If Intel can do it, then so could AMD and NVIDIA. I'm not sure there would be much weeping, but there would certainly be some solid competition for budget/low end graphics.
Isn't Intel rolling their own with Micron - MCDRAM?
(to be clear: marketing *claims* it's just an order of magnitude slower, but that may turn out to be overly optimistic, especially for writes, and especially if a somewhat complex controller is required)
intel is already putting large L4 cache's on their IRIS GPU's. .It's pretty damn fast by itself. increasing the amount will help some, but not a revolution in performance like you are imagining.
It'd be nice if Intel was able to overcome the bandwidth limitations of current shared system memory with their graphics as well as supplying the processor with fast memory that's physically close to it. I think the benefit in that is a further reduction in the need to package dedicated graphics that demand additional power and PCB space while adding complexity that's only needed when performing graphics intensive tasks. It might hurt the bottom line at the two big dGPU manufacturers and make people like me happy as long as the heat output doesn't dramatically increase so that it spoils the trend toward simpler, passive cooling. It's been a nice change and I'd hate to see that focus on battery life and simplification die on the vine because of a new memory technology that isn't strictly necessary as current Intel HD graphics are very good performers already.
Yes, I think I would trade 64MB eDRAM for 2GB HBM2. Or Heck just use a Single 8GB HBM2 as main memory. The current Iris Pro is really limited by its bandwidth.
It's sale price has little to do with it's manufacturing cost. It's hard to say how much the price would go up, if any. They are priced, basically, as high as the market will bear, regardless of the cost.
this would also be one clear cut way for amd to get back into the small formfactor / mobile market.
push out a 2/4 core low power zen apu with cgn 1.3 onboard and stack one 4gb hbm2 model on the package: et voila done is your x86 tablet chip. no extra ram, small package, power efficient. if they manage that (and of course if zen lives up to expectations in terms of performance and efficiency) this would be a real contender for intels atom.
same thing goes for laptops up to and including surface pro class devices. imagine a zen-cpu with decent cgn based onboard graphics (that still blows intels iris out of the water easily) backed up by 16GB HBM2, shared between cpu/gpu. this has real beast potential.
also i would guess a zen-cpu with 16/24/32 gb hbm2 on board would be a real nice base for a workstation and also gaming pc for vr combined wiht a fitting discrete graphics board. there is no need for ddr-ram slots in 99.9% of these cases with 32gb easily being enough for current demands. only server and very specialized workstation scenarios would even need more than that so this in turn opens up a market for mainboards without those slots, reducing board-complexity, space requirements and also production-price.
great potential for good stuff whichever way you look at it.
It would be really interesting to see an HBM APU. Though I guess this would be targeted for small-form-factor (read: soldered CPU, even non-upgradable RAM) or some other custom solutions that fancies some. Of course there's the next gen of consoles. Would be good to see in Surface-like devices as pretty much locked-up upgradability-wise (for debatable reasons). Only thing left waiting is for AMD to gather their shit up and manage to get some design wins.
that's a pretty fascinating prospect. even with a smaller amount of RAM, with modern or next gen SSDs and storage interfaces, swap may be feasible again...
Most likely, we'd use 3D-XPoint or classic RAM for that swap when a chip has enough HBM. Really not much reason to have swap if you have enoug memory. That said, for most users, 8-16GB on-package HBM would be plenty for essentially everything
Couple that with Zen and you got me extremely excited. I seriously hope Zen wont be a letdown. If it provides 90%+ ST of whatever Intel has to offer by that time, I'd consider it a huge win. You either get Zen based CPUs or Intel dramatically drops its prices.
That said, my Haswell chip is still good for a couple more years. Lets see what happens by then.
There are two things I'd love to see this integrated into:
1) An APU. A socket-replaceable AMD APU with HBM for the graphics would be quite potent. At that point you'd be looking at a serious jump in performance because current implementations of AMDs iGPU require access to system ram. A dedicated 2GB of HBM would solve that bottleneck easily.
2) System RAM and iGPU in notebooks. 8GB in a laptop for system RAM with another 2GB or 4GB for the iGPU would enable some stupidly fast configurations while also saving power. Right now DDR4 maxes out at about 50-60GB/sec, you could drive HBM system RAM at about 1/4 the speed it's capable of for power savings and still increase the memory bandwidth by about 4x. Most laptops already are soldered RAM setups, so this wouldn't be any different.
I am wondering about using HBM as the main system memory too. Can it be done, standard wise, or does JEDEC restrict it to video memory usage only?
Then again, could it be that there isn't much of a need for such high bandwidth memory for desktop/laptop use cases?
On the other hand, seeing that XOne and PS4 are using much higher bandwidth memory as system memory, shouldn't that gap be closed on PCs?
Surely porting a PS4 game, that takes full advantage of its memory, to PC wouldn't be an easy task, and the developers might be forced to sacrifice some graphical fidelity or details or performance along the way.
I doubt we will see HBM used as main system memory for quite some time, if ever. Desktop CPU's just don't need memory to be much faster than it is. Server CPU's with lots of cores do need quite a bit of B/W but they also need a TON of capacity. It would be impossible to get 256GB of HBM + a big server CPU on an interposer. Maybe see it used as a level of cache, sure, but not main memory.
A large server chip with 8gb of "L4" cache would give a serious boost to certain ops - especially anything that likes caching (web servers, pure databases, etc) with a few hundred gig of 'swap' in reserve.
You really don't want HBM2 as main system memory. It needs a silicon interposer interconnect to function. So you can't put it down on a PCB. So either you need to route all data calls through the CPU package or you need to have a prohibitively expensive mainboard area. Neither are great options.
Nonsense. Clearly they have been holding back to milk every dime out of you sheep. Especially AMD they have been secretly holding back HBM tech for decades. They had HBM while they were still selling you overpriced DDR1 systems. Even now, they're intentionally losing money merely as a ruse to make you really believe the tech isn't ready for release.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
42 Comments
Back to Article
iwod - Wednesday, January 20, 2016 - link
Assuming using the same power as HBM1 while providing same bandwidth, I could expect for HBM2 4GB Memory to use 3.75W power?I wonder if this will be a good thing for mobile graphics. ( In terms of Laptop )
saratoga4 - Wednesday, January 27, 2016 - link
Seems like it would be a no-brainer on laptops, and probably even smartphones once the power can be scaled low enough.Aspiring Techie - Wednesday, January 20, 2016 - link
Put 4GB of this stuff with Iris Pro graphics and watch Nvidia and AMD weep. OR use it as a massive L4 cache on an i7...nathanddrews - Wednesday, January 20, 2016 - link
If Intel can do it, then so could AMD and NVIDIA. I'm not sure there would be much weeping, but there would certainly be some solid competition for budget/low end graphics.Isn't Intel rolling their own with Micron - MCDRAM?
Jimbo11 - Wednesday, January 20, 2016 - link
Micron also has HMC, it will soon to use Xpoint, which is passive and substantial larger memory ( 128GB), low power, it will be a killeremn13 - Friday, January 22, 2016 - link
XPoint is at an order of magnitude slower, and although it has brilliant endurance compared to nand it has terrible endurance compared to dram.It doesn't sound likely to me that it'll replace dram very quickly, if ever. Perhaps eventually it might push dram into an l4 cache niche...
Of course, what you call cache and what memory is a bit of an artificial distinction.
emn13 - Friday, January 22, 2016 - link
(to be clear: marketing *claims* it's just an order of magnitude slower, but that may turn out to be overly optimistic, especially for writes, and especially if a somewhat complex controller is required)T1beriu - Wednesday, January 20, 2016 - link
Yeah, because the GPU's power is based on bandwidth...Qwertilot - Wednesday, January 20, 2016 - link
Well, no, but it has been a pretty horrible bottleneck for iGPU's :)jasonelmore - Wednesday, January 20, 2016 - link
intel is already putting large L4 cache's on their IRIS GPU's. .It's pretty damn fast by itself. increasing the amount will help some, but not a revolution in performance like you are imagining.nikaldro - Wednesday, January 20, 2016 - link
Actually, iris pro is quite memory starved, and GT4e will be even more starved.StevoLincolnite - Saturday, January 23, 2016 - link
Not as bandwidth starved as you think though. Even Intel admitted to that.http://www.anandtech.com/show/7364/memory-scaling-...
Intel stated that the L4 once it exceeded 32Mb in size started to show diminishing returns, but quad-drupled the size for future scenario's.
BrokenCrayons - Wednesday, January 20, 2016 - link
It'd be nice if Intel was able to overcome the bandwidth limitations of current shared system memory with their graphics as well as supplying the processor with fast memory that's physically close to it. I think the benefit in that is a further reduction in the need to package dedicated graphics that demand additional power and PCB space while adding complexity that's only needed when performing graphics intensive tasks. It might hurt the bottom line at the two big dGPU manufacturers and make people like me happy as long as the heat output doesn't dramatically increase so that it spoils the trend toward simpler, passive cooling. It's been a nice change and I'd hate to see that focus on battery life and simplification die on the vine because of a new memory technology that isn't strictly necessary as current Intel HD graphics are very good performers already.iwod - Wednesday, January 20, 2016 - link
Yes, I think I would trade 64MB eDRAM for 2GB HBM2. Or Heck just use a Single 8GB HBM2 as main memory. The current Iris Pro is really limited by its bandwidth.Samus - Wednesday, January 20, 2016 - link
As Iris Pro is already overpriced, I don't see bundling it with an exotic memory solution helping much...extide - Wednesday, January 20, 2016 - link
It's sale price has little to do with it's manufacturing cost. It's hard to say how much the price would go up, if any. They are priced, basically, as high as the market will bear, regardless of the cost.dustwalker13 - Sunday, January 24, 2016 - link
this would also be one clear cut way for amd to get back into the small formfactor / mobile market.push out a 2/4 core low power zen apu with cgn 1.3 onboard and stack one 4gb hbm2 model on the package: et voila done is your x86 tablet chip. no extra ram, small package, power efficient. if they manage that (and of course if zen lives up to expectations in terms of performance and efficiency) this would be a real contender for intels atom.
same thing goes for laptops up to and including surface pro class devices. imagine a zen-cpu with decent cgn based onboard graphics (that still blows intels iris out of the water easily) backed up by 16GB HBM2, shared between cpu/gpu. this has real beast potential.
also i would guess a zen-cpu with 16/24/32 gb hbm2 on board would be a real nice base for a workstation and also gaming pc for vr combined wiht a fitting discrete graphics board. there is no need for ddr-ram slots in 99.9% of these cases with 32gb easily being enough for current demands. only server and very specialized workstation scenarios would even need more than that so this in turn opens up a market for mainboards without those slots, reducing board-complexity, space requirements and also production-price.
great potential for good stuff whichever way you look at it.
Mugur - Tuesday, February 2, 2016 - link
Imagine a next gen console with a Zen derivative SoC and 16GB of HBM in the package... 4k gaming on 4k TVs.Eden-K121D - Friday, June 17, 2016 - link
Won't the latency affect CPU based workloadsOEMG - Wednesday, January 20, 2016 - link
It would be really interesting to see an HBM APU. Though I guess this would be targeted for small-form-factor (read: soldered CPU, even non-upgradable RAM) or some other custom solutions that fancies some. Of course there's the next gen of consoles. Would be good to see in Surface-like devices as pretty much locked-up upgradability-wise (for debatable reasons). Only thing left waiting is for AMD to gather their shit up and manage to get some design wins.plonk420 - Wednesday, January 20, 2016 - link
that's a pretty fascinating prospect. even with a smaller amount of RAM, with modern or next gen SSDs and storage interfaces, swap may be feasible again...ZeDestructor - Wednesday, January 20, 2016 - link
Most likely, we'd use 3D-XPoint or classic RAM for that swap when a chip has enough HBM. Really not much reason to have swap if you have enoug memory. That said, for most users, 8-16GB on-package HBM would be plenty for essentially everythingBeany2013 - Wednesday, January 20, 2016 - link
HBM2 with a decent APU and an M.2 SSD would allow something seriously small and seriously, seriously potent, proving the cooling could keep up.Interesting times ahead, I suspect.
lilmoe - Wednesday, January 20, 2016 - link
Couple that with Zen and you got me extremely excited. I seriously hope Zen wont be a letdown. If it provides 90%+ ST of whatever Intel has to offer by that time, I'd consider it a huge win. You either get Zen based CPUs or Intel dramatically drops its prices.That said, my Haswell chip is still good for a couple more years. Lets see what happens by then.
Michael Bay - Sunday, January 24, 2016 - link
>amd>90% of intel STP
Who`s your dealer? I`m bying!
Not even their own insane marketing is promising this.
bill.rookard - Wednesday, January 20, 2016 - link
There are two things I'd love to see this integrated into:1) An APU. A socket-replaceable AMD APU with HBM for the graphics would be quite potent. At that point you'd be looking at a serious jump in performance because current implementations of AMDs iGPU require access to system ram. A dedicated 2GB of HBM would solve that bottleneck easily.
2) System RAM and iGPU in notebooks. 8GB in a laptop for system RAM with another 2GB or 4GB for the iGPU would enable some stupidly fast configurations while also saving power. Right now DDR4 maxes out at about 50-60GB/sec, you could drive HBM system RAM at about 1/4 the speed it's capable of for power savings and still increase the memory bandwidth by about 4x. Most laptops already are soldered RAM setups, so this wouldn't be any different.
eddman - Wednesday, January 20, 2016 - link
I am wondering about using HBM as the main system memory too. Can it be done, standard wise, or does JEDEC restrict it to video memory usage only?Then again, could it be that there isn't much of a need for such high bandwidth memory for desktop/laptop use cases?
On the other hand, seeing that XOne and PS4 are using much higher bandwidth memory as system memory, shouldn't that gap be closed on PCs?
Surely porting a PS4 game, that takes full advantage of its memory, to PC wouldn't be an easy task, and the developers might be forced to sacrifice some graphical fidelity or details or performance along the way.
extide - Wednesday, January 20, 2016 - link
I doubt we will see HBM used as main system memory for quite some time, if ever. Desktop CPU's just don't need memory to be much faster than it is. Server CPU's with lots of cores do need quite a bit of B/W but they also need a TON of capacity. It would be impossible to get 256GB of HBM + a big server CPU on an interposer. Maybe see it used as a level of cache, sure, but not main memory.extide - Wednesday, January 20, 2016 - link
(I was responding to #2) -- totally agree with an APU having HBM as graphics memory/cache, but DDR3/4 for system memory.Meaker10 - Wednesday, January 20, 2016 - link
There is no reason an HBM chip can't be pin compatible with current sockets so it could be done.Beany2013 - Wednesday, January 20, 2016 - link
A large server chip with 8gb of "L4" cache would give a serious boost to certain ops - especially anything that likes caching (web servers, pure databases, etc) with a few hundred gig of 'swap' in reserve.Very, very interesting times head...
stadisticado - Wednesday, January 20, 2016 - link
You really don't want HBM2 as main system memory. It needs a silicon interposer interconnect to function. So you can't put it down on a PCB. So either you need to route all data calls through the CPU package or you need to have a prohibitively expensive mainboard area. Neither are great options.damianrobertjones - Wednesday, January 20, 2016 - link
Another jump forward which, probably, could have happened years back but when it's all about the $$$$... .extide - Wednesday, January 20, 2016 - link
Well, the TSV's and microbumps that enabled this tech are relatively new. Perhaps it could have been a bit earlier, but not really much.Alexvrb - Wednesday, January 20, 2016 - link
Nonsense. Clearly they have been holding back to milk every dime out of you sheep. Especially AMD they have been secretly holding back HBM tech for decades. They had HBM while they were still selling you overpriced DDR1 systems. Even now, they're intentionally losing money merely as a ruse to make you really believe the tech isn't ready for release.HollyDOL - Thursday, January 21, 2016 - link
Yep, and Intel had secret 7nm fabs since 1979 :-)Oxford Guy - Friday, January 22, 2016 - link
Holding back is a thing, though.stephenbrooks - Friday, January 22, 2016 - link
Usually it'll be the front-runner that holds back, with the 2nd-place competitor throwing everything they've got.menting - Thursday, January 21, 2016 - link
stacking dies that thin with that many TSVs isn't easy. Not to mention just fabbing the TSVs are not easy either.Murloc - Thursday, January 21, 2016 - link
yeah right....if these companies exist it's because someone puts money into them in order to earn more money.
Michael Bay - Sunday, January 24, 2016 - link
Newsflash: IT`S ALL ABOUT THE MONEY.XmppTextingBloodsport - Saturday, March 19, 2016 - link
"Why should I not mix controllers-, timings-, or denisities?"