Slashdot Mirror
NVIDIA Announces New Quadro M6000 With 24GB Memory Buffer For Heavy Workloads (hothardware.com)
Reader MojoKid writes: Some might say there's no such thing as a one-size-fits-all solution when it comes to computing, and that's especially true for workstation graphics professionals who need varying levels of performance and memory space. For that reason, NVIDIA is now offering a version of its Quadro M6000 graphics card with 24GB of GDDR5 memory, twice as much memory as much as the original model. According to NVIDIA, customers rendering datasets larger than 12GB can experience up to 5X faster performance compared to the previous Quadro M6000. Like the 12GB version, the new 24GB Quadro M6000 is based on NVIDIA's Maxwell architecture. It has 3,072 CUDA cores, a 384-bit memory bus, four DisplayPort 1.2 connectors, a single DVI-I connectors, and a maximum power consumption rating of 250W. In addition to the doubling the memory buffer, NVIDIA added a few other features, including more GPU clock options, greater software temperature control to keep the GPU temp below the point where throttling occurs, and a new under-power boot message if the card is ever under powered.
I remember being all excited to have a video card with 1MB of RAM which would do 1024x768.
My desktop has 16GB of RAM.
What kind of porn are you guys watching you need a 24GB video card?
Lost at C:>. Found at C.
What surprises me(especially since AMD is clearly capable of it, since the PS4 does it, and it'd be a bit of a shock if Nvidia and Intel were incapable of following suit or already have something ready to go) is that CUDA(or OpenCL)-targeted systems still haven't made any moves(at least not commercially visible ones) toward unified memory architecture; and/or allowing the relatively expansive supply of motherboard memory expansion to use GDDR and connecting the GPU via a faster interface than PCIe if needed.
As much as it would probably involve selling more Nvidia GPUs than they would like, AMD seems like the one that might actually be particularly interested in this, since it would at least get opterons into servers: They've been pushing HyperTransport as a multi-vendor high speed interconnect(first under 'Torrenza', now under the 'Heterogenous System Architecture Foundation', with some fancy FPGAs available that can interact with Opterons in multi-socket boards.
If GPU compute cards are shipping with increasingly monstrous amounts of RAM, that seems like an opportunity for HSA to move the GPU closer to being a first-class citizen along with the CPU; and allowing RAM expansion.
NVIDIA is in fact working on something like that. It's called NVLink and will, according to that page, enable "data sharing at rates 5 to 12 times faster than the traditional PCIe Gen3 interconnect". There's no commercial hardware yet that supports it, but it should be coming soon.
"I'm too busy to research this and form an educated opinion, but I do have time to tell everyone my uninformed opinion."