NVIDIA Unveils The Quadro P6000, The "World's Fastest GPU"

[phoronix]

Phoronix: NVIDIA Unveils The Quadro P6000, The "World's Fastest GPU"

Just days after NVIDIA announced the new GTX TITAN X powered by Pascal that clocks in at 11 TFLOPS, NVIDIA unveiled at SIGGRAPH today what they call "the world's fastest GPU" and is capable of 12 TFLOPS...

http://www.phoronix.com/scan.php?page=news_item&px=NVIDIA-Quadro-P6000

[DTM_NV]

"The P6000 has a power draw of 2450 Watts..."

Man that's a lot of power, it would have to come in its own powered and watercooled case! :P

[thelongdivider]

Not to be a nuisance, but

"The P6000 has a power draw of 250 Watts, 12GB of GDDR5X video memory while the P5000 has 16GB of GDDR5X."

Why does the lower class GPU have more memory? Does the P6000 have HBM?

[efikkan]

Not to be a nuisance, but

"The P6000 has a power draw of 250 Watts, 12GB of GDDR5X video memory while the P5000 has 16GB of GDDR5X."

Why does the lower class GPU have more memory? Does the P6000 have HBM?

No, it's another typo. P6000 have 24 GB GDDR5X, and it's based on GP102 so it can't have HBM.

[Delgarde]

Heh... we've been having a bunch of argument in the office lately, about whether the next order of developer workstations should have 16GB or 32GB RAM. The IT guys weren't entirely appreciative when I pointed out that you can get *video cards* with more than 16GB now...

[efikkan]

Heh... we've been having a bunch of argument in the office lately, about whether the next order of developer workstations should have 16GB or 32GB RAM. The IT guys weren't entirely appreciative when I pointed out that you can get *video cards* with more than 16GB now...

It depends on the workload, but I've been using 64 GB for four years now...
The nice thing about workstations running Linux is that all free RAM is used for IO cache, which is way faster than any SSD. So a Linux computer should always have more RAM than it "needs", and how long will it take before the employee earns that RAM in increased productivity?

[Delgarde]

It depends on the workload, but I've been using 64 GB for four years now...
The nice thing about workstations running Linux is that all free RAM is used for IO cache, which is way faster than any SSD. So a Linux computer should always have more RAM than it "needs", and how long will it take before the employee earns that RAM in increased productivity?

There would be little benefit to 64 GB for us. Most of the load is from compiling code and running automated tests, and even allowing for disk caching, that doesn't require too much RAM. But it *does* keep the CPU at the red-line for a large portion of the day...

[microcode]

There would be little benefit to 64 GB for us. Most of the load is from compiling code and running automated tests, and even allowing for disk caching, that doesn't require too much RAM. But it *does* keep the CPU at the red-line for a large portion of the day...

64GB is standard for me as well. No reason to cheap out on RAM if you have a controller which supports it. Keep in mind that buffers are not often shared between the block cache and running processes. If you have a 3GB codebase in a dynamic language, your runtime could have an entire copy of the source code in RAM, as well as several compiled versions of that code, and all of the working dataset.

16 more GB of RAM these days costs a pittance, a small fraction of a day's compensation for professionals; there's scarcely a reason to bother doubting its utility, it could easily save that much time in disk accesses if you're working on a database-attached application as many are.

[bug77]

Heh... we've been having a bunch of argument in the office lately, about whether the next order of developer workstations should have 16GB or 32GB RAM. The IT guys weren't entirely appreciative when I pointed out that you can get *video cards* with more than 16GB now...

Eh, I can work with 8GB RAM if I deploy remotely. Higher capacity RAM tends to have more latency and it can hurt compilation (one of the few instances where CPU cache won't compensate for RAM's latency).