Why doesn't AMD just solder on 512-1024MB of HBM on APUs? Currently all the 28nm A10 APUs out there are fucking dogshit because of how massively bandwidth starved they are.
Is HBM that fucking expensive?
If AMD made an APU with 640 gnc cores and 1024MB of 128GB/s of HBM, it would be fucking perfect for entry 1080p vydia gaems. They could do this for the desktop excavator APUs they're suppose to release in 2016.
Are they scared such a batshit powerful chip would cannibalize their R7 GPU sales or something?
>>51916700
because Ahmed is not working for them
>>51916700
Yes it is that expensive, you need an extra 90nm interposer die and the memory stacks are not easy to put together. We might still see something like that but I doubt it will happen before it's in all the GPUs.
>>51916700
Honestly I don't think it is that simple. To just solder HBM to the die and call it a day. Would require a complete architecture change.
That said, amd doesnt even need to go that far. Current Godvari chips are actually very good when paired with ddr3 2400mhz. At 1080p they easily keep up with games like Leage, cs:go, etc. But as you said they're bandwidth starved. All amd should do is release godvari with ddr4 capabilities. Ddr4 3200mhz kits arent super expensive anymore and I've personally seen people reach ddr4 4000mhz just by bumping the voltage and loosening up the timings a tiny bit.
Performance would go up a shit ton with HBM of course, but a an APU with ddr4 is much more realistic
>>51916700
It will not happen because people are not willing to spend ~$300 on a shitty APU
Even Broadwell i5s are a hard sell on the DIY market and it's almost $310.
>>51916700
>he literally solders CPU's
fuckin epic!!!!!
>>51916803
>>51916857
Do you think APUs are the future or will we be stuck with cpu + gpu combos for all of eternity?
>>51916700
why don't you do it?
>>51916875
OP's imaginary apu wouldn't be shitty at all. It would be fuckton expensive but not shitty especially since it would use an excavator cpu.
>>51916948
Hopefully the latter
>>51916700
>Is HBM that fucking expensive?
currently, yes.
it's about bandwidth per watt, not bandwidth per dollar, especially in lower segments.
the thing is that interposers and HBM chips are still a lot more expensive to produce than the equivalent (G)DDRx chips plus addition PCB complexity, though that should change somewhat over time.
>>51916948
As in will APUs replace everything? maybe in the diatant future sure. HBM and better igpu is inevitable but so isbcrazy performance for gpus. Unless all companies agree to start making huge chips with like Titan X levels of performance on die which wouod never happen, we will have dedicated gpus for years to come.
Plus you have the problem of upgradability. Say your apu has the power of an r9 270. Years go by now the thing ia falling on its face in newer games. Now you have to go out and buy the new $250 chip and possibly new board because of new socket.
Where as now people are still running sandy bridge i5s and just going out to spend $150 or so on a newer mid range card.
Pcie may eventually be phased out if we ever fully saturate gen 3 (we are a ways off), but I dont see dedicated gpus leaving soon.
>>51916978
It would be shitty because the actual CPU itself would be based on a 3-year old architecture that was weak to begin with and a GPU architecture that's not well-suited for weaker cored CPUs. Coupling a GCN GPU with a much stronger CPU would be the best thing since sliced bread, but that combination doesn't exist because the necessary CPU doesn't exist yet.
But that is why a Zen-based APU will be a massive leap forward thanks to the die shrink and new CPU architecture. A flagship Zen APU will outperform current consoles with less power draw and heat. With HSA utilization and DDR4, Zen APUs might be able to minimize the bandwidth gap enough to make HBM unnecessary. Intel's Skylake with Iris Pro is coming out around the same time, so I fully expect the Zen APU to get BTFO by it, but it will almost certainly cost twice as much for less than 20% performance improvement.
>>51917076
>we ever fully saturate gen 3
We've fully saturated Gen 2, with the dual Fury and Titan X cards expected to touch the Gen 3's ceiling. PCIe Gen4 needs to take over after the next generation of cards.
>>51916700
>AMD
>caring about meme companies that won't exist in 5 years
AMD doesn't even have the right to transfer it's x86 license if it gets sold lmao, when it goes bankrupt in an year or two no one will want to buy this shit company.
>>51917125
Fujitsu grorious nippon folded steel fabs will save us
Thought the process was overall,nvidia cuda Intel and hunt of Gpu processing com parallel comparable unit test builds of CPU,why model untuned manufacturing obviate is not produce? Could name runs Titan,be Intel uses of cray processing and why?...
>>51917158
Wat
>>51917158
What?
Real hard core dar.bibalex.org?ISIS,az iz?.....orig amists
>>51916700
They will surely do that in the next couple of years if they don't go bankrupt in the meantime.
>>51916857
The bottleneck in the current PC architecture is not memory speed but the length of time it takes for the CPU to communicate with RAM, latency matters way more than RAM speed.
>>51916700
Thats the plan, one of the sole reasons behind HBM's structure was to serve their APUs instead of being a strictly GPU oriented memory.
Their next APU platform is called Raven Ridge, and its likely that it will include some HBM on package, though at this point nothing is confirmed. The immense area scaling benefits from 28nm down to 14nm could facilitate a relatively enormous IGP in a die that is still modest in size. They could pair 4 Zen cores with an IGP the size of whats inside the PS4 and end up with a 100mm2 95w part.
>>51916857
HBM isn't connected to the die. Its connected to an interposer on the package.
>>51917357
BS. If that were true then upgrading from 1600mhz to 2133mhz ram for an APU would have no impact on gaymen performance.
Why don't AMD just solder Intel CPU and AMD CPU together to create a hybrid cpu?
>>51916700
Yeah dude wtf xD, why don't AMD just like solder more HBM on their CPU? They're so dumb lol xD
>>51917125
The day that happens intel will go down with AMD unless they can jump on the ARM train which they failed to do.
>>51917357
Youre not wrong. But hbm on die is a dream at this point and having much faster and higher bandwidth ram that is cheapish and readily avilable (ddr4) will yield great results. Its a mix really. Speed, latency, and even physical distance.
>>51917407
He is partially right. Ram speed plays an inportant role but so does ram timing and the physical distance of ram from the chip believe it or not. Look it up. Interesting read about how cpus interact with different ram types and configurations.
>>51917407
That ebin poster was talking about CPUs. GPUs are highly parallel and generally do sequential reads which means prefetching hides the latency.
You can use the same strategy to optimize software that runs primarily on the CPU though.
>>51917458
This. AMD has intel by the balls with the AMD64 thing.
>>51917125
This isn't true. At all. There are so many nuances here is not even worth getting into.
The FTC ruled that intel must negotiate "in good faith" with any party who acquires either AMD or VIA, and they reserved the right to define what they meant by "in good faith."
In fact VIA's X86 license ran up, and they continued to produce CPUs uninterrupted. Guess why? Intel has no real claim over either AMD or VIA. The X86 license is basically 100% defunct.
>>51916700
>Why doesn't AMD just solder on 512-1024MB of HBM on APUs?
Because they only just now managed to do a test run for HBM, and they still have shit yields.
HBM2 is still said to be high-end only on GPUs, so I doubt they'll manage to put them on CPUs.
Maybe somewhere during the second or third 14mm refreshes, we'll see more HBM crap added... but only as glorified L4 caches for the onboard video.
>>51917793
>Guess why
Because there's no point in suing a company that barely exists as it is. Basically Intel will just have to pay for VIA to get destroyed. So losing money and losing face at the same time. Naturally they wouldn't do it.
But there is no reason why they won't run anyone who shows interest in AMD into the ground if he's even barely conceived as a potential threat to their industry domination.
For instance, Samsung. Samsung ain't getting the right to produce their own x86 CPUs ever. Pretty much any company on that scale and Intel will sue them to the ground.
So basically AMD is a meme, and will continue being one as long as it serves the purpose of monopoly lawsuit evasion for both Intel and Nvidia. It's a joke company.
>>51917959
Can't say I'm surprised to see even more childish shitposting.
intel has no legal ground whatsoever to prevent, or even impede a firm from acquiring either AMD or VIA. They literally do not have any say in the matter whatsoever. They could try to take the issue to court, which they would have to if they raised any objection at all, and it would fall back to the FTC who would simply tell intel that their case had no grounds to stand.
The X86 license is defunct. It is pointless. It is irrelevant. The only IP intel has that they can protect has to do with specific implementation.
But feel free to continue shitposting and regurgitate the same old forum trash over and over. If you want to be instantly discredited just keep bringing up the X86 license.
>>51916700
Because AMD is a shit.
>>51918032
Only the initial instruction set that was licensed 20 years ago cannot be patented anymore retard. Everything done after is very much under patent.
>Keeping in mind how important the cross-license agreement between Intel and AMD is, it is likely that any of AMD’s possible M&A actions will be discussed not only by the two companies, but also with the U.S. authorities, which prohibit imports of certain technologies to certain countries. In any case, it is noteworthy that the CFO of AMD does not reject the possible plan in general.
Sure Intel can't stop them from getting sold.
But they can stop them from using 9/10 of the fucking instruction set, whoop de doo.
>>51918122
Again, "in good faith."
Also again, the only IP intel has that they can protect has to do with specific implementation.
Intel, if they fought a long drawn out legal battle could have a shot at preventing a would be future Samsung X86 CPU with AVX or other intel created supplemental instructions. They could not prevent a Samsung X86 CPU, and the likely result would be that the FTC would quash any objections raised by intel for the sake of keeping market competition alive.
Like everyone has been saying, HBM is just too expensive. DDR4 is the only way we are getting faster vram for APUs. Hope AMD let's people use 3,200 Mhz DDR4, that would make the bandwidth starvation problem significantly less severe.
>>51917113
not really anon
https://www.techpowerup.com/reviews/NVIDIA/GTX_980_PCI-Express_Scaling/
https://www.techpowerup.com/reviews/AMD/R9_Fury_X_PCI-Express_Scaling/
>>51920420
>Dual Fury and Titan X
I meant the upcoming R9 Gemini and GTX 990.
>>51920530
in regards to dual gpu cards, the information stays on card due to the plx chip. if running 2 dual gpu cards (quadfire/quad sli), you'd still only need a tiny amount of bandwidth. bear in mind nvidia uses an external bridge for some of their data, so they're pcie usage is lower. even with amd's xdma engine, quadfire runs fine. you're talking single percentage points of improvement for more lanes
more lanes for pcie ssds though... now you're talking
>>51920632
their. how'd i fuck that one up?
>>51920632
The GTX 990 will use a PLX chip like the Fury Gemini, but I remember reading that the pre-production trial run for the dual Fiji card would almost saturate the bandwidth of a PCIE3 16x lanes when benchmarked at 1080p. The Titan X is rumored to saturate about 65% alone, so it's not that hard to imagine that a dual GM200 card would increase the bandwidth by 30%. I think the R9 295x2 was using about a quarter more bandwidth than a single R9 290X at 1080p, so it's not hard to imagine that more powerful cards would be able to use up almost the entire bandwidth of the PCIE3 lanes.
>>51920977
Doesn't nvidea have their own high bandwidth connection? Link or something?
>>51921667
nvlink. it's not for the consumer space. hpc only
