AMD FirePro the First OpenCL 2.0 dGPU?

[patrickjp93]

http://www.newegg.com/Product/Product.aspx?Item=N82E16814195134&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=

 

There are a couple questions for AMD, namely because unified system memory access was nowhere on their roadmaps for computer or workstation graphics cards. Here's why: you can't do that with a GPU. You can only do that with a CPU. This is why Nvidia has been talking about putting ARM cores on its cards for a while now, to do this unified system access and independent scheduling. OpenCL 2.0 requires this level of access. So, AMD claims it EXPECTS the card to be fully OpenCL 2.0 compliant, meaning it's not sure right now, which raises more questions. 

 

Intel's broadwell chips with Gen 8 Graphics iGPUs are fully OpenCL 2.0 compliant having implemented unified memory, but that's possible because the CPU and GPU are attached to the same memory bus and are literally attached at the hip. AMD's Kaveri Architecture has unified memory, but the chips were released long before the ratification of OpenCL 2.0, so even with unified memory it's not currently clear if AMD can rig Kaveri to handle it all in hardware.

 

So, can anyone dig up some more info? did AMD put ARM cores on their FirePro cards, X86 cores maybe, or is there some other trick they implemented to try to do this?

 

For reference, Nvidia Maxwell is on OpenCL 1.2

http://en.wikipedia.org/wiki/GeForce_900_series

 

So are Intel Gen 7.5 Graphics in Haswell and below and the AMD Rx 200 series.

[flibberdipper]

Kinda neat, I guess.

[patrickjp93]

Kinda neat, I guess.

think about your CPU never having to spend clock cycles telling your RAM bus to send things to the GPU and instead only passing the GPU an address of where to start and then the CPU can just go jump to the next part of the code meant for it while the GPU churns.

[Darkman]

Interesting.

Am I the only one wondering how the hell they're single slot cards?

[flibberdipper]

think about your CPU never having to spend clock cycles telling your RAM bus to send things to the GPU and instead only passing the GPU an address of where to start and then the CPU can just go jump to the next part of the code meant for it while the GPU churns.

Oh. REALLY FUCKING NEAT I WANT.

[patrickjp93]

Interesting.

Am I the only one wondering how the hell they're single slot cards?

Look at the Stream Processor count, bus width, and clock speed. I'm not surprised AMD can pull off a single-slot FirePro at this level. You know what's weird though? It can't do XFire... Nvidia's Quadros can do SLI. WTF?

[Techhog]

Look at the Stream Processor count, bus width, and clock speed. I'm not surprised AMD can pull off a single-slot FirePro at this level.

Hm... Tonga?

[patrickjp93]

Hm... Tonga?

Possibly given it's far too soon to see Fiji/Bermuda FirePro cards.

[zacRupnow]

good looking card

[patrickjp93]

Thats pretty sweet. The fact that its single slot is awesome because then they can fit in 1U rackmount servers that have a riser for them. 

Hardly a point in doing that when you have 10.4 TFlop Xeon Phis at just 350W and the 2nd generation Knight's Landing about to drop, though the OpenCL 2.0 thing is definitely appealing for AMD's side.

[patrickjp93]

True, but I have yet to see a single lot Xeon Phi, also this has 4 displayports on it if you're into that sort of thing. In something like a PowerEdge 2950, this would be awesome. Imagine the posibilities in something like this:

(EDIT: Yes I know its socket 771, but its still a powerhouse)

I'm sure if Intel felt the need/draw of high margin profits. it could and would make a single-slot one.

 

Also, socket 771...HAHAHAHAHAHAHAHAHA! God it's a relic

[RexinOridle]

Wish I could get one of these for my 3DS Max. Life of a poor student.

[patrickjp93]

Wish I could get one of these for my 3DS Max. Life of a poor student.

Research grant proposal/

[Deletive]

Is 285 OpenCL 2.0? because if so I soooo am buying that.

[patrickjp93]

Is 285 OpenCL 2.0? because if so I soooo am buy that.

Currently only rated for 1.2, but that could change if the released FirePro here is based on Tonga and is fully OpenCL 2.0 compliant.

[Deletive]

Currently only rated for 1.2, but that could change if the released FirePro here is based on Tonga and is fully OpenCL 2.0 compliant.

awesome. Because I really want to try and mess around with OpenCL.

[patrickjp93]

awesome. Because I really want to try and mess around with OpenCL.

It's ugly. Don't go in totally blind. Have some learning resources and lots of code examples. Once you master loading kernels in the various ways, most of it is no harder than a parallel algorithms or OS course at a college/university. I had to learn on 1.1 using the C++ 11 standards, and that REALLY sucked.

[Deletive]

It's ugly. Don't go in totally blind. Have some learning resources and lots of code examples. Once you master loading kernels in the various ways, most of it is no harder than a parallel algorithms or OS course at a college/university. I had to learn on 1.1, and that REALLY sucked.

I'm probably going to do that because I start messing around. with the language.