, much appreciated!

 

[Cooe14]

Excellent write up Steve. This will be very handy for people with only 8 GB of system Ram.
When using an APU, you have as much vRam as you do system Ram. They are the same speed, plain and simple!

It is important to note that BF1 used only 7 GB combined at 1080p. The only instance where 8 GB may suffer, is with COD: WW2. For eSport games, 8 GB will be plenty.

I concluded this using your other material, but posted on a different site:

https://hardforum.com/threads/reviews-for-amds-apu-ryzen-2400g-are-in.1954364/page-3#post-1043491136

Please let me know if everything was properly sourced / linked!

Thanks mate. Note VRAM usage is higher in the other test because we aren't using low settings. The GTX 1060 is much much more powerful than the Vega 8 GPU in the 2200G. That said with ultra quality textures and everything else set to low we still got the same results with the Ryzen 3 2200G, that is to say allocation size made no difference.

Hey Steve, I was wondering if you could possibly answer a question of mine that's been driving me crazy for FOREVER now. Why do the memory bandwidth benchmarks in SiSoft Sandra & AIDA64 report such drastically different results??? AIDA's "Cache & Memory Benchmark" tends to report values near exactly the same as those reported in the official JEDEC figures (I.e. DDR-3200 = 25GB/s per channel, AIDA's results on my Ryzen 7 PC with dual-channel 3200MHz ≈49.5GB/s reads & writes, and slightly less for copies), whereas Sandra's are DRAMATICALLY less, with my results falling around much the same area as yours, aka ≈35GB/s. Why the massive discrepancy?

* Also, it appears that not all boards handle this setting the same way, as I have seen the same exact setting have some pretty freaking dramatic effects on performance depending on what it's set to in the testing of other highly reputable tech reviewers, whereas it's obviously not for you. Super weird sauce.

 

[Steve]

Hey Steve, I was wondering if you could possibly answer a question of mine that's been driving me crazy for FOREVER now. Why do the memory bandwidth benchmarks in SiSoft Sandra & AIDA64 report such drastically different results??? AIDA's "Cache & Memory Benchmark" tends to report values near exactly the same as those reported in the official JEDEC figures (I.e. DDR-3200 = 25GB/s per channel, AIDA's results on my Ryzen 7 PC with dual-channel 3200MHz ≈49.5GB/s reads & writes, and slightly less for copies), whereas Sandra's are DRAMATICALLY less, with my results falling around much the same area as yours, aka ≈35GB/s. Why the massive discrepancy?

* Also, it appears that not all boards handle this setting the same way, as I have seen the same exact setting have some pretty freaking dramatic effects on performance depending on what it's set to in the testing of other highly reputable tech reviewers, whereas it's obviously not for you. Super weird sauce.

AIDA64 reports maximum throughput where as SiSoft Sandra I feel is a bit more realistic as it reports sustained throughput. It’s a lot like how SSD makers say their drives are good for 550MB/s read and 500 MB/s write but those are the peak throughputs, you’ll never ever sustain that in a file transfer for example.

Now for the second part of your question, what board are you using and where are you seeing a pretty freaking dramatic effect? I’ve tried an Asrock, MSI and Gigabyte board now, all provided the exact same results.

Also which other highly reputable tech reviewers are showing differences between the various frame buffer sizes? Differences greater than 3-5% that is. So far I’ve only seen tests confirming my findings.

 

[Cooe14]

AIDA64 reports maximum throughput where as SiSoft Sandra I feel is a bit more realistic as it reports sustained throughput. It’s a lot like how SSD makers say their drives are good for 550MB/s read and 500 MB/s write but those are the peak throughputs, you’ll never ever sustain that in a file transfer for example.

Now for the second part of your question, what board are you using and where are you seeing a pretty freaking dramatic effect? I’ve tried an Asrock, MSI and Gigabyte board now, all provided the exact same results.

Also which other highly reputable tech reviewers are showing differences between the various frame buffer sizes? Differences greater than 3-5% that is. So far I’ve only seen tests confirming my findings.

Not my board hahaha. I haven't purchased Raven Ridge yet, but will be soon! (A mITX backpack fitting totally OC'd Ryzen 2200G budget LAN box will be the perfect complement to my massive, all liquid AMD R7/Fury X desktop). I'm just referring to stuff I'd seen or read. And I can't quite remember exactly which exactly did the testing on 1GB vs 2GB, but I do remember the weird results in question either compared to other testing they did, or other's using the official test kit coming from people using currently possessed boards (for ex. 1% & .1% lows plummeting with the 2400g vs the 2200g in many games for a person testing at 1GB, which isn't present in other people's results using the AMD sent board that defaults to 2GB for instance despite otherwise similar results), but I'll get back to you on that after I get a chance to retrace some steps so I can be more specific .

And much thanks! That actually explains it perfectly, as I had too much trust in both the respective companies to believe either'd ever put out even remotely factually incorrect results; hence my confusion hahaha. (One using peak, and the other sustained though; freaking duh doi, of course that's what it is)!

 

[Danny101]

"You would then force people to spend a lot of extra money on a motherboard for a feature that may not be used later. This defeats the purpose of an APU - an inexpensive package with descent graphics This works for laptops, but not desktops."

Just looking forward to a stronger integrated solution. In this case, memory is the bottleneck. No complaints really. Just observing the limits of the APU. You can't process what can't be delivered.


"Um, first this isn't the article that draws those conclusions. Second, Ryzen APUs are a big leap over previous APU generations and rival Intel chips."

True. Ryzen APU's aren't slouches. They're not yet designed to drive gaming and if AMD doesn't find a serious solution for memory size and bandwidth, then APU'S will never be a current generation gaming chip product.

 

[Cooe14]

"You would then force people to spend a lot of extra money on a motherboard for a feature that may not be used later. This defeats the purpose of an APU - an inexpensive package with descent graphics This works for laptops, but not desktops."

Just looking forward to a stronger integrated solution. In this case, memory is the bottleneck. No complaints really. Just observing the limits of the APU. You can't process what can't be delivered.


"Um, first this isn't the article that draws those conclusions. Second, Ryzen APUs are a big leap over previous APU generations and rival Intel chips."
True. Ryzen APU's aren't slouches. They're not yet designed to drive gaming and if AMD doesn't find a serious solution for memory size and bandwidth, then APU'S will never be a current generation gaming chip product.

Lol if they aren't "meant to drive gaming" neither are the 30 odd million OG Xbox One's & S's which performance near identically in AAA titles (720p Ultra, 900p medium/high, 1080p low/medium), and not far off the PS4, so you've pretty much invalidated nearly all the gaming market with that statement. Not everyone needs, or is willing to pay for maxed visual goodies @ 1080p (and thus why the PS4 Pro, Xbox One X, and dGPU's over $200 aren't the biggest sellers by a massive gulf. Baseline consoles & entry level GPU's DOMINATE sales, and that's ALWAYS been the case).

I totally, 100% understand where you & your thinking is coming from though, and I most often immediately jump to similar thoughts myself. This is because after using a high end gaming rig as an enthusiast gamer for a long period, one inevitably becomes desensitized to anything notably beneath that hardware quality level, along with the perception of those who's gaming needs are far more casual/mainstream than your average high end rig owner (the average joe's who play Overwatch, Fifa, CS, maybe some RL & COD, etc... occasionally in their spare time and that's all the gaming they need, and who'd normally be totally statisfied with a console to fill it). This makes what would have been satisfactory to you at one time long before, now absolutely not, no matter how much you try and convince yourself it still is. Your baseline of expectations has shifted through exposure, and your associated very high level of gaming interest colors those even further. Same principle applies to things like moving to a higher screen resolution, at first it feels extraordinary & beyond what's absolutely needed for a good experience (in a GOOD way) but after long enough, that BECOMES part of your mental image of a "good gaming experience" and you can't ever be satisfied going back to less the way you originally were as it doesn't meet that new mental standard. (For ex. I simply couldn't go back to gaming w/o FreeSync at this point if someone literally paid me to haha. For me now, a regular monitor = not enough for gaming, case closed. But if I said that's also true for everyone else, that'd be kinda ridiculous).

I'd wager the vast majority of the people saying these APU's aren't enough for gaming are those that have been gaming at a far higher level than that on the base consoles for a significant period (and would likely feel similarly dis-satisfied by them), and thus have innately different expectations & minimum criteria for satisfaction. You gotta go back to before all you knew was 60fps+ w/ everything cranked and put yourself in average casual gamer Joe's shoes for who that unsatisfactory experience for you, is more than plenty for him. These new chips, meet all of the "satisfactory gaming" needs for a huge market of people (a large chunk of the 80 odd million happy with a base model console), and as such, opens up the door to the freer, open, more customizabile, cheaper day to day, and ever more popular world of PC gaming at a similar performance level & as similar a cost as possible, all of which makes it FAR more likely to be one of people in this groups default choices for a gaming platform in addition to the venerable XBO & PS4. And then if they end up getting sucked in like those far past the point of fathoming an APU/console level return then they'll get to not only experience the vastly superior technical experience in addition to all the other benefits they got from going PC via APU vs XBO/PS4, but those again will be the minority of that group. Just as a base console is plenty o' gaming umph for most, so too will these APU's.

TL;DR - It can be really hard to think from outside the perception of long time enthusiast PC gamer's & builders; long since used to gaming on vastly better hardware than that in the baseconsoles, so it's for one to try and understand the appeal of entry level products like these no matter how good the value, you really have to shift one's thinking back to the needs & wants of a more causal non-PC (yet) gamer instead; like those that actually make up the vast majority of the gaming market (including PC, like I said entry level cards utterly dominate total sales numbers, even if that can easily sounds nuts/not possibly true to people who hang around places like this for entertainment). We are not the norm, not even remotely freaking close to it lol. Though the niche environments like these we frequent can convince us that we are (and I screw that one up just as much as any PC building & enthusiast would).

 

[Danny101]

"Lol if they aren't "meant to drive gaming" neither are the 30 odd million OG Xbox One's & S's which performance near identically in AAA titles (720p Ultra, 900p medium/high, 1080p low/medium), and not far off the PS4, so you've pretty much invalidated nearly all the gaming market with that statement."

Yeah, I was well aware of the dichotomy that existed about APU'S being used in console boxes. Also, keep in mind that these APU's are also customized and are geared toward gaming performance. I imagine that the Ryzen line will also have it's version of a customized chip for a new generation of consoles. I usually stay a generation behind anyway. When I maximize, it's most often not due to gaming as much as it is to extend the internet browsing capability of the machine. They pack more data that has to be processed over the internet along with forced upgrades. I had an Athlon XP machine that could process everything the internet could throw at it up to 2010 and thereafter couldn't process squat. A dual processor was required at that point. I generally turn over former gaming machines to friends and relatives for internet work after games have surpassed it's gaming capabilities.

It may be that I'm overestimating the need of video memory for experience beyond visual clarity. If so, I was always happy with 720. Nothing I can't stand more than stuttering.

 

[Cooe14]

"Lol if they aren't "meant to drive gaming" neither are the 30 odd million OG Xbox One's & S's which performance near identically in AAA titles (720p Ultra, 900p medium/high, 1080p low/medium), and not far off the PS4, so you've pretty much invalidated nearly all the gaming market with that statement."

Yeah, I was well aware of the dichotomy that existed about APU'S being used in console boxes. Also, keep in mind that these APU's are also customized and are geared toward gaming performance. I imagine that the Ryzen line will also have it's version of a customized chip for a new generation of consoles. I usually stay a generation behind anyway. When I maximize, it's most often not due to gaming as much as it is to extend the internet browsing capability of the machine. They pack more data that has to be processed over the internet along with forced upgrades. I had an Athlon XP machine that could process everything the internet could throw at it up to 2010 and thereafter couldn't process squat. A dual processor was required at that point. I generally turn over former gaming machines to friends and relatives for internet work after games have surpassed it's gaming capabilities.

It may be that I'm overestimating the need of video memory for experience beyond visual clarity. If so, I was always happy with 720. Nothing I can't stand more that stuttering.

These don't have any more stutter issues than any other GPU of it's performance level until you tap it out or put it under crazy VRAM loads that only really come from using much higher settings that it never could run well regardless; dedicated memory or not. You wouldn't be able to tell the difference between it and using a discrete GT 1030 w/ 2GB of GDDR5 played at optimal settings. Having discrete memory in and of itself doesn't make a GPU superior to one that doesn't. (The GT 1030's GDDR5 only gives it about 100GB/s max bandwidth, 2x channel DD4 3200 gives 50GB/s, but Vega's much more powerful in other places to make up that gap).

Also lol think you missed my point about the console part. The don't just also use APU's, they perform almost exactly the same as Raven Ridge in today's AAA titles. Basically, if an Xbox One provides you a satisfactory gaming experience, Ryzen-G will give you largely an equivalent one (better in some titles, worse in others, but with gaming performance in general being very similar), but on a PC instead. That's what makes them such huge freaking game changers! It's like buying a CPU for a general use computer, and getting an Xbox packed in totally for free! (which is PLENTY of graphics power for most non-hardcore PC gamers; it seems to be only those mega gamers long used to high end hardware they somehow think it's a gaming disappointment.)

 

[msroadkill612]

These don't have any more stutter issues than any other GPU of it's performance level ... /QUOTE]

Do you guys think this is something Freesync monitors would help greatly with? I get the impression they are a boon for taxed gpuS.

I vaguely recall a report from a guy using a 75MHZ freesync and getting outstanding results from a cheap mainstream freesync monitor.

If we are talking entry level pricepoint PCs, then we are talking people who want good, affordable monitors also. That means Freesync, and Freesyc means AMD.

 

[Polycount]

What about multi tasking surely there maybe a latency issue

Not that I can see. Most gamers don't multi-task when gaming though.

Psh. I beg to differ. Multiple monitors, my friend, multiple monitors. Netflix + The Witcher 3!

Of course, I'm hardly the majority.

Great write-up, as usual.

 

[Dragonstongue]

A GPU while they may have crazy high memory speed there is no doubt about that, but, they are not feeding that full speed to the pci-e bus either, that speed is used internally by the GPU then when the processing is done they feed that information to the display or whatever, there is always a "loss" and as you stated "theoretical" speed, a cpu may only be able to access the main system memory at maximum of somewhere in the range of 56-72gb/s, so they are not "as fast" but then again the graphics card relies on the cpu to feed it the required instructions or bottlenecks will happen, there is all kinds of clock gating, multipliers, link speed up or down etc so all the various parts "sync" with each other.

If a graphics card were truly able to be fed off of its memory as fast as they numbers suggest, than the ones using HBM or HBM2 on a 4096 or 1024 bit memory bus would absolutely demolish everything else, such is not always the case, things are never "perfect" ^.^

That being said, am sure AMD is working whatever magic they can behind the scenes with their motherboard partners to update BIOS etc to make proper use of the system memory to feed Raven Ridge as best as possible, to reduce overhead or whatever, it just does not make sense at least in my mind that a game such as BF1 would show basically no difference at 64mb or 2gb available graphics memory, to me that says the "software" controlling the "hardware" is at a very early stage, they just wanted to make sure it works first and foremost, not to mention there does not seem to be a very large difference in actual performance between the 2200 and 2400 and yet their "specs" show a distinct advantage of the 2400 over the 2200.....seems to me they need fine tuning, otherwise what is the point of paying an extra $100 or more for something that is only netting you at best 10% more given performance O.o

 

[Dragonstongue]

"The Raven Ridge APUs for example are limited to a memory bandwidth of around 35GB/s for system memory when using DDR4-3200."

Okay, I went back and re-read the relevant part of your post. So, the APU has about 1/3 the memory bandwidth as a discrete RX 550 graphics card. So, I still find myself puzzled. How is the equalization between the two being achieved? The APU uses the Infinity Fabric as a faster alternative to system ram? Maybe the system memory accesses required by the APU are much less often than I imagining. I was thinking it was "all the time" since an iGPU has no memory of it's own. But that fact alone, doesn't require that it be used 24/7 I guess.

Bandwidth is just part of the equation and the RX 550 doesn't necessarily require the bandwidth it has. The read/write performance of the RX 550 in reality is more like 88 GB/s, but that's still about 2.5x greater. It's a bit like low end graphics cards that have big frame buffers, that's nice and all but for the most part they don't really need them as they can't take full advantage.

Keep in mind the GT 1030 often beats the RX 550 and it only has a theoretical peak of 48 GB/s.

depends on the game, the 1030 is NOT faster than the 750Ti in most cases, and the RX 500 is faster than the 750Ti in most cases, so, IMO the RX 550 is BETTER in most cases (more shaders, higher total bandwidth etc etc etc)

 

[Dragonstongue]

Excellent write up Steve. This will be very handy for people with only 8 GB of system Ram.
When using an APU, you have as much vRam as you do system Ram. They are the same speed, plain and simple!

It is important to note that BF1 used only 7 GB combined at 1080p. The only instance where 8 GB may suffer, is with COD: WW2. For eSport games, 8 GB will be plenty.

I concluded this using your other material, but posted on a different site:

https://hardforum.com/threads/reviews-for-amds-apu-ryzen-2400g-are-in.1954364/page-3#post-1043491136

Please let me know if everything was properly sourced / linked!

Thanks mate. Note VRAM usage is higher in the other test because we aren't using low settings. The GTX 1060 is much much more powerful than the Vega 8 GPU in the 2200G. That said with ultra quality textures and everything else set to low we still got the same results with the Ryzen 3 2200G, that is to say allocation size made no difference.

Hey Steve, I was wondering if you could possibly answer a question of mine that's been driving me crazy for FOREVER now. Why do the memory bandwidth benchmarks in SiSoft Sandra & AIDA64 report such drastically different results??? AIDA's "Cache & Memory Benchmark" tends to report values near exactly the same as those reported in the official JEDEC figures (I.e. DDR-3200 = 25GB/s per channel, AIDA's results on my Ryzen 7 PC with dual-channel 3200MHz ≈49.5GB/s reads & writes, and slightly less for copies), whereas Sandra's are DRAMATICALLY less, with my results falling around much the same area as yours, aka ≈35GB/s. Why the massive discrepancy?

* Also, it appears that not all boards handle this setting the same way, as I have seen the same exact setting have some pretty freaking dramatic effects on performance depending on what it's set to in the testing of other highly reputable tech reviewers, whereas it's obviously not for you. Super weird sauce.

take the 25.6 multiply x2 if being used bi-directional vs unidirectional, maybe the one program is reading the "peak" on a single channel whereas the other is reading at maximum sustained throughput, 25.6gb/s is exactly right for DDR4 3200 class, however if you take that number and multiply by 2, it becomes 51.2gb/s, which also is very much a true number, depends on the program, depends on the ram you are using (iif single rank, single channel, or dual rank dual channel unganged etc etc)

IMO is like GDDR5 is, take the number they give (memory clock) multiply this by 4, so for example, my 7870 has a base memory clock of 1200, however the "effective speed" is actually 4800Mhz. system memory we use these days is double data rate and can be dual rank, can be single rank, can of course be read as single channel, or dual channel (even when 2 stick are being used depends on the software using it) etc etc..

either way, those programs seem to be agreeing with each other +/- only about 1.7Gb/s for the mentioned PER CHANNEL, speed..only AMD knows how the internal speed is being calibrated, for many years AMD Hypertransport was "capable of"

HyperTransport Version Year Max. HT frequency Max. link width Max. aggregate bandwidth (GB/s)
bi-directional 16-bit unidirectional 32-bit unidirectional*
1.0 2001 800 MHz 32-bit 12.8 3.2 6.4
1.0 2001 800 MHz 32-bit 12.8 3.2 6.4
1.1 2002 800 MHz 32-bit 12.8 3.2 6.4
2.0 2004 1.4 GHz 32-bit 22.4 5.6 11.2
3.0 2006 2.6 GHz 32-bit 41.6 10.4 20.8
3.1 2008 3.2 GHz 32-bit 51.2 12.8 25.6

infinity fabric is a "new version" of what Hypertransport was, so, IMO, those numbers still very much match what AMD design is capable of, in Intel terms with QPI or PCI-E calling "Gigatransfers" Hypertransport" is up to 5.1 GT/s ^.^

Either way, IMO those numbers are exactly where they should be "within margin of error" as they are completely different programs, maybe one is reading the actual speed per channel per dimm, whereas the other might just be comparing the "spec" and if the number is what it should be (example JEDEC specs 3200 as 25gb/s the module is reporting that is it following the JEDEC number so it just "reporting" the number that is flagged vs actually running at full load to see what the given performance ACTUALLY IS...bring up my graphics card as an example, some programs only report the baseline memory speed of my card at 1200Mhz whereas others are reporting the ACTUAL effective clock speed 4800Mhz ^.^

Also JEDEC "spec" for DDR4 is given as PC4 followed by the number PC4-25600, the 25600 is the "peak bandwidth" so therefore, the "actual" peak speed of 3200 class DDR4 per channel is 25600, or 25.6GB/s, so in one case the number is seeming "low" whereas the other to you seems "low" but in the first case, the number is actually lower than it should be in the first place, whereas the other is lower as well.
25GB instead of 25.6
49.5 instead of 51.2
+/- margin of error either way, if anything, the "internal speed" of Ryzen 2xxx saying 35gb is actually not bad especially if this is only reporting at 1/2 rate, if it is "based on my numbers stated above..if it applies the same thinking" 70GB is actually VERY good O.O

 

[Nightfire]

TMI

 

[msroadkill612]

A hypothetical question driving me nuts?

Yes, questionable economics to many I know (16GB is cheaper and better option), but just say, take one normal minimalist 8GB 2400g, booting on whatever, but add a dedicated 256 GB $120US Evo nvme for swap/scratch/temp files.

I.e, an optimally formatted nvme for large sequential swapping/paging R/W (as opposed to a smaller default block size suited to a boot system disk) - 32MB block size e.g. (any thoughts on optimal block size?)

My my question is basically, to what extent would a dedicated, tweaked block size nvme drive improve windows swapping etc., compared to a swap file on the same model drive optimised as a bootable system drive?

The lag may not be so bad if windows is as smart as I suspect - delaying slower writes, and pre-fetching cache reads into memory.

Occasional swapping may be tolerable to many. With the amazingly faster samsung nvme drives, swapping ain't what it used to be.

My 8GB current non gaming PC never seems to max out 8GB

 

[DragonEgg]

Late to that discussion...

I tought VRAM was to prevent other system tasks to slowdown your game on a low RAM system Ex Antivirus and/or a browser left open or whatever...

 

[neeyik]

AIDA64 reports maximum throughput where as SiSoft Sandra I feel is a bit more realistic as it reports sustained throughput.

I think the difference lies in that AIDA64 will use every possible extension supported by the CPU and 'hand tweaked' assembly code to perform the best possible memory read/write/move performance, whereas SiSoft Sandra is more of a generic multithreaded process that uses a dynamic RAM load to create the sustained throughput. Sandra could be coded in a similar manner to AIDA, but what apps/games are coded for every possible CPU variant in that manner?

I tought VRAM was to prevent other system tasks to slowdown your game on a low RAM system Ex Antivirus and/or a browser left open or whatever...

Essentially, yes - integrated GPUs, with no dedicated memory, are allocated a set range of memory locations, to prevent the OS from using them. However, while the allocated memory is contiguous, is still set within system. Having some memory, separate to the system's, means that it can be fully allocated to the GPU, with no hardware or software, bar the CPU, GPU drivers, and graphics API, having access to it.[/QUOTE]

 

[msroadkill612]

I had a bit of a hard time following this comment but I don’t think there is a fundamental logic fail here at all.

“Shifting data in and out of system memory is significantly slower than VRAM. The Raven Ridge APUs for example are limited to a memory bandwidth of around 35GB/s for system memory when using DDR4-3200. So in the case of the RX 550, it has a bandwidth of 112GB/s when accessing data locally using the VRAM, but when accessing data from system memory it's limited to 16GB/s (PCIe 3.0 x16 limit), which is to say that it takes at least seven times longer to process the same data.”

The Vega 8 and 11 GPUs don't use the PCIe bus but I think you realize that based on your comment, I'm just missing whatever the point is sorry.

Belatedly, but incase it helps future readers...

No steve, thats my error. The loss of 8 lanes by using an apu fooled me into thinking it was using 8 pcie lanes as the gpu<>ram bus.

As I now know, the gpu uses Fabric, & impressively saturates available ram bandwidth - which is getting up to 55GB/s+ these days on am4.

Ta for ur reply.