Zen and HBM potential?

psyclist

Gawd
Joined
Jan 25, 2005
Messages
844
Hey guys, so seeing the benches surrounding the Fiji GPU about to drop and with HBM on board with 512GB/s links the upcoming ZEN architecture and DX12 could really open AMD up a lead in certain operations im thinking. Anyone else think so?
 
By the time Zen comes out we might already see HBM2. Zen won't have HBM and DirectX12 will run "better" with AMD piledriver cpu anyway. The difference is that IPC is no longer benefiting from the crippled way things were working in DX11 (1 core could only talk to the gpu).
For gaming on Mantle/DX12/Vulkan Zen will have little impact.
 
I don't understand why everyone sees a slower and wider memory like HBM and suddenly thinks it should be included in everything.

It's not magical. It's not lower latency than existing DDR3/DDR4, so it won't benefit your average user.

Every time someone beings this up, I just point to Intel's Crystalwell: it's an improvement for some workloads, but not all. Furthermore, Intel charges at least a $40 premium for that single 128MB memory chip, and most people (those not using the integrated graphics) would find it difficult to see $40 worth of value in that eDRAM chip.
 
I think it is more based on the APUs being a touch bandwidth starved and later APUs will have stronger iGPUs therefore increasing the issue. HBM would alleviate that issue in gaming mostly and prob hinder other operations, but doesn't stop the hope and dreams.
 
I think it is more based on the APUs being a touch bandwidth starved and later APUs will have stronger iGPUs therefore increasing the issue. HBM would alleviate that issue in gaming mostly and prob hinder other operations, but doesn't stop the hope and dreams.

Well, that's the thing - How much more do you want to pay for your APU?

AMD already charges a $50 premium for getting the same processor spec with a GPU included. You can be sure they'll charge you even more for additional HBM memory on-interposer.

Would you pay $175 for a powerful APU, when you could get a more capable discrete card for $100, and slap it on that GPU-less model?
 
Well, that's the thing - How much more do you want to pay for your APU?

AMD already charges a $50 premium for getting the same processor spec with a GPU included. You can be sure they'll charge you even more for additional HBM memory on-interposer.

Would you pay $175 for a powerful APU, when you could get a more capable discrete card for $100, and slap it on that GPU-less model?

I was speaking more for Small form factor like laptops and such. but still you can even consider consoles as great need of such.
 
I was speaking more for Small form factor like laptops and such. but still you can even consider consoles as great need of such.

I guarantee you at 19w AMD is hitting the power wall before they come anywhere near the bandwidth wall. And AMD is not switching-out DDR main memory for HBM, so it's still going to be there. Might as well use it.

Hell, we both know that Intel is hitting the power wall at 15w with just 4 threads and HD 6000 graphics (orange). That's why they made a NUC using the 28w Iris. And that performance improvement all happens with just normal dual-channel DDR3.

http://techreport.com/review/28332/intel-nuc5i7ryh-mini-pc-with-iris-graphics-reviewed/4

Don't think for a moment that AMD has it any better.

Game consoles are a viable option, however :D

My point is, there's limited use case for HBM outside of tasks that utilize vast quantities of dedicated GPU power. Discrete GPUS and game consoles are easy sells for this new tech, but APUs (which are price sensitive)? Not so much.
 
I was speaking more for Small form factor like laptops and such. but still you can even consider consoles as great need of such.

Also lower power consumption which helps laptops quite a bit.
 
Game consoles are a viable option, however :D

I guarantee Sony and Microsoft are already thinking about HBM 2.0 in their next consoles.
Especially the XboX camp, they don't consistently make it to the 1080P resolution in games as it is now with DDR3 in the console, they have been sucking hind tit in graphics.
They are going to have to address 4K and 8K UHDTVs the next time around.

With an next generation nVidia or AMD core and HBM, they could make the mark.
 
I don't understand why everyone sees a slower and wider memory like HBM and suddenly thinks it should be included in everything.

slower? 512gb/s compared to 20gb/s.

It's not magical. It's not lower latency than existing DDR3/DDR4, so it won't benefit your average user.

should we go back to ddr1? it has lower latency than ddr2/3/4.

Every time someone beings this up, I just point to Intel's Crystalwell: it's an improvement for some workloads, but not all. Furthermore, Intel charges at least a $40 premium for that single 128MB memory chip, and most people (those not using the integrated graphics) would find it difficult to see $40 worth of value in that eDRAM chip.

but it wouldnt be eDRAM. it was be 2-4gb of ram for the gpu. just think an 8 core cpu with 1000+ shaders and 2-4gb of gpu ram with 256/512gb/s of bandwidth. that would be a bad ass apu.
 
I don't understand why everyone sees a slower and wider memory like HBM and suddenly thinks it should be included in everything.

A single HBM1 stack at 500Mhz/1Ghz has a bandwidth of 128GB/sec, and most likely uses a lot less power than a DDR3/DDR4 interface that provides the same bandwidth.

For DDR4 systems to generate that much bandwidth, you need a quad-channel (256 bit) interface running at a 4GHz data rate. That's not even taking trace lengths and routing into consideration.
 
By the time Zen comes out we might already see HBM2. Zen won't have HBM and DirectX12 will run "better" with AMD piledriver cpu anyway. The difference is that IPC is no longer benefiting from the crippled way things were working in DX11 (1 core could only talk to the gpu).
For gaming on Mantle/DX12/Vulkan Zen will have little impact.

The misconception that DX12 will magically make Piledriver competitive is silly. The most powerful FX series chips still lose out in multhreaded CPU benchmarks to the i7 2600K, a 4-year old processor. Direct X 12 will make games less reliant on processor performance, yes, but it's not a magic bullet that will make up for Piledriver's poor integer performance and utterly atrocious floating-point performance.
 
A single HBM1 stack at 500Mhz/1Ghz has a bandwidth of 128GB/sec, and most likely uses a lot less power than a DDR3/DDR4 interface that provides the same bandwidth.

For DDR4 systems to generate that much bandwidth, you need a quad-channel (256 bit) interface running at a 4GHz data rate. That's not even taking trace lengths and routing into consideration.

But we don't need the bandwidth we already have on the average gaming desktop. That's why I made the distinction of "slow and wide" when describing HBM - it's the exact opposite of DDR3/DDR4 memory, and thus not an improvement for GENERAL DESKTOP SYSTEM MEMORY in any way, shape or form.

See here where dual-channel only adds "up to" 5% performance on a Core i5 for non-workstation tasks. For the majority it's under 3%:

http://www.gamersnexus.net/guides/1349-ram-how-dual-channel-works-vs-single-channel/Page-3

When you're doing things with more random load patterns (games, web browsing, file compression), memory SPEED matters a whole lot more than the WIDTH.

And while HBM may benefit certain workloads, you're not going to see it replacing standard DDR3/4 in any mainstream systems anytime soon. It's more expensive than DDR4, and also a whole lot less flexible (one capacity per system design, and lower memory chip density). IT would be a major undertaking to integrate it into a memory space without slowing down operations that depend on fast random access.
 
The misconception that DX12 will magically make Piledriver competitive is silly. The most powerful FX series chips still lose out in multhreaded CPU benchmarks to the i7 2600K, a 4-year old processor. Direct X 12 will make games less reliant on processor performance, yes, but it's not a magic bullet that will make up for Piledriver's poor integer performance and utterly atrocious floating-point performance.

Yeah that keeps me up at night integer performance. Why do you ignore the star swarm demo? It shows that AMD Piledriver can keep up with games.

I have a simple answer if it requires computational power then dicking around on a cpu while something as OpenCL is around is a big waste of time....

Games for me are the reason why I should buy a new cpu and Integer performance not so much...
 
But we don't need the bandwidth we already have on the average gaming desktop. That's why I made the distinction of "slow and wide" when describing HBM - it's the exact opposite of DDR3/DDR4 memory, and thus not an improvement for GENERAL DESKTOP SYSTEM MEMORY in any way, shape or form.

See here where dual-channel only adds "up to" 5% performance on a Core i5 for non-workstation tasks. For the majority it's under 3%:

http://www.gamersnexus.net/guides/1349-ram-how-dual-channel-works-vs-single-channel/Page-3

When you're doing things with more random load patterns (games, web browsing, file compression), memory SPEED matters a whole lot more than the WIDTH.

And while HBM may benefit certain workloads, you're not going to see it replacing standard DDR3/4 in any mainstream systems anytime soon. It's more expensive than DDR4, and also a whole lot less flexible (one capacity per system design, and lower memory chip density). IT would be a major undertaking to integrate it into a memory space without slowing down operations that depend on fast random access.

i dont see hbm being used to replace system ram. it would be used as gpu ram. in that usage it would really make an apu shine. who is saying it would be used to replace system ram?
 
Well, if it's dedicated ram it won't be a very cheap apu, now will it?

The whole intent of "cost savings" with the APU is to make use of the bandwidth you already have. Anything more puts extra cost on a price sensitive device

You dreamers don't seem to understand: there's no such thing as a free lunch. Adding 8 to 16 memory chips still costs money even on interposer. Adding more GPU units to an APU cost money. Who is going to buy this $300 APU that can't be moved to another system, and uses way more power t than any OEM would dream of fitting into a SFF pc?
 
Last edited:
i dont see hbm being used to replace system ram. it would be used as gpu ram. in that usage it would really make an apu shine. who is saying it would be used to replace system ram?

Actually I could see it replacing shared ram the system ram being super fast might not show that much benefit but dropping ddr3/4 from the system and one of the taller pieces on the motherboard goes away if hbm can be good enough for the full system with hsa blending gpu and system tasks in memory a budget system board could be reduced in complexity to just the various io controllers and power regulators. I am thinking for super small system like all in one or smart tv or nuc or those wall wart sized pc.

Imagine a pc the size of a deck of cards able to play crisis 3 on high detail with 60+FPS that is where I think hbm can take us.



Well, if it's dedicated ram it won't be a very cheap apu, now will it?

The whole intent of "cost savings" with the APU is to make use of the bandwidth you already have. Anything more puts extra cost on a price sensitive device

You dreamers don't seem to understand: there's no such thing as a free lunch. Adding 8 to 16 memory chips still costs money even on interposer. Adding more GPU units to an APU cost money. Who is going to buy this $300 APU that can't be moved to another system, and uses way more power t than any OEM would dream of fitting into a SFF pc?

I was thinking of embedded type systems like the intel nuc or the ps4/x bone or low-mid end laptops no ram slots at all just the soldered in apu with hbm the only things left configurable would be hdd or odd and wireless adaptors but even those could be soldered to the board directly.

And what do you mean uses so much power zen is going to be on 14nm and the power is just going to be sipped...

also hypothetically the apu is 300 retail but only 150 bulk oem. The thing I envision with this is a capable disposeable computer when it gets obsolete or breaks nothing there to repair or replace except the unit as a whole.
 
Last edited:
Actually I could see it replacing shared ram the system ram being super fast might not show that much benefit but dropping ddr3/4 from the system and one of the taller pieces on the motherboard goes away if hbm can be good enough for the full system with hsa blending gpu and system tasks in memory a budget system board could be reduced in complexity to just the various io controllers and power regulators. I am thinking for super small system like all in one or smart tv or nuc or those wall wart sized pc.

Imagine a pc the size of a deck of cards able to play crisis 3 on high detail with 60+FPS that is where I think hbm can take us.





I was thinking of embedded type systems like the intel nuc or the ps4/x bone or low-mid end laptops no ram slots at all just the soldered in apu with hbm the only things left configurable would be hdd or odd and wireless adaptors but even those could be soldered to the board directly.

And what do you mean uses so much power zen is going to be on 14nm and the power is just going to be sipped...

also hypothetically the apu is 300 retail but only 150 bulk oem. The thing I envision with this is a capable disposeable computer when it gets obsolete or breaks nothing there to repair or replace except the unit as a whole.

You have big dreams my friend, but we're nowhere near solving the power issue :D

See the post I shared above comparing the performance of 15w HD 6000 (orange) versus 28w HD 6100 (red orange):

oyCVT4L.gif


The only difference between the two Intel chips is the amount of power allotted, but you can see how power-limited we are right now on 14nm Broadwell. That's failing to run the game smoothly at 1280x720, so you know we're hitting the power wall at 15 watts. All the chips approaching 30fps on that graph use at least 65w, making them unattainable for laptops.

So while HBM will undoubtedly be a part of PCs once we drop below 10w, we're nowhere near that yet. At this point the extra cost and small power consumption improvement can't be justified due to extra cost (along with the small performance hit most applications would take).

AMD is not going to find nirvana or anything. They will be just as power-limited as Intel is at 14nm and 15 watts, even with Zen.

I too used to dream big, but then I learned too much about silicon physics. It sucks :(
 
You have big dreams my friend, but we're nowhere near solving the power issue :D

See the post I shared above comparing the performance of 15w HD 6000 (orange) versus 28w HD 6100 (red orange):

oyCVT4L.gif


The only difference between the two Intel chips is the amount of power allotted, but you can see how power-limited we are right now on 14nm Broadwell. That's failing to run the game smoothly at 1280x720, so you know we're hitting the power wall at 15 watts. All the chips approaching 30fps on that graph use at least 65w, making them unattainable for laptops.
So while HBM will undoubtedly be a part of PCs once we drop below 10w, we're nowhere near that yet. At this point the extra cost and small power consumption improvement can't be justified due to extra cost (along with the small performance hit most applications would take).

AMD is not going to find nirvana or anything. They will be just as power-limited as Intel is at 14nm and 15 watts, even with Zen.

I too used to dream big, but then I learned too much about silicon physics. It sucks :(
That is for laptops but what about all in one and set top boxes those could be run at 35-65w

laptops for a sub 500 gaming laptop could run up to 35w when plugged in unplugged could scale things down below 15w you just won't have the performance needed to do more than surf and extremely light gaming. I know most people get hung up on they should have all the performance on a battery they do while plugged in but there is only so thin the power cost of power can be shaved. Zen will be at least as good as intel hopefully but I don't hold out much hope for amd in the laptop field. Desktop or set top or appliance yes they can excell and in those areas they are not power limited as much as on a laptop.
 
That is for laptops but what about all in one and set top boxes those could be run at 35-65w

That market is already being served by laptops. 15" laptops can have a power budget of around 80w, so there's plenty of room for a Core i7 45w part, and a cherry-picked underclocked GTX 750 or HD 7770 part. There are many set top boxes that simply repackage these laptop parts, so you get the benefits of economies of scale.

The versatility of this configuration (versus just one powerful APU) is many: for one, OEMs can easily replace your current discrete GPU or CPU with another model that is better. Can't do that easily with an APU, which would require new silicon.

And you get just as much synergy using separate components as you do with a single combined piece of silicon. Nvidia and AMD have software that will hand-off the load from the APU to the discrete card.

Finally, if you use discrete cards instead of a particular APU, you can immediately benefit from new technologies you introduced in your desktop graphics cards (like HBM), instead of waiting for a silicon revision to support the new memory.

APUs are a tough market outside of consoles because there's limited cost savings, and you're restricted to X+Y, instead of any combination of CPU and GPU. To make a new piece of silicon profitable, you have to sell lots of them, so one static configuration has to be "good enough" to catch the eyes of OEMS. AMD has had much trouble convincing OEMS that they should pay a premium for more GPU power, so you can imagine how hard-up AMD had to be when they dropped Kaveri prices by $40 last year. This is why I don't have high hopes for AMD building more powerful APUs anytime soon.
 
Last edited:
Back
Top