Why no more mixed memory density video cards

[wolf]

Testing a GPU I picked up cheap in my retro rig and looking up the specs got me wondering, what are the drawbacks and why can't (or won't) any company make mixed density memory configuration video card these days? It is really just greed, or are there technical reasons and limitations. 1024 MB on a 192 bit bus.

 

[AusWolf]

Probably not worth it. Having only one kind of memory chip reduces the number of different components needed to assemble the card by 1.

I assume it also makes configuring timings easier, although I'm not an expert on that.

 

[ty_ger]

It is usually beneficial to reduce the number of different components because it makes the supply chain easier to measure and manage, and usually the supply chain gives you bulk discounts on the components as a reward.

 

[Macro Device]

In most cases, it's CHEAPER to buy 1.5 GB worth of RAM for your 192-bit bus than it is to only get 1 GB. Not only that, also easier to manufacture and configure. And don't forget the K.I.S.S. principle.

All in all, GTX 550 Ti is so horrible it wouldn't even max 768 MB out in literally 99.9% games from then. 1 GB is unnecessary on GPUs this slow.

 

[Dr. Dro]

I recall reading a very long time ago that doing this caused a negative performance impact as it reduced the effective bandwidth available, much like mismatched size RAM sticks on the motherboard where you'd have 16+32 GB for a total of 48, so 32 GB of this could be accessed at full speed and the remaining 16, since it's physically located in only one of the sticks, can only be mapped and accessed through the channel that it is plugged into, at effectively half speed.

It was on the GTX 660 Ti's review, here:

The GeForce GTX 660 Ti Review, Feat. EVGA, Zotac, and Gigabyte

www.anandtech.com

Of course at a low-level it’s more complex than that. In a symmetrical design with an equal amount of RAM on each controller it’s rather easy to interleave memory operations across all of the controllers, which maximizes performance of the memory subsystem as a whole. However complete interleaving requires that kind of a symmetrical design, which means it’s not quite suitable for use on NVIDIA’s asymmetrical memory designs. Instead NVIDIA must start playing tricks. And when tricks are involved, there’s always a downside.

The best case scenario is always going to be that the entire 192bit bus is in use by interleaving a memory operation across all 3 controllers, giving the card 144GB/sec of memory bandwidth (192bit * 6GHz / 8). But that can only be done at up to 1.5GB of memory; the final 512MB of memory is attached to a single memory controller. This invokes the worst case scenario, where only 1 64-bit memory controller is in use and thereby reducing memory bandwidth to a much more modest 48GB/sec.

 

[freeagent]

In most cases, it's CHEAPER to buy 1.5 GB worth of RAM

*Looks at GTX 580 on the shelf*

 

[Macro Device]

*Looks at GTX 580 on the shelf*

I hammered one with an axe a decade ago because I was stupid mad for I don't remember what reason. No regrets, it wasn't worth repairing anyway.

 

[DaemonForce]

People are wondering about matching VRAM?
I'm still wondering about a non-binary count of total physical memory chips.

 

[Ruru]

All in all, GTX 550 Ti is so horrible it wouldn't even max 768 MB out in literally 99.9% games from then. 1 GB is unnecessary on GPUs this slow.

Most -50 tier cards have always been crap by their price/performance ratio, but 550 Ti was horrible even back in 2011. Hell, even my HD 4850 I had back then was faster.

I recall reading a very long time ago that doing this caused a negative performance impact as it reduced the effective bandwidth available, much like mismatched size RAM sticks on the motherboard where you'd have 16+32 GB for a total of 48, so 32 GB of this could be accessed at full speed and the remaining 16, since it's physically located in only one of the sticks, can only be mapped and accessed through the channel that it is plugged into, at effectively half speed.

Interesting, I didn't know that it affects RAM as well. How about 32GB + 16GB kits where the both kits are in dual channel?

 

[Dr. Dro]

Most -50 tier cards have always been crap by their price/performance ratio, but 550 Ti was horrible even back in 2011. Hell, even my HD 4850 I had back then was faster.


Interesting, I didn't know that it affects RAM as well. How about 32GB + 16GB kits where the both kits are in dual channel?

Similar situation occurs, the imbalance generally rests on the fact that one device has more memory than the other, and this is what causes a problem with memory interleaving. The excess memory cannot be interleaved, and thus is accessed directly at single channel bandwidth.

People are wondering about matching VRAM?
I'm still wondering about a non-binary count of total physical memory chips.

AFAIK, "non-binary" memory sticks still have the same bit capacity across all ranks, banks and chips, meaning that it's possible to interleave them no problem. This is a visual representation of how the technology works in general, the concept is to spread memory addresses evenly across all physical memory devices, which naturally would require them to have the same size. Otherwie, only direct access can be achieved.

 

[Ruru]

Similar situation occurs, the imbalance generally rests on the fact that one device has more memory than the other, and this is what causes a problem with memory interleaving. The excess memory cannot be interleaved, and thus is accessed directly at single channel bandwidth.

So same sized modules it is when upgrading. Nice that you can always learn something new.

 

[AusWolf]

So same sized modules it is when upgrading. Nice that you can always learn something new.

I'd say that depends. A friend of mine uses 2x8 and 2x16 GB modules for a 48 GB setup without a problem. But then, not everybody needs top speeds. I, for example, run my RAM at JEDEC standard 4800 MHz instead of the advertised 6000 because it saves me a good chunk of power and heat for an extremely tiny (if any) real-world performance sacrifice. I'm not saying that everybody should do the same, but it suits me.

 

[Dr. Dro]

So same sized modules it is when upgrading. Nice that you can always learn something new.

Ideally, yes. Mismatched size configurations may work but most of the memory read/write/copy performance goes down the drain, since any excess capacity will be addressed directly in single-channel mode. In some cases this configuration is expected and explicitly supported (I believe Intel calls it Flex Memory Access, don't remember AMD's nomenclature if it has one), but the concept and drawbacks remain. This is the reason why you get iffy performance if the RAM sticks aren't of the same size, it's not even (just) the potential timing conflicts since the CPU's memory controller will even it out on its own.

I'd say that depends. A friend of mine uses 2x8 and 2x16 GB modules for a 48 GB setup without a problem. But then, not everybody needs top speeds. I, for example, run my RAM at JEDEC standard 4800 MHz instead of the advertised 6000 because it saves me a good chunk of power and heat for an extremely tiny (if any) real-world performance sacrifice. I'm not saying that everybody should do the same, but it suits me.

You shouldn't have a problem, it's just the performance that suffers. Even then, it's best to install the identically sized modules in each separate channel, so in this case 8+8 and 16+16, that way memory interleaving still occurs even if you eventually exceed one channel's capacity and only a single channel is accessed for the next segment of memory

This explains better than I can

System Memory Controller Organization Mode (DDR4/5 Only) - 001 - ID:655258 | 12th Generation Intel® Core™ Processors Datasheet, Volume 1 of 2

edc.intel.com

 

[AusWolf]

Ideally, yes. Mismatched size configurations may work but most of the memory read/write/copy performance goes down the drain, since any excess capacity will be addressed directly in single-channel mode. In some cases this configuration is expected and explicitly supported (I believe Intel calls it Flex Memory Access, don't remember AMD's nomenclature if it has one), but the concept and drawbacks remain. This is the reason why you get iffy performance if the RAM sticks aren't of the same size, it's not even (just) the potential timing conflicts since the CPU's memory controller will even it out on its own.



You shouldn't have a problem, it's just the performance that suffers. Even then, it's best to install the identically sized modules in each separate channel, so in this case 8+8 and 16+16, that way memory interleaving still occurs even if you eventually exceed one channel's capacity and only a single channel is accessed for the next segment of memory

This explains better than I can

System Memory Controller Organization Mode (DDR4/5 Only) - 001 - ID:655258 | 12th Generation Intel® Core™ Processors Datasheet, Volume 1 of 2

edc.intel.com

Is that what's happening on the 550 Ti in OP's example? It seems to have a perfect 192-bit bus in the database. In case of system RAM, you have 2/3 of it at 128-bit (dual channel) and 1/3 at 64-bit (single).

 

[Dr. Dro]

Is that what's happening on the 550 Ti in OP's example? It seems to have a perfect 192-bit bus in the database. In case of system RAM, you have 2/3 of it at 128-bit (dual channel) and 1/3 at 64-bit (single).

Yeah, it's the same thing that happens with the 660 Ti as well, my post above has the link to the old Anandtech review that goes in depth as to why this causes inconsistent performance

 

[wNotyarD]

I recall reading a very long time ago that doing this caused a negative performance impact as it reduced the effective bandwidth available, much like mismatched size RAM sticks on the motherboard where you'd have 16+32 GB for a total of 48, so 32 GB of this could be accessed at full speed and the remaining 16, since it's physically located in only one of the sticks, can only be mapped and accessed through the channel that it is plugged into, at effectively half speed.

Aka the GTX 970 approach?

 

[iameatingjam]

Ah good memories. A gtx 560 was (kind of) my first GPU. Not really, but it was the first dedicated GPU that I put a significant amount money down on and made sure it had other compatible and optimal components to go with it. I remember it blowing me away. I don't think it had the same mixed memory as the 550 ti as the math works out for 8x 128MB. But I could be wrong, idk.


You can definitely see the trend of shrinking bus sizes when looking at that card. a 256 bit 60 series card. Though I understand why faster memory and more cache make that unnecessary at this point. Still, wish it was more than 128 bit. That seems so.... 50 or 40 series card. Even though those classes don't really exist anymore. And I suppose thats not a bad thing, since they were mostly e-waste. Still, even the best APUs come nowhere even close to bottom of the barrel new GPUs.

 

[Vayra86]

The 660 had an asymmetric bus too and I ran two in SLI.
1.5 GB+0.5GB. 128bit for the first 1.5GB, 64bit for the last half, so kinda like a 970 in reverse.

It was a stutterfest and this would grow more pronounced when running into the 'above 0.5GB' and then again into the >1.5GB range, in SLI that is. The latency between those two chips is different, and its a frametime killer., especially if you add SLI on top. With a single card, it just slows down, or you got a small framedrop (which could be smoothed out). In SLI? Utter madness.

The 970 had the same issue but there the last 0.5GB ran over less bandwidth relatively. Whichever way it goes, it creates two differently performing halves of a GPU; the shaders are still equally fast, the memory knocks it all down the lowest common denominator.

After the 970 I think Nvidia concluded the requirements for GPUs were no longer allowing asymmetric bus/densities, too much risk for too little gain. The gain is product positioning, I don't think its entirely a cost thing, because for the component cost you save, you have a more complicated GPU that probably also requires more aftercare, like the 970 did, I think its more / also related closely to the rest of the GPU stack and how things must be positioned; Nvidia's 'cuts' of the die have been all over the place.. Basically with 970 Nvidia gambled on people not running into the last 0.5GB too often and any game that would need up to 3.5GB would run exclusively over that part of the memory, something Nvidia had to manage in bios or driver. Its also part of the reason it could be sold for the price it had, which was staggeringly low for the first 4GB 'bottom high end' GPU on an extremely efficient architecture for its time.

It might very well be true that the class action that followed the 970 has convinced Nvidia it was enough with that. Because the bottom line really is that they're not selling what's on the box. You don't have 4GB at the advertised speeds/bandwidth. Or 2GB. You get less, and its hard to quantify, but you know you get less.

 

[Dr. Dro]

Aka the GTX 970 approach?

The 970 had a different issue, it was related to how the memory controller works internally and how the core configuration was sliced down from the full GM204 to match its specification. Basically it cannot access both memory segments at once due to that, even though all memory channels are populated

GeForce GTX 970: Correcting The Specs & Exploring Memory Allocation

www.anandtech.com

 

[agent_x007]

Because last time NV tried that they were class action sued, and lost.
Mixed memory density is dead with GDDR7 (or at least should).
I'm happy because I got "fixed" GTX 660

 

[80-watt Hamster]

Ah good memories. A gtx 560 was (kind of) my first GPU. Not really, but it was the first dedicated GPU that I put a significant amount money down on and made sure it had other compatible and optimal components to go with it. I remember it blowing me away. I don't think it had the same mixed memory as the 550 ti as the math works out for 8x 128MB. But I could be wrong, idk.


You can definitely see the trend of shrinking bus sizes when looking at that card. a 256 bit 60 series card. Though I understand why faster memory and more cache make that unnecessary at this point. Still, wish it was more than 128 bit. That seems so.... 50 or 40 series card. Even though those classes don't really exist anymore. And I suppose thats not a bad thing, since they were mostly e-waste. Still, even the best APUs come nowhere even close to bottom of the barrel new GPUs.

x50 cards could be great for users desiring better-than-integrated or those that were simply frugal. The 550 ti getting so much flak here carried me for several years, and while the $125 I paid for it new was probably too much, I can't argue with the sum-total value it provided. I'm not terribly familiar with AMD's historical models, but Nvidia's 6-, 7-, 9- and 1050 families were all dang good for modest rigs, IMO. Even the GTS 250, being a G92 part, was reasonably capable given its segment. Of course, all of those came out before every release was trying to be the next Crysis-level visual spectacle.

 

[iameatingjam]

x50 cards could be great for users desiring better-than-integrated or those that were simply frugal. The 550 ti getting so much flak here carried me for several years, and while the $125 I paid for it new was probably too much, I can't argue with the sum-total value it provided. I'm not terribly familiar with AMD's historical models, but Nvidia's 6-, 7-, 9- and 1050 families were all dang good for modest rigs, IMO. Even the GTS 250, being a G92 part, was reasonably capable given its segment. Of course, all of those came out before every release was trying to be the next Crysis-level visual spectacle.

Yeah I do not disagree, necessarily.... Its just when you get down so low the performance drop off becomes very big as fixed costs stay the same, especially recently. But you're right there is a hole in the market for 150 - 200 cards right now. But I mean, that can be filled with used/old cards can it not? That does even seem to be nvidia's official strategy with the 3050 6gb. Though certainly there are better used choices for the price.

50 series cards these days (like 1050, 3050 6gb) are generally only barely usable at launch, and very quickly become completely obsolete and discarded. That was my concern with keeping a line like that going I feel like it just creates more waste. Where something in say the 70 series could be used for years and years as long as the user is willing to make compromises.

But I don't disagree with you either. I understand people are price constrained, especially in certain parts of the world so sure if we could get a decent performing card in that price range then all is good but I just don't think its profitable anymore, infact who knows if it would even break even. So used/older card seems like a better option than new but so cut down. For example you can buy a 2060 super used for cheaper than the 3050 6GB, and get more way more performance, and give new life to somebody else's junk if you are just a casual 1080p gamer. You could probably do even better if you look on the red side as well.

 

[AusWolf]

But you're right there is a hole in the market for 150 - 200 cards right now. But I mean, that can be filled with used/old cards can it not?

No. An old x90 card still consumes x90 power, not to mention the lack of warranty and the abuse it's probably been through by miners and overclockers.

 

[iameatingjam]

No. An old x90 card still consumes x90 power, not to mention the lack of warranty and the abuse it's probably been through by miners and overclockers.

I did say old/used not just used. If you don't like used, go for old, you can still buy new 6600s. But if say nvidia were to have launched a 4050 I don't think anybody would have liked the result.

I just don't think its profitable to release decent new cards in that price range or even break even anymore, so the cards we get, are a few years old, and heavily discounted to clear them out. Those are the ones we should go for, not the 6500s or 6400s right as they release. Do you not agree?

 

[Chrispy_]

Testing a GPU I picked up cheap in my retro rig and looking up the specs got me wondering, what are the drawbacks and why can't (or won't) any company make mixed density memory configuration video card these days? It is really just greed, or are there technical reasons and limitations. 1024 MB on a 192 bit bus.

View attachment 374638
View attachment 374639

Because it's a lie.

Cards like the 550Ti fell apart when using >768MB of VRAM because the last 256MB were accessed at 64-bit bus width. That last 256MB was at 1/3rd the bandwidth and FPS dropped by 70-80% or more as the bottleneck was so severe it had knock-on effects to cache hit rates that extended far further than the 67% reduction in bandwidth alone.

You'd be bumbling along at 60-70fps and suddenly it'd drop to 12fps which felt like your PC had just completely broken. The drivers did everything in their power to hide the fact it was really only a 768MB card and refused to use the remaining 256MB unless you absolutely brute-forced it.

The GTX970 was so bad (32-bit bus instead of the 256-bit bus) for the last 512MB that a class-action lawsuit was successful

NVIDIA Settles Class-Action Lawsuit Over GTX 970 Memory

NVIDIA settled in a 2015 class-action lawsuit against it, for misrepresenting the amount of memory on GeForce GTX 970 graphics cards. The company has agreed to pay every buyer of the card USD $30 (per card), and also cover the legal fees of the class, amounting to $1.3 million. The company...

www.techpowerup.com

TL;DR, mixed density VRAM configurations are pure lies, Nvidia were caught lying and it cost them real money, so they won't do that particular flavour of dishonesty again.