no kidding...maybe heat was the issue, or like most things, the maker in this case samsung is claiming numbers that are simply under the "best case scenario" not in an actual on product (GPU-CPU completed model) or I suppose, maybe AMD and now Nv decide to use a slower clocked variant at lower voltage so as reduce cost to THEM to maximize revenue?
Who knows, what I do know, the memory makers need stop being foolish and give the manufacturing capacity to keep end consumer pricing in check, very few DRAM/Flash makers out there, not hard for them to be in collusion i.e drive the price up massively benefits them and royally screws the end customer (the ones they sell it to) or at the very least screws us the consumer...the whole point of DDR to DDR2 etc and a smaller node was to keep pricing in check reduce heat/power, performance going up.
They seem to have gotten the memo on the reduce heat/power, not so much on the first and hit and miss on the 3rd point ^.^
I'm probably too late to this comment party to get a reply but if any of you uber nerds out there have an opinion on this I'd be interested to hear it.
I have 12GB of DDR3 in my system which is mostly used for gaming, web browsing, etc. Occasionally I do some video editing but it's not often enough to warrant more RAM. I have not noticed any issues relating to RAM bottlenecking performance (that's more likely to be down to the ageing but still competent GTX 780).
Given a RAM upgrade to DDR4 would cost me a new mobo, probably a new CPU (especially given it's one that Intel has warned their bug patching will affect considerably) as well as the RAM, does anyone think it's actually worth it? I personally can not see a reason to upgrade it but does anyone think it's going to be holding things back? The DDR3 is running at 1600MHz.
No need to upgrade to more RAM or DDR4 unless you are upgrading your whole system, which you already pointed out. Upgrading you will of course get performance enhancements, but only you know if the money you spend is worth the performance increase you receive.
As far as amount, you have plenty of RAM for gaming, web browsing. As far as your RAM speed goes, it is not likely holding you back any more that the CPU it is paired with. This is particularly true if you have triple channel memory. I would suggest your money is much better spent upgrading your still competent, but no longer exceptional GTX780. The GTX780 is a little short on VRAM for the newest games at high resolutions as well. It is possible that a particularly low end or very old processor and/or a lower resolution monitor (you didn't specify these) will place the bottleneck on the CPU, but it is less likely and RAM is still unlikely to be the issue. Also keep in mind that at this point it looks like gaming performance loss is usually around 3% with a single outlier of around 9%. If you play mostly CPU bound games (I.E. RTS at low resolution), then you may want to consider this. Your safest bet is to simply wait for the patch and play your favorite games. If they don't feel much different, then don't worry about the CPU (from a gaming perspective).
Video editing might possibly suffer with the new security patch. It depends on how IO intensive your software / projects are as well as how fast the storage is. Faster storage (I.E. NVME SSD) will see a larger hit while the hit to HDDs may not be noticeable in this scenario. This won't be related to RAM, though.
I suppose you rather mean 2 Gbps (clocked at 1 Ghz), right? It's possible that there were serious TDP and/or power consumption issues with the Flarebolt dies, since they were effectively overclocked. Samsung raised the voltage above spec to 1.35 V, so the power and thermal budget probably went through the roof, as always happens when you raise voltages. However Samsung's new 2.4 Gbps dies went back to 1.2 V, so this time there will be no TDP and power issues. The Aquabolt dies are both much faster and markedly more power efficient. If SK Hynix's 2.4 Gbps dies are in the ~1.35 V range they are probably toast.
p.s. I checked the linked pdf of SK Hynix. Their 2.4 Gbps dies are also voltaged at 1.2 V, both their 4- and 8-stack. That's good, because it means that there will be healthy competition and 2.4 Gbps HBM2 prices will not go through the roof.
Would it be possible to put something like this on top of a regular Intel CPU? 1.2 TB memory bandwidth would be pretty insane. Although I'm not sure it will matter for regular usage like gaming?
Although it would certainly be great for graphics, I was thinking more along the lines of wether it could be used as a complete replacement for regular DDR4 memory. So we could get rid of the dimm slots altogether and just have 32GB ram sitting on top of the CPU.
"On top" is very problematic for a high performance chip, as the chip would heat the HBM directly instead of the heat sink. This would be unhealthy for the memory and present a huge added thermal resistance for the CPU and hence drive its temperature up (given similar cooling).
HBM can be added besides the HPC chip, though, as others have already commented. There are at least 3 issues with that:
- Look at current DRAM frequency scaling. There's not much benefit in higher bandwidth. Which makes sense in the way that the current CPUs are well balanced designs and not starved for bandwidth. Sure, more helps a little bit here and there, but a single HBM stack would already suffice.
- Capacity lock-in. Who are you to dictate everyone's going to be fine with x GB of memory? They could offer some variation with 1 to 8-Hi stacks, but that's not much capacity and they'd create seperate SKUs for all those capacity points (not good for the manufacturer, as they'll overproduce more CPUs which turn out to sell worse than anticipated). And look at it from your side: would you want to "throw away" your probably expensive CPU everytime you want/need more memory?
- Cost: rumor puts a current high performance HBM2 stack at 80$. Sounds excessive but probably includes the relatively expensive packaging and sub-optimal yields. add some vendor margin and Imagine adding 4 of these to a CPU - suddenly it costs 400$ more. That's a tough pill to swallow and way above DDR4 pricing.
It is alwasy suggested to put the memory on top of the cpu (or gpu) die which of course leads to heat issues. Why not put the cpu on top of the memory?
The board needs the CPU, the CPU needs the memory. If you wanted to put the CPU on top of the memory, you'd need to route everything through the memory to the motherboard, which would be kind of weird and probably pretty expensive.
The problem is that the quad Core ix series pretty well tops out around 20GBps. Anything below that slows the CPU down, anything much above that isn't used.
Now once you start adding more cores and improving the memory controller that's different, but right now DDR4 is feeding it quite well.
There's a benchmark from a couple of years ago that shows this. The numbers may have increased a bit, but we're still not going to see a huge improvement going from 20GBps to 100GBps bandwidth on this architecture.
One of the very nice things about putting the CPU and GPU on the same RAM, even if the CPU does not need that bandwidth, is that there is never a need to copy the data. It is just there in RAM available for either processor to use.
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
17 Comments
Back to Article
DanNeely - Thursday, January 11, 2018 - link
It'll be interesting to see if this shows up anywhere at its rated speed. Neither AMD or NVidia have managed to get HBM2 at 2ghz into anything yet.Dragonstongue - Thursday, January 11, 2018 - link
no kidding...maybe heat was the issue, or like most things, the maker in this case samsung is claiming numbers that are simply under the "best case scenario" not in an actual on product (GPU-CPU completed model) or I suppose, maybe AMD and now Nv decide to use a slower clocked variant at lower voltage so as reduce cost to THEM to maximize revenue?Who knows, what I do know, the memory makers need stop being foolish and give the manufacturing capacity to keep end consumer pricing in check, very few DRAM/Flash makers out there, not hard for them to be in collusion i.e drive the price up massively benefits them and royally screws the end customer (the ones they sell it to) or at the very least screws us the consumer...the whole point of DDR to DDR2 etc and a smaller node was to keep pricing in check reduce heat/power, performance going up.
They seem to have gotten the memo on the reduce heat/power, not so much on the first and hit and miss on the 3rd point ^.^
philehidiot - Saturday, January 13, 2018 - link
I'm probably too late to this comment party to get a reply but if any of you uber nerds out there have an opinion on this I'd be interested to hear it.I have 12GB of DDR3 in my system which is mostly used for gaming, web browsing, etc. Occasionally I do some video editing but it's not often enough to warrant more RAM. I have not noticed any issues relating to RAM bottlenecking performance (that's more likely to be down to the ageing but still competent GTX 780).
Given a RAM upgrade to DDR4 would cost me a new mobo, probably a new CPU (especially given it's one that Intel has warned their bug patching will affect considerably) as well as the RAM, does anyone think it's actually worth it? I personally can not see a reason to upgrade it but does anyone think it's going to be holding things back? The DDR3 is running at 1600MHz.
SilthDraeth - Saturday, January 13, 2018 - link
No need to upgrade to more RAM or DDR4 unless you are upgrading your whole system, which you already pointed out. Upgrading you will of course get performance enhancements, but only you know if the money you spend is worth the performance increase you receive.BurntMyBacon - Monday, January 15, 2018 - link
As far as amount, you have plenty of RAM for gaming, web browsing. As far as your RAM speed goes, it is not likely holding you back any more that the CPU it is paired with. This is particularly true if you have triple channel memory. I would suggest your money is much better spent upgrading your still competent, but no longer exceptional GTX780. The GTX780 is a little short on VRAM for the newest games at high resolutions as well. It is possible that a particularly low end or very old processor and/or a lower resolution monitor (you didn't specify these) will place the bottleneck on the CPU, but it is less likely and RAM is still unlikely to be the issue. Also keep in mind that at this point it looks like gaming performance loss is usually around 3% with a single outlier of around 9%. If you play mostly CPU bound games (I.E. RTS at low resolution), then you may want to consider this. Your safest bet is to simply wait for the patch and play your favorite games. If they don't feel much different, then don't worry about the CPU (from a gaming perspective).Video editing might possibly suffer with the new security patch. It depends on how IO intensive your software / projects are as well as how fast the storage is. Faster storage (I.E. NVME SSD) will see a larger hit while the hit to HDDs may not be noticeable in this scenario. This won't be related to RAM, though.
Santoval - Thursday, January 11, 2018 - link
I suppose you rather mean 2 Gbps (clocked at 1 Ghz), right? It's possible that there were serious TDP and/or power consumption issues with the Flarebolt dies, since they were effectively overclocked. Samsung raised the voltage above spec to 1.35 V, so the power and thermal budget probably went through the roof, as always happens when you raise voltages. However Samsung's new 2.4 Gbps dies went back to 1.2 V, so this time there will be no TDP and power issues. The Aquabolt dies are both much faster and markedly more power efficient. If SK Hynix's 2.4 Gbps dies are in the ~1.35 V range they are probably toast.Santoval - Thursday, January 11, 2018 - link
p.s. I checked the linked pdf of SK Hynix. Their 2.4 Gbps dies are also voltaged at 1.2 V, both their 4- and 8-stack. That's good, because it means that there will be healthy competition and 2.4 Gbps HBM2 prices will not go through the roof.btb - Thursday, January 11, 2018 - link
Would it be possible to put something like this on top of a regular Intel CPU? 1.2 TB memory bandwidth would be pretty insane. Although I'm not sure it will matter for regular usage like gaming?Elstar - Thursday, January 11, 2018 - link
Intel's Xeon Phi product line has 16 GiB of high-bandwidth "MCDRAM" in addition to normal memory.ianmills - Thursday, January 11, 2018 - link
The new Intel chip with radeon graphics has HBM on the package but as a separate die. It that what you mean?btb - Friday, January 12, 2018 - link
Although it would certainly be great for graphics, I was thinking more along the lines of wether it could be used as a complete replacement for regular DDR4 memory. So we could get rid of the dimm slots altogether and just have 32GB ram sitting on top of the CPU.MrSpadge - Friday, January 12, 2018 - link
"On top" is very problematic for a high performance chip, as the chip would heat the HBM directly instead of the heat sink. This would be unhealthy for the memory and present a huge added thermal resistance for the CPU and hence drive its temperature up (given similar cooling).HBM can be added besides the HPC chip, though, as others have already commented. There are at least 3 issues with that:
- Look at current DRAM frequency scaling. There's not much benefit in higher bandwidth. Which makes sense in the way that the current CPUs are well balanced designs and not starved for bandwidth. Sure, more helps a little bit here and there, but a single HBM stack would already suffice.
- Capacity lock-in. Who are you to dictate everyone's going to be fine with x GB of memory? They could offer some variation with 1 to 8-Hi stacks, but that's not much capacity and they'd create seperate SKUs for all those capacity points (not good for the manufacturer, as they'll overproduce more CPUs which turn out to sell worse than anticipated). And look at it from your side: would you want to "throw away" your probably expensive CPU everytime you want/need more memory?
- Cost: rumor puts a current high performance HBM2 stack at 80$. Sounds excessive but probably includes the relatively expensive packaging and sub-optimal yields. add some vendor margin and Imagine adding 4 of these to a CPU - suddenly it costs 400$ more. That's a tough pill to swallow and way above DDR4 pricing.
SunnyNW - Friday, January 12, 2018 - link
It is alwasy suggested to put the memory on top of the cpu (or gpu) die which of course leads to heat issues. Why not put the cpu on top of the memory?mkozakewich - Friday, January 12, 2018 - link
The board needs the CPU, the CPU needs the memory. If you wanted to put the CPU on top of the memory, you'd need to route everything through the memory to the motherboard, which would be kind of weird and probably pretty expensive.0ldman79 - Saturday, January 13, 2018 - link
The problem is that the quad Core ix series pretty well tops out around 20GBps. Anything below that slows the CPU down, anything much above that isn't used.Now once you start adding more cores and improving the memory controller that's different, but right now DDR4 is feeding it quite well.
There's a benchmark from a couple of years ago that shows this. The numbers may have increased a bit, but we're still not going to see a huge improvement going from 20GBps to 100GBps bandwidth on this architecture.
Zan Lynx - Monday, January 15, 2018 - link
One of the very nice things about putting the CPU and GPU on the same RAM, even if the CPU does not need that bandwidth, is that there is never a need to copy the data. It is just there in RAM available for either processor to use.[email protected] - Thursday, June 28, 2018 - link
What is the formula and explanation that you get these values above?307.2 GB/s 256 GB/s 204.8 GB/s
I know for IOPs etc.... but this one I am not sure.