Is there a way to augment the VRAM on a GPU?

[couture]

The memory capacity on the GPU board is fixed... is there a way to somehow spoof the location of memory the GPU references? In doing so.. could you redirect the GPU to use an external SSD as its RAM?


I doubt this is possible... but the idea of a GPU accessing a 2TB SSD is surely highly appealing?

[bathrobehero]

The memory capacity on the GPU board is fixed... is there a way to somehow spoof the location of memory the GPU references? In doing so.. could you redirect the GPU to use an external SSD as its RAM?


I doubt this is possible... but the idea of a GPU accessing a 2TB SSD is surely highly appealing?

Even if it was possible, it would be super slow (compared to RAM speeds).

[YIz]

It's possible (in theory) but you're limited by the speed of the SSD, the SATA, and the PCIe. yeah.. not a great idea.

[drrobert]

it is something like +60.000 mb/s with ram
and ssd 500 mb/s (lucky)

[couture]

hmm shame

it would be nice to be able to plug in auxiliary storage - of suitable speed

[vm1990]

erm just use the RAM.... games do it all the time run out of VRAM and start eating away at system ram. hell youve go M.2 which is faster than SSD, Ram disks, and PCI-e Storage options. SSD looks like a floppy disk compared to the options you could use

[QuintLeo]

It is NOT possible with current cards - however, AMD is talking about the capability of accessing offboard memory being part of their new VEGA cards, though all of the issues about "very slow access" mentioned by others will apply.

 For a mining card, this is NOT an ability you should care about as it would KILL hashrate to try to use external memory while mining - the external memory IN ANY FORM is going to be quite a bit slower than the GDDR on the card itself.


 M.2 is NOT faster than a RAM disk running from main memory.
 M.2 is a form of SSD - it's speed depends on the interface used compared to a non-M.2 SSD.
 M.2 fastest drives use one or more PCI-E lane(s) to talk to the CPU, they are NOT faster than PCI-E based SSD but are commonly the SAME speed - depends on how many lanes each drive in a comparison uses.
 M.2 slower drives use SATA to talk to the CPU, those are SLOWER than PCI-E based SSD drives and the same speed as non-M.2 SATA SSD drives.



 For a directly applicable comparison - AMD A10-7890K (or 70K or 60K, referred to after this as the A10-78x0K since the GPU side on all 3 is the same) has 512 cores that can be clocked at 800 Mhz.
 AMD HD 7750 also has 512 cores that clock at 800 Mhz, but is a discrete card with (usually) 1 GB of GDDR5 on the card (a few early models used GDDR 3 instead).

 The HD 7750 is commonly 1.5-3 times faster on most usage than the GPU side of the A10-78x0K - there are a very few VERY narrow exceptions, like D.net R5-72 work where the memory is very little used and they perform almost identically, but in most cases the APU using system-board memory is quite a bit slower than the dedicated card using much faster GDDR on the card.

 The A10-78x0Ks are newer by a few years and are still in current production (though not for a lot longer), but are built on the same process node and are a hair newer GCN version than the HD 7750.

 My A10-7890K running DDR3 2133 system ram that I tried once on ZEC mining worked - but pulled something like 11 or 12 sol/sec - my HD 7750 on the SAME machine pulled a bit over 50 sol/sec running the SAME version of Claymore at the SAME settings and core clock.

[couture]

Thanks all for replies, really helpful and informative

So as for building a new rig (without some crazy Frankenstein vddr enhancement) should I wait for vega? (I bet the price is gonna be steep....) or just continue with a usual sapphire nitro+ rx470/480 build?

If graphics cards were under the £180/$220 range life would be a bit easier haha