• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

AMD to launch 300W GPU with High-Bandwidth-Memory

bandwidth on a card has precisely nothing to do with pci-e bandwidth. Current HAwaii bandwidth is 300GB/s +, 1 stack of HBM gives 128GB/s of bandwidth and afaik the first round of HBM will be 1GB per stack, so 4GB HBM = 512GB/s bandwidth.

A pci-e 3 16x slot provides around 16GB/s of bandwidth. The reason to have memory on the gpu is so you can load data over there as fast as pci-e allows at the start of a level and then limit pci-e traffic to mostly just telling the gpu what it needs to do with the data it has rather than providing new data.

On the eco angle of everything, reducing idle power usage is fantastic, because wasting power when you aren't using the device using it is literally wasteful. When you're using it, how much power it uses is pretty much irrelevant. You pick and chose how much power you want to use based off the device you want. But 90-95% of new chips, be it mobile soc's, gpu's, cpu's, by design they are generally more power efficient than the previous chip.
 
Last edited:
flopper said:
Its not all they "bend" reality around like showing cards with woodscrews etc...

I look forward what amd has cooked up.
I am hungry so its going to be good.
Click to expand...

The 290X was good to run @94c so maybe this new card with 300w+ available will be good to hit 100c+ and come with an optional frying pan.:D

In all seriousness the above is one good reason why the new AMD card will probably not use any more power than the 290X.:)
 
drunkenmaster said:
bandwidth on a card has precisely nothing to do with pci-e bandwidth. Current HAwaii bandwidth is 300GB/s +, 1 stack of HBM gives 128GB/s of bandwidth and afaik the first round of HBM will be 1GB per stack, so 4GB HBM = 512GB/s bandwidth.

A pci-e 3 16x slot provides around 16GB/s of bandwidth. The reason to have memory on the gpu is so you can load data over there as fast as pci-e allows at the start of a level and then limit pci-e traffic to mostly just telling the gpu what it needs to do with the data it has rather than providing new data.

On the eco angle of everything, reducing idle power usage is fantastic, because wasting power when you aren't using the device using it is literally wasteful. When you're using it, how much power it uses is pretty much irrelevant. You pick and chose how much power you want to use based off the device you want. But 90-95% of new chips, be it mobile soc's, gpu's, cpu's, by design they are generally more power efficient than the previous chip.
Click to expand...

You don't see a lot of motherboards fully equipped with all PCI-E 3.0 x16 slots.:D
 
drunkenmaster said:
load data over there as fast as pci-e allows at the start of a level and then limit pci-e traffic to mostly just telling the gpu what it needs to do with the data it has rather than providing new data.
Click to expand...

I'm really sorry, but I'm gunuinely interested in this, would you, or somone be kind enough to reword this part in the hope that my feeble mind might comprehend it?

Where is there, CPU? what do you mean by level?
 
Kaapstad said:
The 290X was good to run @94c so maybe this new card with 300w+ available will be good to hit 100c+ and come with an optional frying pan.:D

In all seriousness the above is one good reason why the new AMD card will probably not use any more power than the 290X.:)
Click to expand...

20% shader/rop increase over 290x, decentish stock cooler, HBM and DP1.2a and it should be a decent release I think.
 
Predictably the usual suspects have this thread in the gutter with-in minutes.

This forum makes up 10% trying to talk about Hardware by 90% of it's members, 90% trying set-off arguments instigated by 10% of it's members.
 
Last edited:
humbug said:
Predictably the usual suspects have this thread in the gutter with-in minutes.

This forum makes up 10% trying to talk about Hardware by 90% of its members, 90% trying set-off arguments instigated by 10% of its members.
Click to expand...

Main reason why I only ever now just read the main article and leave my thoughts on my wallet being open or not and leave straight away :D
 
new boy said:
I'm really sorry, but I'm gunuinely interested in this, would you, or somone be kind enough to reword this part in the hope that my feeble mind might comprehend it?

Where is there, CPU? what do you mean by level?
Click to expand...

I'm talking about, lets say you load up a game, it takes X seconds to load, this is largely the data being shoved across a relatively slow pci-e bus to the gpu where it stores massive amounts of texture data in the memory on the card.

When a game actually starts rather than the cpu processing data and sending the data in each frame, it just sends a bunch of instructions, load texture X here, colour texture Y there. Texture Y may be 400MB, but telling the gpu to process texture Y in whatever way may be 0.2MB of data. So while the GPU requires the bandwidth on the gpu to access the 400mb of texture data from it's own memory, the cpu sends this command to the gpu over the pci-e bus and only requires a few bits of data to do so.

All the data for every frame as you play a game isn't sent over the pci-e bus, just a very limited set of commands to the gpu. The gpu then executes these commands on the data it has in it's own pool of memory.

So gpu memory bandwidth isn't reliant on pci-e bandwidth. Though many thousands of commands still starts to take up a lot of data and some texture information is certainly sent over to the card as a game is on going. As a game moves into a new area the gpu may delete a bunch of stuff from memory and load in new data, mostly this is limited by either well designed streaming of data, so knowing when and where to do this from the game developer side, or is done when loading a new area of a game with a loading screen. Loading it upfront makes more sense as it reduces the chance of stuttering, when the gpu runs out of memory or has to find the texture data it needs from system memory you get major slow down because that pci-e bus is slow. If it was as fast as gpu memory you could get rid of gpu memory and just use the pci-e bus but it's inefficient from a power standpoint.

Effectively the further you send a signal and the more connections that signal traverses(like the pci-e slot, pci-e to cpu, cpu to memory, etc) the more power it uses, significantly more power.

You would also need the main system memory in that scenario to match the speed gpu memory does... which would mean more expensive power hungry system memory.

Ultimately because we can front load the data into level sized chunks and store it on the gpu it makes sense to have the required faster gpu memory as locally as possible, cheaper slower system memory for the CPU(as the cpu doesn't require uber fast memory) and just deal with the loading time of games.

This is another reason for HBM, because it's moving it even close to the GPU(from the pcb to on package where you reduce the length of the traces from ~50cm to ~5cm you end up with a drastic reduction in power.
 
Current Graphics memory does not have anything like the texture transfer performance modern engines can output.

I'm learning that in my limited development experience, i'm constantly trying to work my way round bandwidth bottlenecks.
 
You must log in or register to reply here.
Back
Top Bottom