hello
Opened this issue · 14 comments
Hi, oh wow, Tesla. The size is truncated I didn't realize the type I use is limited to 4GB, I'll fix this in couple of minutes.
Ok, man.. we are wait...
Syahmi, what do u from?
Please test if the issue has been fixed here: https://github.com/prsyahmi/GpuRamDrive/releases/tag/v0.2
я из Малазии
GpuRamDrive_x64.exe
GpuRamDrive-cuda_x64.exe - no fix problem
Any error shown or what size does it allocate?
in first screen shot.
GpuRamDrive-cuda_x64.exe
Hello. I have been looking for someone to create this integration. I have been able to get it set up on my 1080 but the benchmark scores are lower than I would have thought. I know PCIE is not going to have the same bandwidth as on board ram, but I would have expected faster than SSD from the Passmark Benchmark. Awesome to see you create this but wanted to see if anyone else has experienced better results on read and write speed. Also, if we were able to remove the bottleneck, I would love to see a RAMcache version. I got my hands on an early developer release of the new 16G Tesla Pascal with HMB2 onboard ram. Looking to find alternative ways to leverage the card and would love to use the high speed ram as my system cache.
Hi @RBUCKDAVIS, there is also some bottleneck between ImDisk and its Proxy feature. When using ImDisk alone the seq read speed is around 4000-5000MB/s while using proxy feature (still using ram) the speed barely reach ~2000MB/s. This probably due to synchronized read/write and slow IPC mechanisms. Even if the PCIe bandwidth somehow increased, there is still bottleneck on ImDisk and little I can do about that.
Hello prsyahmi
From your above comment it seems as though there is a way to use ImDisk without using the Proxy feature?
I did install this quickly on a client machine (Restored pre install state after a quick benchmark or 2) but don't recall a choice to not use the proxy feature..?
I think you may be underestimating the value of this idea:
The upcoming Radeon VII GF card is coming out with 16GB of HBM2 RAM.
I believe this is more about achieving a wide mem bus than an actual use for all that RAM at this point.
An app that dynamically allocated excess GPU RAM to caching software that speeds up HDD/SSD performance in a kind of RAID 0 arrangement would be a great selling point for such a card. Especially as there seems to be a lot of untapped performance to be had!?
As is; the all important random 4K #s are much better than most SATA SSDs get.
An app called EBoostr
http://www.eboostr.com/
May well be able to give you some idea of what I mean. (untested)
PrimoCache has a deferred write option (dangerous without UPS) that dramatically saves on writes to SSDs, saving on lifespan.|
PrimoCache does not 'see' the ImDisk however.
I have alerted them to your software here:
https://www.romexsoftware.com/bbs2/en-us/viewtopic.php?f=35&t=4637
They seem pretty interested in the idea.
Also remember to implement parallel data access. Many ramdisk products only open one channel for data transfer whereas Skylake cores can handle up to 9 concurrent RAM transfers per clock.
When you benchmark ramdisks you will notice that different. The only ramdisk I know of to do this right is Softperfect RamDisk (https://www.softperfect.com/products/ramdisk).