The One Vision That Intel, AMD And Nvidia Are All Chasing – Why Heterogeneous Computing Is The Futur

Pottsey · 13 Oct 2015 at 18:42

layte said:
I'm talking the technology itself and the functionality it brings. http://developer.amd.com/resources/...hat-is-heterogeneous-system-architecture-hsa/ Not the AMD led group who have decided to hitch themselves together (which is talked about in the latter part of the link /hat).

What do you mean AMD lead group? AMD are members yes but I was not aware they are in charge or that's its a AMD lead.

AMD didn't found the group alone. Imagination have been pushing for this just as long if not longer.

drunkenmaster · 13 Oct 2015 at 18:45

layte said:
NVlink on its own yes, I'll agree with that. But as I've said, it enables other technology that Nvidia have already and are working on to bring similar shared computing to the table. I even alluded to that in my original post where all I said was NV and IBM were cooking something up built upon NVlink, then things went crazy.

What the HSA alliance are setting out to do is certainly different, but that's something else entirely. I'm talking about the technology itself.

Still none of it does what you say it does. Even your first rebuttal where you managed to absolutely pointlessly highlight the "allows gpu to access system memory at the same BANDWITH as the cpu" isn't what HSA does and it's exactly the problem. Bandwidth isn't an issue, latency is. Notice it didn't say the GPU could access with the same latency. The latency is the ball game and none of these technologies address the latency problem.

LIterally none of the white papers (which is a pretty hilarious description of basic blog style posts) you've linked suggest any functionality that HSA provides.

The 'unified' memory access Nvidia is (yet) to provide will not enable any of the functionality HSA is aiming for. IF and only if Nvidia was providing absolutely and completely full unified memory(which would mean no caveats about only matching the gpu memory and the rest of the system memory not being unified for instance), it would NOT provide the same functionality. Nvidia themselves saying the latency is not noticeably reduced, you are still talking across the bus.

You've yet to state which technologies will provide the same or similar functions to all of the listed HSA features.

I wonder which of these technologies will be useful when AMD is producing APU's with 256GB/s bandwidth stacks of HBM2 on die providing real and full unified memory with on die latency and bandwidth beyond anything IBM or NVLink can achieve.

HSA is primarily geared around local acceleration of any code, on a dsp, a gpu, on other cpus, ANY compute IP block that can all talk to each other and use the same code.

NVLink is about providing better multi-gpu scaling ONLY to IBM computers for GPGPU compute.

humbug · 13 Oct 2015 at 18:47

drunkenmaster said:
Still none of it does what you say it does. Even your first rebuttal where you managed to absolutely pointlessly highlight the "allows gpu to access system memory at the same BANDWITH as the cpu" isn't what HSA does and it's exactly the problem. Bandwidth isn't an issue, latency is. Notice it didn't say the GPU could access with the same latency. The latency is the ball game and none of these technologies address the latency problem.

LIterally none of the white papers (which is a pretty hilarious description of basic blog style posts) you've linked suggest any functionality that HSA provides.

The 'unified' memory access Nvidia is (yet) to provide will not enable any of the functionality HSA is aiming for. IF and only if Nvidia was providing absolutely and completely full unified memory(which would mean no caveats about only matching the gpu memory and the rest of the system memory not being unified for instance), it would NOT provide the same functionality. Nvidia themselves saying the latency is not noticeably reduced, you are still talking across the bus.

You've yet to state which technologies will provide the same or similar functions to all of the listed HSA features.

I wonder which of these technologies will be useful when AMD is producing APU's with 256GB/s bandwidth stacks of HBM2 on die providing real and full unified memory with on die latency and bandwidth beyond anything IBM or NVLink can achieve.

HSA is primarily geared around local acceleration of any code, on a dsp, a gpu, on other cpus, ANY compute IP block that can all talk to each other and use the same code.

NVLink is about providing better multi-gpu scaling ONLY to IBM computers for GPGPU compute.

Isn't that what i said?

http://forums.overclockers.co.uk/showpost.php?p=28679354&postcount=44

You put it better tho....

drunkenmaster · 13 Oct 2015 at 18:55

Pottsey said:
What do you mean AMD lead group? AMD are members yes but I was not aware they are in charge or that's its a AMD lead.

AMD didn't found the group alone. Imagination have been pushing for this just as long if not longer.

For the record AMD did start HSA, they were working on HSA for well, a year publicly, probably 2-3 privately before the HSA foundation was announced. I mean that is a relatively big deal but this is fairly common. Most industry standards start by being spearheaded by someone before being, not sure of the best term for it, being handed off to a controlling group to push forward. If AMD kept hold of it then other companies don't get involved.

OpenCL was started by Apple then handed off to a group to push forward together as a group. These things kept hold of by one company die a death because the industry doesn't want to support something they have no say at all in.

AMD, Apple, Intel, Samsung, even people like Sony have a history of creating things then 'giving' them to the industry. Some have a long history of many things, some companies give a few things away, hold on to others.

In general no one wants to be the guy to start spending money only for someone else to beat them(hd dvd/blu ray, vhs/betamax, etc) and a group together starting from scratch will get no where fast because no one can agree on anything. A head start from one company with a good idea then being handed off is pretty much the best way new things get started and actually supported.

Pottsey · 13 Oct 2015 at 19:02

Imagination have been publicly talking about and working on HSA for years. I thought years before AMD started publicly talking about it. To be clear I don't mean to say it came from Imagination as I have no idea who thought of it first. Its just I thought AMD was late to the game talking about it, years after everyone else, so I was surprised to hear them getting credit for starting it.

EDIT: Seems like it dates back a lot longer then I thought and frankly I cannot be bothered to go back past 2012 to see who came up with it first. It doesn't really matter who came up with the idea now they are working as a group.

D.P. · 13 Oct 2015 at 19:19

Pottsey said:
Imagination have been publicly talking about and working on HSA for years. I thought years before AMD started publicly talking about it. To be clear I don't mean to say it came from Imagination as I have no idea who thought of it first. Its just I thought AMD was late to the game talking about it, years after everyone else, so I was surprised to hear them getting credit for starting it.

EDIT: Seems like it dates back a lot longer then I thought and frankly I cannot be bothered to go back past 2012 to see who came up with it first. It doesn't really matter who came up with the idea now they are working as a group.

HSA was originally conceived by IBM and Toshiba back in 2001, privately developed by IBM and there before that. The first big commercial product was the CELL process used in PS3 and then servers. Intel have been working hard on HSA since the same time frame, we now have Intel Phi. Nvidia have been doing it since 2005 with Tesla GPUs and now have a solid partnership with IBM.

AMD tried to jumpstart on the bandwagon by paying a ridiculous amount for ATI.

adolf hamster · 13 Oct 2015 at 19:30

humbug said:
Intel have perfectly capable GPU cores of their own, they don't need AMD's GPU cores, all Intel need to do is update their architecture to be fully HSA compliant.

True, i was thinking more in terms of what will be available to us gamers, that said i suppose theres not much stopping intel expanding their igp range to compete with amd/nvidia in the gaming scene when this takes off.

drunkenmaster · 13 Oct 2015 at 19:39

There is HSA the acronym in general and HSA the acroynm for a specific incarnation of it with specific details of implementation. HSA, the general version, has been worked on since computers were made, accelerating certain functions by having different blocks can be thought about at as low levels as separating cpu cores into FPU/Int logic.

HSA as a specific incarnation and now managed by the HSA foundation was started by AMD. As said the who isn't that important, though a necessary step. That the concept was moved to a point others were willing to sign on and manage it as a standard is what's important.

Had it stayed AMD it would have gone nowhere as the software implementation side is and will continue to happen purely because enough players are doing it to make it very worthwhile to support through software.

This is fundamentally where Fusion failed, because while a lot of the hardware was there software support wasn't forthcoming with all the people trying to make gpu acceleration were all doing it differently. Fusion in reality was AMD's first attempt at HSA.

Fusion I think you could call trying to take what were fundamentally established concepts of CPU and GPU and combine them on die. HSA is taking any and all compute units and providing a basis with which to stitch them together and code for them in an easy to implement way. Fusion was fairly specific where HSA is much more general and takes the big step of becoming a standard to not just provide the hardware, but bring the software support with it.

pete910 · 13 Oct 2015 at 20:14

** Post deleted **

May I ask what you are smoking ? :confused:

humbug · 13 Oct 2015 at 20:16

Good Grief... again.

wazza300 · 13 Oct 2015 at 20:19

homosexualgenious is better

humbug · 13 Oct 2015 at 20:51

lol ^^^^

Howling · 13 Oct 2015 at 21:00

You had me at homo.

Raviolay · 14 Oct 2015 at 01:21

** Post deleted **

Mantle has been handed off, and is (apparently) in both DX12 and Vulkan in some form. Two API's that use the same code yet work on any compliant hardware. With the goal of increasing performance through software, by removing/mitigating bottlenecks & latency. Sound familiar to HSA (software) perchance?

AthlonXP1800 · 14 Oct 2015 at 08:20

Wonder is AMD future of homosexualgenious computing doomed to failure?

AMD Corporate Fellow Phil Rogers was with ATI and AMD for 21 years and also was the president of HSA Foundation had now left AMD and HSA Foundation, he now work as Compute Server Architect at NVIDIA!

http://www.theregister.co.uk/2015/10/14/phil_rogers_amd_nvidia/
http://wccftech.com/phil-rogers-leaves-amd-nvidias-chief-software-architect-compute-server/

This is no doubt a massive blow for AMD, was Phil Rogers created HSA and software? He probably not happy with AMD future direction or HSA lacked of progress and he believed AMD has no future in homosexualgenious computing.

AthlonXP1800 · 14 Oct 2015 at 08:20

wazza300 said:
homosexualgenious is better

drunkenmaster · 14 Oct 2015 at 11:24

Not happy with the direction AMD was going, he was president of HSA foundation, he had the largest say in the direction and HSA is gaining significant momentum.

As established by this thread(for those without blinkers) AMD has an incredibly coherent plan going forward for compute architectures, Nvidia made a bigger bus and as usual is behind the industry in architecture because they prefer sticking to close standards and no industry support which incidentally is why they gave up with Tegra in servers... the industry doesn't want to buy servers that aren't compatible with anything else they buy nor from an unreliable supplier(providing chips that don't meet spec for 4-5 years in a row with Tegra).

You poach people like that by throwing money at them, Nvidia required someone with his level of expertise and knowledge of how to push architectures forward. A guy like him doesn't go begging for a job to Nvidia because his current job sucks. Nvidia needed him and likely offered him huge money. If he wasn't doing a good job directing HSA and working with AMD, why would Nvidia poach him?

People move between these companies all the time, 99% of the time it's someone poached for an area they need to improve. AMD took Nvidia's server sales VP or whatever he was, they needed a bigger presence in selling servers and HPC type units.

LoadsaMoney · 14 Oct 2015 at 20:16

AthlonXP1800 said:
Wonder is AMD future of homosexualgenious computing doomed to failure?

AMD Corporate Fellow Phil Rogers was with ATI and AMD for 21 years and also was the president of HSA Foundation had now left AMD and HSA Foundation, he now work as Compute Server Architect at NVIDIA!

http://www.theregister.co.uk/2015/10/14/phil_rogers_amd_nvidia/
http://wccftech.com/phil-rogers-leaves-amd-nvidias-chief-software-architect-compute-server/

This is no doubt a massive blow for AMD, was Phil Rogers created HSA and software? He probably not happy with AMD future direction or HSA lacked of progress and he believed AMD has no future in homosexualgenious computing.

Orangey · 14 Oct 2015 at 20:52

What was left for him to do... it hit 1.0 recently. AMD's chips are already fully compliant. It should be on autopilot by now.

Get dat money. Then come back in a years time as a double agent.