-
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 going DDR4 any time soon?
- Thread starter RobHague
- Start date
Caporegime
- Joined
- 18 Sep 2009
- Posts
- 30,503
- Location
- Dormanstown.
2016 hopefully.
Currently and for some time to come, DDR4 is and will continue to be vastly overrated, pretty much useless and anyone buying a DDR4 system now are effectively flushing their money down the drain. That will remain true until it's properly supported by AMD, whose superior implementation of it will be amazingly awesome and make all current DDR4 owners deeply regret wasting their money and bringing such shame upon themselves and their forever-tainted enthusiast reputations.
...Seriously tho, fingers-crossed they make a new platform with DDR4, SATA-E, much improved IPC on 8+ cores, and price it to humiliate Broadwell or Skylake quads. Ugh, if Intel are still pushing expensive quads by 2016 and AMD don't have much of note, part two of my killing spree will commence!
...Seriously tho, fingers-crossed they make a new platform with DDR4, SATA-E, much improved IPC on 8+ cores, and price it to humiliate Broadwell or Skylake quads. Ugh, if Intel are still pushing expensive quads by 2016 and AMD don't have much of note, part two of my killing spree will commence!
Last edited:
+1 to this DDR4 atm is not worth buying you wont see much noticeable difference. Important set of words there, 'noticeable difference'.
http://www.anandtech.com/show/7843/testing-sata-express-with-asus
I haven't been keeping up the past few months but I don't think there are any proper SATA-E SSDs on the market yet.
Heheh, admittedly I wasn't completely joking with that
Soldato
- Joined
- 21 Jul 2005
- Posts
- 21,104
- Location
- Officially least sunny location -Ronskistats
And for most of us, going DDR4 and intel is seriously going overbudget! 
Soldato
- Joined
- 9 Jun 2011
- Posts
- 3,639
DDR4? lol... that's last of their worries.... They should worry about making a CPU which can beat a Overclocked Sandy Bridge....
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
DDR4 kind of died in terms of excitement/need. I can't remember the exact speeds but DDR4 was targeted to launch around 2400Mhz speeds and scale upwards of 3200Mhz(officially and higher later on) but effectively with the same bandwidth DDR3 could achieve at the same speeds, just lower voltage/power. Problem was DDR3 was only designed for what 2133Mhz max so there was a nice large gap in performance. DDR3 kept getting improved up to and beyond 2800Mhz so DDR4 original speeds were all but useless and the current speeds aren't much better.
IE 16gb of 2400Mhz ddr3 is currently cheaper than 2400Mhz ddr4 , but with tighter timings and without having played with or not seen many benchmarks, wouldn't be surprised if performance was a bit better due to tighter timings but comparing different platforms makes that difficult.
Now if DDR3 had never gone past 2133Mhz and ddr4 hit 2400Mhz a couple years ago and was at 3000Mhz now and 20-30% beyond anything ddr3 could do it would be great and the industry would transition to it much more quickly.
DDR4 is more like a very minor power saving tech version of ddr3 rather than something brand spanking new.
Realistically the next big bumps in bandwidth/performance will come from HBM/HMC. Sticking anything from 4-8GB on die and designing cpu's to use some extremely high bandwidth memory alongside slower and larger pools of memory off die.
In 2 years you'll be able to stick one stack of 8gb memory with 256GB/s of bandwidth on package, not cheaply but not that expensive. I'm not sure which way they'd go, 1-2 small stacks of memory giving between 2-4gb adding probably £30-40 to cpu cost and using it more like cache. Or going 8gb and using it less like cache and more like switching entire programs in an out of memory. IE load a game and it loads into the 8gb on package while moving IE/firefox/some IDE/explorer/video/whatever into the rest of system memory so switching between things is still fast but when something wants full performance it will sit in memory with 5-10 times the performance.
Intel seems intent on the HMC version which is fast but higher power and sits off package, it would likely need to be soldered to a mobo meaning increased costs and no upgradability. AMD has gone with HBM which goes on package with the CPU, it's way lower power and offers basically the same high speed. Both have downsides, both add costs because at the very least you're adding memory, but increasing performance is increasing performance. At some point to get significantly higher bandwidth you have to move away from interchangeable slotted memory to something more permanent and closer to the cpu. Depending on how it's used, cache or more like adding to the total memory, it could mean spending less on normal memory anyway.
The main limit in normal memory bandwidth is both the length of traces between the memory controller and the memory(increased trace length means more signal power required which increases significantly, as does increasing the number of traces, the number of traces also increase motherboard complexity and cost and pinout on the die. Ultimately all these things limit the number of slots within a given cost and increase power. Just look at how much power a quad or hexcore Haswell-E use, the massive jump in power is in large part down to the quad channel memory.
You can double bandwidth by doubling traces, power usage, drastically increasing motherboard cost and increasing cost in memory(needing double the number of sticks). HBM lets you stick as low as 1GB per stack but giving 128gb/s of bandwidth with it. The main reason being traces at the silicon scale let you go from dozens to thousands of traces without a problem. Power wise the distances involved are reduced by a magnitude or two(more like 50CM to 0.5cm, keep in mind the straight line distance isn't the real distances, traces go along multiple levels of the mobo and over far greater distances). HMC would likely be a bank of chips alongside the socket. They use a serial bus(think sata ports over parallel, changing the signalling to go from 200+ pins to significantly less. While closer to the socket it's still closer to 5CM traces than 0.5CM, HBM offers another magnitude reduction in distances which is why it's so much lower power. The bandwidth increase on HMC comes from the signalling, rather than adding 5 times the number of traces it makes each trace much more efficient.
We're unlikely to see significant performance increases in general system memory in the future, HBM/HMC and future similar technologies will be where memory performance comes from.
So basically DDR4(ddr5----ddr15) are likely to be minor improvements now, little better power, but will become a step between ssd/on die memory rather than the primary source of in use data for the CPU as it is now.
IE 16gb of 2400Mhz ddr3 is currently cheaper than 2400Mhz ddr4 , but with tighter timings and without having played with or not seen many benchmarks, wouldn't be surprised if performance was a bit better due to tighter timings but comparing different platforms makes that difficult.
Now if DDR3 had never gone past 2133Mhz and ddr4 hit 2400Mhz a couple years ago and was at 3000Mhz now and 20-30% beyond anything ddr3 could do it would be great and the industry would transition to it much more quickly.
DDR4 is more like a very minor power saving tech version of ddr3 rather than something brand spanking new.
Realistically the next big bumps in bandwidth/performance will come from HBM/HMC. Sticking anything from 4-8GB on die and designing cpu's to use some extremely high bandwidth memory alongside slower and larger pools of memory off die.
In 2 years you'll be able to stick one stack of 8gb memory with 256GB/s of bandwidth on package, not cheaply but not that expensive. I'm not sure which way they'd go, 1-2 small stacks of memory giving between 2-4gb adding probably £30-40 to cpu cost and using it more like cache. Or going 8gb and using it less like cache and more like switching entire programs in an out of memory. IE load a game and it loads into the 8gb on package while moving IE/firefox/some IDE/explorer/video/whatever into the rest of system memory so switching between things is still fast but when something wants full performance it will sit in memory with 5-10 times the performance.
Intel seems intent on the HMC version which is fast but higher power and sits off package, it would likely need to be soldered to a mobo meaning increased costs and no upgradability. AMD has gone with HBM which goes on package with the CPU, it's way lower power and offers basically the same high speed. Both have downsides, both add costs because at the very least you're adding memory, but increasing performance is increasing performance. At some point to get significantly higher bandwidth you have to move away from interchangeable slotted memory to something more permanent and closer to the cpu. Depending on how it's used, cache or more like adding to the total memory, it could mean spending less on normal memory anyway.
The main limit in normal memory bandwidth is both the length of traces between the memory controller and the memory(increased trace length means more signal power required which increases significantly, as does increasing the number of traces, the number of traces also increase motherboard complexity and cost and pinout on the die. Ultimately all these things limit the number of slots within a given cost and increase power. Just look at how much power a quad or hexcore Haswell-E use, the massive jump in power is in large part down to the quad channel memory.
You can double bandwidth by doubling traces, power usage, drastically increasing motherboard cost and increasing cost in memory(needing double the number of sticks). HBM lets you stick as low as 1GB per stack but giving 128gb/s of bandwidth with it. The main reason being traces at the silicon scale let you go from dozens to thousands of traces without a problem. Power wise the distances involved are reduced by a magnitude or two(more like 50CM to 0.5cm, keep in mind the straight line distance isn't the real distances, traces go along multiple levels of the mobo and over far greater distances). HMC would likely be a bank of chips alongside the socket. They use a serial bus(think sata ports over parallel, changing the signalling to go from 200+ pins to significantly less. While closer to the socket it's still closer to 5CM traces than 0.5CM, HBM offers another magnitude reduction in distances which is why it's so much lower power. The bandwidth increase on HMC comes from the signalling, rather than adding 5 times the number of traces it makes each trace much more efficient.
We're unlikely to see significant performance increases in general system memory in the future, HBM/HMC and future similar technologies will be where memory performance comes from.
So basically DDR4(ddr5----ddr15) are likely to be minor improvements now, little better power, but will become a step between ssd/on die memory rather than the primary source of in use data for the CPU as it is now.
Last edited:
Soldato
- Joined
- 5 Sep 2011
- Posts
- 12,881
- Location
- Surrey
Nice armchair.
DDR4 bandwidth is higher. 2400c10 for example is slower than DDR4 2400c13. The leap however comes from both the frequencies achievable and the density thanks to Intels IMC
You won't see any DDR3 platform Intel or otherwise capable of running 64gb of memory at 2666T1.
DDR4 bandwidth is higher. 2400c10 for example is slower than DDR4 2400c13. The leap however comes from both the frequencies achievable and the density thanks to Intels IMC
You won't see any DDR3 platform Intel or otherwise capable of running 64gb of memory at 2666T1.
Last edited:
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
Intel doesn't support more than 8gb per dimm on mainstream platforms, AMD supports 16GB
Bandwidth is barely higher at all, we're talking marginal and not really noticeable in anything but a memory benchmark. Where as DDR2 offered massively more bandwidth compared to DDR1... the difference in DDR3 and 4 is marginal. But DDR3 2400Mhz is cheaper than DDR4 2133Mhz.
http://www.corsair.com/en-us/blog/2014/september/ddr3_vs_ddr4_synthetic
Ivy-e dual bank 59,831MB/s, Haswell-e dual bank 60,231MB/s...
That is less than 1% more bandwidth, doesn't account for improvements in the memory controller and potentially higher efficiency. It's within the margin of benchmarking error and certainly won't make noticeable performance differences at all.
Keep in mind you need 4 channels of dual bank sticks to achieve 60GB/s....
In 5 years at DDR4's peak we're probably talking 80GB/s in 4 sticks giving say 32GB of memory, a more like 16gb with single bank 4GB sticks giving 75GB/s.
A single stack of 8GB HBM will be achieving 256GB/s in under 2 years from now. 2 stacks would achieve 512GB/s, you could have 1TB/s bandwidth with 32GB using HBM.
1TB/s vs 80GB/s, one is a marginal generational increase, one is a freaking revolution in performance and what you can do with a cpu architecturally when you can back on 10 times the bandwidth.
More to the point, 1TB/s of HBM will use less power than 80GB/s of ddr4. WIth little to nothing else on the horizon 1-5% bandwidth improvements in system memory was actually interesting, when we're pretty much a few months away from the first HBM gpu's and probably 6-12 months from the first HBM using APU's...... current system memory is pretty much a joke and holds absolutely zero interest going forwards. It's rather like HDD's vs SSD's, random 4kb reads jumping from 0.5mb's to 25MB's, night and day in terms of responsiveness and performance. We're talking about extremely "last gen" product versus a magnitudes faster new gen product. HBM/HMC will entirely change ram and how CPU's use it. The sub 1% performance increase currently from same speed ddr3 vs ddr4(with even higher speed ddr3 available and has been available for a long time already don't forget) is not exciting, the roughly speaking 450% performance increase from ddr4 to HBM is just a tad more interesting. It's not an AMD or Intel thing, Intel has HMC with the performance benefits and a decent portion of the power benefits.
You've also got to remember that mainstream platforms have significantly lower bandwidth, 30,000MB/s and HBM will be the same there, no need for ultra expensive quad channel motherboards or high end cpu's to go with it. The same performance will be available on your £150 quad cores as your £900 hex/octo cores. Jumping from 30,000MB/s to 256,000MB/s using less power on mainstream platforms, that 30,250MB/s from DDR4 looks a touch less interesting.
Bandwidth is barely higher at all, we're talking marginal and not really noticeable in anything but a memory benchmark. Where as DDR2 offered massively more bandwidth compared to DDR1... the difference in DDR3 and 4 is marginal. But DDR3 2400Mhz is cheaper than DDR4 2133Mhz.
http://www.corsair.com/en-us/blog/2014/september/ddr3_vs_ddr4_synthetic
Ivy-e dual bank 59,831MB/s, Haswell-e dual bank 60,231MB/s...
That is less than 1% more bandwidth, doesn't account for improvements in the memory controller and potentially higher efficiency. It's within the margin of benchmarking error and certainly won't make noticeable performance differences at all.
Keep in mind you need 4 channels of dual bank sticks to achieve 60GB/s....
In 5 years at DDR4's peak we're probably talking 80GB/s in 4 sticks giving say 32GB of memory, a more like 16gb with single bank 4GB sticks giving 75GB/s.
A single stack of 8GB HBM will be achieving 256GB/s in under 2 years from now. 2 stacks would achieve 512GB/s, you could have 1TB/s bandwidth with 32GB using HBM.
1TB/s vs 80GB/s, one is a marginal generational increase, one is a freaking revolution in performance and what you can do with a cpu architecturally when you can back on 10 times the bandwidth.
More to the point, 1TB/s of HBM will use less power than 80GB/s of ddr4. WIth little to nothing else on the horizon 1-5% bandwidth improvements in system memory was actually interesting, when we're pretty much a few months away from the first HBM gpu's and probably 6-12 months from the first HBM using APU's...... current system memory is pretty much a joke and holds absolutely zero interest going forwards. It's rather like HDD's vs SSD's, random 4kb reads jumping from 0.5mb's to 25MB's, night and day in terms of responsiveness and performance. We're talking about extremely "last gen" product versus a magnitudes faster new gen product. HBM/HMC will entirely change ram and how CPU's use it. The sub 1% performance increase currently from same speed ddr3 vs ddr4(with even higher speed ddr3 available and has been available for a long time already don't forget) is not exciting, the roughly speaking 450% performance increase from ddr4 to HBM is just a tad more interesting. It's not an AMD or Intel thing, Intel has HMC with the performance benefits and a decent portion of the power benefits.
You've also got to remember that mainstream platforms have significantly lower bandwidth, 30,000MB/s and HBM will be the same there, no need for ultra expensive quad channel motherboards or high end cpu's to go with it. The same performance will be available on your £150 quad cores as your £900 hex/octo cores. Jumping from 30,000MB/s to 256,000MB/s using less power on mainstream platforms, that 30,250MB/s from DDR4 looks a touch less interesting.
Last edited:
haha yeah - Lets hope they can create a CPU which can beat a 2nd gen intel chip while everyone else uses 5th gen chips......
Soldato
- Joined
- 5 Sep 2011
- Posts
- 12,881
- Location
- Surrey
X99 supports 16gb DIMMS that's what matters 
. The bandwidth increase is marginal Yes I don't think there are any huge leaps but when you experiement with the platform and its potential application, you can't ignore that 2666 / 2800 with large memory density is pretty impressive.
In fact with T2 and being conservative I'm pretty sure you could get a large number of samples to achieve 3000 with 64gb. Although that may require some patience
. The bandwidth increase is marginal Yes I don't think there are any huge leaps but when you experiement with the platform and its potential application, you can't ignore that 2666 / 2800 with large memory density is pretty impressive.In fact with T2 and being conservative I'm pretty sure you could get a large number of samples to achieve 3000 with 64gb. Although that may require some patience
Last edited:
Intel doesn't support more than 8gb per dimm on mainstream platforms, AMD supports 16GB
Bandwidth is barely higher at all, we're talking marginal and not really noticeable in anything but a memory benchmark. Where as DDR2 offered massively more bandwidth compared to DDR1... the difference in DDR3 and 4 is marginal. But DDR3 2400Mhz is cheaper than DDR4 2133Mhz.
http://www.corsair.com/en-us/blog/2014/september/ddr3_vs_ddr4_synthetic
Ivy-e dual bank 59,831MB/s, Haswell-e dual bank 60,231MB/s...
That is less than 1% more bandwidth, doesn't account for improvements in the memory controller and potentially higher efficiency. It's within the margin of benchmarking error and certainly won't make noticeable performance differences at all.
Keep in mind you need 4 channels of dual bank sticks to achieve 60GB/s....
In 5 years at DDR4's peak we're probably talking 80GB/s in 4 sticks giving say 32GB of memory, a more like 16gb with single bank 4GB sticks giving 75GB/s.
A single stack of 8GB HBM will be achieving 256GB/s in under 2 years from now. 2 stacks would achieve 512GB/s, you could have 1TB/s bandwidth with 32GB using HBM.
1TB/s vs 80GB/s, one is a marginal generational increase, one is a freaking revolution in performance and what you can do with a cpu architecturally when you can back on 10 times the bandwidth.
More to the point, 1TB/s of HBM will use less power than 80GB/s of ddr4. WIth little to nothing else on the horizon 1-5% bandwidth improvements in system memory was actually interesting, when we're pretty much a few months away from the first HBM gpu's and probably 6-12 months from the first HBM using APU's...... current system memory is pretty much a joke and holds absolutely zero interest going forwards. It's rather like HDD's vs SSD's, random 4kb reads jumping from 0.5mb's to 25MB's, night and day in terms of responsiveness and performance. We're talking about extremely "last gen" product versus a magnitudes faster new gen product. HBM/HMC will entirely change ram and how CPU's use it. The sub 1% performance increase currently from same speed ddr3 vs ddr4(with even higher speed ddr3 available and has been available for a long time already don't forget) is not exciting, the roughly speaking 450% performance increase from ddr4 to HBM is just a tad more interesting. It's not an AMD or Intel thing, Intel has HMC with the performance benefits and a decent portion of the power benefits.
You've also got to remember that mainstream platforms have significantly lower bandwidth, 30,000MB/s and HBM will be the same there, no need for ultra expensive quad channel motherboards or high end cpu's to go with it. The same performance will be available on your £150 quad cores as your £900 hex/octo cores. Jumping from 30,000MB/s to 256,000MB/s using less power on mainstream platforms, that 30,250MB/s from DDR4 looks a touch less interesting.
Have to question your opinion when, as the poster above said, Intel already support 16GB dimms on X99....
Either way, DDR4 > DDR3. Technological progression is nice to see, even if it isn't as much of a leap as we'd hope for.
At least Intel are releasing new CPU's, have fun waiting until 2016 for AMD's.
The technicality is the 'mainstream platforms', x99 being enthusiast.
Progress is always good to see. Personally cannot wait to see some APU's with HBM. Solve bandwidth issue, whack on a 7870 perf level gpu, 4 decent cores.
Happily sidegrade. Whack it in an SFF case, maybe with some watercooling love too so I can clock the transistors off of it.
Progress is always good to see. Personally cannot wait to see some APU's with HBM. Solve bandwidth issue, whack on a 7870 perf level gpu, 4 decent cores.
Happily sidegrade. Whack it in an SFF case, maybe with some watercooling love too so I can clock the transistors off of it.
DDR4 is total crap
Here is the best physics score for Firestrike done by Kingpin using LN2 and a massively overclocked 5960X @5775mhz and DDR4
Physics Score 27605
http://www.3dmark.com/fs/3262926
To put that in prospective, if we take a humble Haswell 4790k using DDR3 and half the cores it is possible for people using just water and a much lower oc to get half of Kingpins score.
Also the physics tests in the futuremark benches scale very well indeed with regard to the number of cores used.
Here is the best physics score for Firestrike done by Kingpin using LN2 and a massively overclocked 5960X @5775mhz and DDR4
Physics Score 27605
http://www.3dmark.com/fs/3262926
To put that in prospective, if we take a humble Haswell 4790k using DDR3 and half the cores it is possible for people using just water and a much lower oc to get half of Kingpins score.
Also the physics tests in the futuremark benches scale very well indeed with regard to the number of cores used.
If Haswell-E and DDR4 are 'total crap' then I ponder the wording with which you'd describe the FX line of cpu's