Details about Kaveri and how the Heterogeneous uniform memory access works, or Da new bang from AMD yo

AMD’s “heterogeneous Uniform Memory Access” coming this year in Kaveri | Ars Technica

interesting article with lots of words, they have pictures too

 

At least GTA 4 won't complain again, only 512MB memory available. It will see all system memory .

 

Unleash the power amd !

 

Too early for all those words... but will read. From my brief glimpse, it looks interesting and a good setup for a system primarily concerned with graphics operations.

 

What about Kaveri's single-threaded CPU performance?
AMD spoke about fixing that with Steamroller (on which Kaveri is based on).
While the modular architecture has been retained... each module containing 2 cores for example should behave a lot more like 2 intel i cores instead of a single core.

 

having finally read all those words and looking at those snazzy and not so snazzy (really who makes those ppt presentations should be applauded and shot to death, such a mix of good job and terrible one, the same goes for intel), I can say:

1) This is what they were aiming for

2) I hope that people actually code with hUMA in mind, they might actually benefit from coding done to crystalwell

3) This is similar to what IBM did with power 6 and 7, but with a gpu in there, and this is what crystalwell is as well

My take on this is that this will serve as a L4 cache or even in the end substitute the L3, which is good, L3 is basically more space than less latency like L1 and L2, you need coding done specifically to it so that it would enjoy the benefits of HSA

In the end there is a clear trend that both intel and AMD are pursuing the same thing here, so this might be done sooner rather than latter.

PS: crystalwell is the GT3e gpu, actually the embedded ram on it

I can say that kaveri will be very anticipated on my part, I have some friends in dire need to change their laptops, I hope that AMD get some very good design wins on this and that it appears on quality notebooks. In the end I just hope that they improve A LOT their cpu power, competing with dual cores is not healthy, when you are supposed to be quad, so that I can actually consider then buying again in notebooks

 

I think it's going to improve somewhat, especially with the decoder bottleneck fix.

Although it's not going to compete against Intel's IB, single-threaded performance is starting to lose importance. Tom's Hardware replaced the dual core IB Pentium with an old x4 Athlon in the recommended budget CPU, citing that many of the newer games see better performance boost with even slow quad-cores.

 

Sounds very promising, although it doesn`t say anything why IGPs will be stronger. Just optimized architecture that let the CPU and GPU read and write simultaneously to get better integration = faster. Or am I missing something here?
CPU wise it does seems like a big step up. I`m really interested in seeing what the Kaveri can do

 

there is nothing saying about cpu power, there is so much cache that a cpu can use with consumer load, you dont see 20mb of L3 cache in a consumer cpu. Though the shared cache might give a little boost in performance, but needs to be, probably, coded in to work as intended

the gpu should benefit from the fast and low latency ram

so basically, great news for gpu, and probable good news for the cpu in mid to long term, with no info on how steamroller will work or if its just a rehash of piledriver, which I think it is, with tweaks to the arch

all in all that would mean that this is going to take the igpu performance crown with a good distance from the gt3e core

 

Because of access to faster RAM? How does that suddenly give better GPU performance?

 

Erm, isn't that exactly what Intel is also doing?

 

No, they are adding 2.5x more graphic units while adding RAM. RAM alone doesn`t make a difference if the units isn`t bottlenecked by the RAM. I assume laws of physics goes the same way for both companies?

 

we know the igpus are bottlenecked by the ram.

intel and amd are going to add low latency high speed ram, I dont know about the bandwidth, it should be probably be very wide

AMD is going to use GCN cores on kaveri or so I read

so all in all, this should give a boost to gpu performance alone on the embedded ram part, if they do use GCN on kaveri this should be another boost. If all of this comes true I expect a double in igpu performance from trinity or the speed bump richland so from 1300 to 2600

GCN is much more competent than vliw4

basically Im waiting on what broadwell will bring to the table, will intel actually go and compete with AMD on their turf of igpu power with the new arch?

this should be interesting and no questions asked on who the leader of igpu performance will be when kaveri arrives

 

Oh snap, I totally forgot that adding faster DDR actually makes both Trinity and Ivy Bridge push out more FPS. Sorry about that.
Right so its gonna be based on the newest GPU architecture. That is something thats gonna indeed make a difference too. I`m curious to exactly what changes they make though. Broadwell is doing the same, updating IGP to a new architecture. Not sure if Intel excels AMD with this though.

I know AMD plan to sell the APU thats inside the PS4, a cut down version of it that is. 8-core Jaguar with a 7860 GPU on the same die

 

now that is nasty, I wonder how awesome its going to be able to play mainstream games and fail horribly in my beloved strategy games

now I wish that more tablets would take jaguar, I cant stop imagining a macross (robotech for ye americans) shooter with that cpu, no more thinking and hesitating to buy a psone just for that

 

I'm interested to see if there are any mobos or laptops that come with at least 6 GB of GDDR5. Memory bottleneck excluding absurdly high anisotropic filtering or resolution? Bah humbug.

 

RAM does bottleneck the IGP's, very easy to test... and been proven numerous times before.

 

Could they make the unified memory in kaveri "user upgradeable" like sodimms in an ordinary laptop? GDDR5 dimm modules right?

Sent from my SPH-M580 using Tapatalk 2

 

I think they are just SODIMMs in an ordinary laptop. At a fundamental level, this is still just integrated graphics. All HUMA does is remove the partition between the CPU's portion of your RAM and the IGP's portion.

Rumor has it that AMD is working on a PS4-style GDDR5 main memory controller for their APUs, but whether or not such a project even exists is unconfirmed at this point, much less whether or not it'll make it into Kaveri.

 

that would defeat the purpose a bit, high bandwidth ram, with fast speeds and other niceties

they would need a major revamp on a lot of things

 

High bandwidth RAM is great, but that bandwidth doesn't come for free. Whereas the power draw from DDR3 is almost negligable (the difference in battery life from having 2 vs 4 SODIMMs installed in your laptop is a few minutes at most), GDDR5 sucks a lot of power.

AMD could also go the other way and adopt a tri/quad-channel memory controller, but that has its own set of problems, the biggest of which is the need for every laptop to have enough physical space to accommodate 3 or 4 SODIMMs.

 

They won't use the fastest, most power hungry GDDR5 modules, low level GDDR5 will still offer much more bandwidth than DDR3 modules. These GDDR5 modules (up to 8GB) will be integrated to the motherboard, to save space and make thinner laptops.
As previously were mentioned a Radon 7750M level graphics will be integrated, that is about 2000 3DMark points.

 

Don't forget that gddr5 has hideous latencies to achieve its high bandwidth. A guy actually tried to downclock the gddr5 amd 4870 to gddr3 speeds and found significantly poorer performance. I can only imagine how bad the random performance of gddr5 will be, especially for branchy CPU operations.
The modern igps are definitely memory bandwidth constrained. However, I think the unified address will be most noticeable when ddr4 comes around instead of ham fisting gddr5 to use as system ram.

 

Yeah I still think GDDR5 as system RAM is not the greatest idea, and with GDDR4 it should vastly improve on the iGP memory bottleneck.

 

I think it was mentioned that AMD found a way around GDDR5 latency issues what with being on the chip itself (among other things).
Not sure of the specific details, but I would imagine that it wouldn't be much of an issue.

 

VDC Research Unveils the Ninth Annual Embeddy Award Winners at Design West
Honoring the Most Cutting Edge Products and Services in the Embedded Design Community, LynuxWorks and AMD Take Home the Coveted Award.
Natick, Massachusetts (PRWEB) May 01, 2013

 

k that is good and useless at the same time

 

meh, some recognition for their troubles. Maybe these things will help win product designs with OEMs.

 

I doubt it would be on the chip. However, they could use additional cache to compensate for the increased latency.

 

Actually, I do think it will be on the chip.
AMD used this approach for the modified APU Sony will use in Playstation 4.
That, and one of the things AMD mentioned to be featured on Kaveri (based on Steamroller) is larger cache

 

Likely on the chip PCB not on the die.

 

die is a very very very expensive area, its either on the cpu pcb, stil very expensive or very near it