HSA Is About To Go Viral

Toggle sidebar Toggle sidebar
Status
Not open for further replies.
A technology of the future again.

AMD said about the similar thing for GPGPU five years ago. I didn't see big progress yet.
 
A technology of the future again.

AMD said about the similar thing for GPGPU five years ago. I didn't see big progress yet.
Click to expand...

Are you joking? GPGPU is huge and used almost everywhere from media creation/editing like Photoshop even into games with OpenCL lighting features.
Click to expand...

He is right, GPGPU is used significantly more now than it used to be and undergone a considerable amount of development.
 
Does this mean AMD's cpu division might not die off on us? even if it is only in the phone/tablet sphere, I'd be glad to see them survive. Still, I can't help but dream about Intel getting some competition in the desktop world. Might be the smart move in the long run though, as I expect that 10-20 years from now home users will not use desktops for gaming.

 
Does this mean AMD's cpu division might not die off on us? even if it is only in the phone/tablet sphere, I'd be glad to see them survive. Still, I can't help but dream about Intel getting some competition in the desktop world. Might be the smart move in the long run though, as I expect that 10-20 years from now home users will not use desktops for gaming.
Click to expand...

Intel's desktop competition is ARM, and it has been for a long time. You just don't realize it yet.

AMD just uses Intel's patent on x86, because of the popularity of IBM PC's back in the day.

IBM doesn't even use Intel anymore.
http://www-03.ibm.com/systems/power/hardware/s822/specs.html
They sold their consumer server business to Lenova awhile back.
 


Last 10 years? I wouldn't say that. In the last 10 years we had Core 2 and Sandy Bridge, both which were pretty big leaps.

I would say more last 3-5 years as Intel hasn't had consistent or aggressive competition in the desktop or server market from AMD.

The one problem ARM faces though is that it is a simplistic design which makes it very power efficient. The more complex they make it the less efficient it becomes. When they become as complex as x86 they will lose that advantage.

That said, HSA is interesting but beyond hardware is software developers and that is where it will be hard to get them to accept it. GPGPU is huge in the design industry but very rarely used in gaming or much else, DX is still the majority holder for gaming.

the idea is that if you can get mass adoption by the software developers you have pretty much won. If not, then you will lose.
 

ARM announcing that they will be integrating HSA support in their reference designs does not really help AMD since most companies designing and manufacturing ARM-based chips are competing against AMD in the mobile and (potentially) server space and use non-AMD IGPs, so no IGP/GPU IP-core sales there either.

The only way this might help AMD a little is by seeding some extra interest in HSA on the x86 desktop side of things but that is not going to be anywhere near enough to save AMD on its own.
 
HSA only works well if the entire system's memory is unified. This works great for mobile because memory is unified by design. In desktops, it won't get headway beyond AMD's APUs, which can't compete in raw performance over a high end GPU. So if you're doing serious number crunching, stuffing everything onto the GPU is more than likely better than using an HSA based APU system (theoretically an R9 Fury is at least 8 times better than an APU, while consuming only about 2-3 times more power).

AMD had great ideas here, but they sort of forgot that detail.
 

HSA's shared memory address space has implications well beyond sharing memory between the IGP and GPU: enabling any device to access any memory space has the potential to eliminate tons of redundant buffer copies between address spaces as data moves from user-land to hardware and back. Tighter communication between the IGP and CPU is only the beginning, sound, network, storage, etc. can all benefit from the reduced overhead opportunities too.

In 2017, I expect to see AMD putting HBM on APUs. That should make APUs very interesting for mid-range gamers.
 
Does this mean AMD's cpu division might not die off on us? even if it is only in the phone/tablet sphere, I'd be glad to see them survive. Still, I can't help but dream about Intel getting some competition in the desktop world. Might be the smart move in the long run though, as I expect that 10-20 years from now home users will not use desktops for gaming.
Click to expand...

Intel's desktop competition is ARM, and it has been for a long time. You just don't realize it yet.

AMD just uses Intel's patent on x86, because of the popularity of IBM PC's back in the day.

IBM doesn't even use Intel anymore.
http://www-03.ibm.com/systems/power/hardware/s822/specs.html
They sold their consumer server business to Lenova awhile back.
Click to expand...


Not quite that simple. intel uses a bunch of amd patents in their x86 chips as well.

Neither company could make modern x86 chips without a whole pile of patents from the other. Of course you are right on the origins, originally intel licensed x86 to amd, and they just made a cheaper x86 chip. But that's not where we are today.

Its like any other big company these days. They each have their equivalent of a nuclear bomb of patents pointed at each other.
 

HSA's shared memory address space has implications well beyond sharing memory between the IGP and GPU: enabling any device to access any memory space has the potential to eliminate tons of redundant buffer copies between address spaces as data moves from user-land to hardware and back. Tighter communication between the IGP and CPU is only the beginning, sound, network, storage, etc. can all benefit from the reduced overhead opportunities too.

In 2017, I expect to see AMD putting HBM on APUs. That should make APUs very interesting for mid-range gamers.
Click to expand...
But the thing is if I have a separate CPU and GPU setup, then there's going to be a penalty for accessing system memory over GPU memory every time I want that data. It's better to just copy the data into GPU memory where it's much faster and requires less hops to access it, negating the benefit of HSA.

It's like if Apple's Fusion Drive filled up the SSD first then the HDD and then never moved data around as appropriate. And developers hate segregated memory pools. They pegged the PS3 for it and they're pegging the XB1 for it. Even if it appears unified, it's not actually unified because you can't guarantee your latency is the same. Consistent latency is a very important aspect to have in programming.
 

You missed the point.

HSA also eliminates some data copies and memory address space conversions between user-space memory, kernel-space memory, driver-space memory, IGP/GPU-space memory and possibly others.

While you still need stuff to be in memory before you can DMA it from/to the GPU, you do not want that data to get copied or translated 10 times along the way, that just wastes time and power. All those intermediate steps between user-space and hardware account for a large part of the reason why GPGPU code is often slower than straight CPU-only code.
 

You missed the point.

HSA also eliminates some data copies and memory address space conversions between user-space memory, kernel-space memory, driver-space memory, IGP/GPU-space memory and possibly others.

While you still need stuff to be in memory before you can DMA it from/to the GPU, you do not want that data to get copied or translated 10 times along the way, that just wastes time and power. All those intermediate steps between user-space and hardware account for a large part of the reason why GPGPU code is often slower than straight CPU-only code.
Click to expand...
I'm going to pull out a [citation needed] flag for this.
 
Status
Not open for further replies.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom