Any clue if Samsung is using TSV in all it's DDR4 (so here too) , wondering about reliability and impact on OC potential, can't say i've seen any data on that topic.
Normally within a series of products it's a fairly smooth continuum with actual latency periods roughly constant; the 4000 module had to really lengthen the latencies though to get those last few MHz out of the chips.
DDR4 is here now and price isn't bad considering you have to buy RAM regardless. Why not use the faster stuff now and switch to HBM when it makes sense? At this point in time integrating their APU designs with an interposer and HBM probably won't be cheap. You can slap it in a high-end graphics card today, but how about a sub-$200 APU?
Personally I wouldn't be surprised if they didn't release HBM-equipped APUs until Zen. I would like to be wrong, and have them at least release a few high-end HBM models for laptops and other compact form factors, where a graphics upgrade isn't a drop-in card away.
AT/Ian already did one for DDR4 on Haswell... There's even more stratospheric speeds now but I don't l doubt it matters, tho there were edge cases where the difference was as much as 10% (not synthetic either).
Probably better off with tighter timings and lower clocks. The price of these DIMMS seems to increase exponentially the higher the timings, even though the performance remains almost the same.
Actually, the time it takes to get a single transfer from RAM (i.e. the latency from the point the command and address are issued until the transferred data resides in the applicable CPU register) is roughly constant. Reducing it by any significant margin would simply be unacceptable, since a higher voltage would be required due to the long bus lines that need to be reliably driven at high speeds, which would, in turn, result in about 5~10W increased power for a 20-30% reduced latency (why 10W of extra power is so unfathomly huge is beyond me though).
However, the increased transfer rate (even with looser timings keeping the total latency constant) normally results in an overall increase in available bandwidth, due to the way the CPU caches and pipelines data and instructions. You can think of it this way, lets say you have a 1km length of single-lane road with a 10km/h speed limit. The minimum time it takes a single car to move from the starting point to the end would then be 6 minutes (aka. the latency). However, this does NOT mean that the maximum throughput is limited to 10 cars per hour (the total bandwidth). If we were to release one car every second, which is (greatly simplified) analogous to the transfer rate of the memory module in Hz, then the total throughput would be 60 cars per hour.
In reality, you need to wait for the data on the line to be stable and latched into a register (normally a flip-flop) before you can send another piece of data (you cannot have multiple bits travelling down the same piece of uninterrupted wire), but the general idea is the same. Unfortunately, the CPU needs to have a very effective caching algorithm (together with accurate branch prediction and instruction reordering) to hide the increased timing latencies, which is why we only see performance improving with certain workloads and only up to a certain point.
Those new Ripjaw spreaders look a lot classier than the old ones, the Trident ones look ever better. Hopefully they're making new kits available by the time Skylake is available...
Right now there's somewhat of a lack of 2x8GB kits at Newegg, specially if you're trying to avoid red and stick to modestly priced 2400 stuff... There's the Ballistix, Corsair Vengeance in black, and I think a 4x8GB Corsair in blue.
I might actually end up buying that last kit and selling two of the 8GB DIMMs for $110 or whatever just to get some blue ones to match my cooler. :p
Funny thing is, it's probably one of the cheapest spreaders to stamp/produce across all brands... Sometimes simple works. I'd take the old clipped on spreaders over much of the over designed ones today.
I'd say, they consume 2W on a combined area that exceeds that of a CPU. Unless you wrap them in insulating blankets and use a passive case without any airflow, I just don't see how you could ever overheat RAM even without and heat spreader.
Im running my ripjawz ddr3 at 2600,they get hot very hot. I had to get a ram fan just to bring down temps so i could stop getting freezes while gaming for hours... You got to have the product first and then talk about it ;)
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
25 Comments
Back to Article
jjj - Wednesday, July 29, 2015 - link
Any clue if Samsung is using TSV in all it's DDR4 (so here too) , wondering about reliability and impact on OC potential, can't say i've seen any data on that topic.DanNeely - Wednesday, July 29, 2015 - link
Looking at those timings I'm wondering if the DDR4-3800 kit won't actually be faster. They really had to loosen the timings for those last 200 MHz.ImSpartacus - Wednesday, July 29, 2015 - link
Seems like that's always the never ending question, eh? Tight timings or high clocks?DanNeely - Wednesday, July 29, 2015 - link
Normally within a series of products it's a fairly smooth continuum with actual latency periods roughly constant; the 4000 module had to really lengthen the latencies though to get those last few MHz out of the chips.littlebitstrouds - Wednesday, July 29, 2015 - link
So, are we going to see a study, like always, showing all these speed bumps really equates to 1-2% increases in performances?Pissedoffyouth - Wednesday, July 29, 2015 - link
I can't wait for the APUs to get this, will be awesomemeacupla - Thursday, July 30, 2015 - link
why? APUs are going to get HBM, which is a lot faster than DDR4Alexvrb - Thursday, July 30, 2015 - link
DDR4 is here now and price isn't bad considering you have to buy RAM regardless. Why not use the faster stuff now and switch to HBM when it makes sense? At this point in time integrating their APU designs with an interposer and HBM probably won't be cheap. You can slap it in a high-end graphics card today, but how about a sub-$200 APU?Personally I wouldn't be surprised if they didn't release HBM-equipped APUs until Zen. I would like to be wrong, and have them at least release a few high-end HBM models for laptops and other compact form factors, where a graphics upgrade isn't a drop-in card away.
yuhong - Friday, July 31, 2015 - link
Not next year's AMD APUs, I think.Impulses - Wednesday, July 29, 2015 - link
AT/Ian already did one for DDR4 on Haswell... There's even more stratospheric speeds now but I don't l doubt it matters, tho there were edge cases where the difference was as much as 10% (not synthetic either).RaistlinZ - Wednesday, July 29, 2015 - link
Probably better off with tighter timings and lower clocks. The price of these DIMMS seems to increase exponentially the higher the timings, even though the performance remains almost the same.Xenonite - Thursday, July 30, 2015 - link
Actually, the time it takes to get a single transfer from RAM (i.e. the latency from the point the command and address are issued until the transferred data resides in the applicable CPU register) is roughly constant.Reducing it by any significant margin would simply be unacceptable, since a higher voltage would be required due to the long bus lines that need to be reliably driven at high speeds, which would, in turn, result in about 5~10W increased power for a 20-30% reduced latency (why 10W of extra power is so unfathomly huge is beyond me though).
However, the increased transfer rate (even with looser timings keeping the total latency constant) normally results in an overall increase in available bandwidth, due to the way the CPU caches and pipelines data and instructions.
You can think of it this way, lets say you have a 1km length of single-lane road with a 10km/h speed limit. The minimum time it takes a single car to move from the starting point to the end would then be 6 minutes (aka. the latency). However, this does NOT mean that the maximum throughput is limited to 10 cars per hour (the total bandwidth). If we were to release one car every second, which is (greatly simplified) analogous to the transfer rate of the memory module in Hz, then the total throughput would be 60 cars per hour.
In reality, you need to wait for the data on the line to be stable and latched into a register (normally a flip-flop) before you can send another piece of data (you cannot have multiple bits travelling down the same piece of uninterrupted wire), but the general idea is the same.
Unfortunately, the CPU needs to have a very effective caching algorithm (together with accurate branch prediction and instruction reordering) to hide the increased timing latencies, which is why we only see performance improving with certain workloads and only up to a certain point.
MapRef41N93W - Wednesday, July 29, 2015 - link
DDR4-4000 already? Man I can't wait to see how fast speeds are going to be when Skylake-E releases next year.Oxford Guy - Wednesday, July 29, 2015 - link
We'll see how much the latency slows them down.nandnandnand - Wednesday, July 29, 2015 - link
16 GB per module DDR4-3400. Good to see a "mainstream" higher module capacity.Impulses - Wednesday, July 29, 2015 - link
Those new Ripjaw spreaders look a lot classier than the old ones, the Trident ones look ever better. Hopefully they're making new kits available by the time Skylake is available...Right now there's somewhat of a lack of 2x8GB kits at Newegg, specially if you're trying to avoid red and stick to modestly priced 2400 stuff... There's the Ballistix, Corsair Vengeance in black, and I think a 4x8GB Corsair in blue.
I might actually end up buying that last kit and selling two of the 8GB DIMMs for $110 or whatever just to get some blue ones to match my cooler. :p
LordanSS - Wednesday, July 29, 2015 - link
Indeed... I wonder what the timings are on that part tho. If they're around 18, that'd sweeten the pot.Oxford Guy - Wednesday, July 29, 2015 - link
19-25-25etamin - Wednesday, July 29, 2015 - link
Haven't read the article yet, but the Trident Z is the first DIMM I've seen whos design alone makes me want it hahaImpulses - Wednesday, July 29, 2015 - link
Funny thing is, it's probably one of the cheapest spreaders to stamp/produce across all brands... Sometimes simple works. I'd take the old clipped on spreaders over much of the over designed ones today.Oxford Guy - Thursday, July 30, 2015 - link
Fins and ridges provide more surface area for heat dissipation. These rather smooth heatsinks look less effective.Gigaplex - Thursday, July 30, 2015 - link
None of the RAM heat spreaders are particularly effective, they just look the part. They don't need to be effective either, they don't get that hot.ShieTar - Thursday, July 30, 2015 - link
I'd say, they consume 2W on a combined area that exceeds that of a CPU. Unless you wrap them in insulating blankets and use a passive case without any airflow, I just don't see how you could ever overheat RAM even without and heat spreader.mrhoward888 - Sunday, August 2, 2015 - link
Im running my ripjawz ddr3 at 2600,they get hot very hot. I had to get a ram fan just to bring down temps so i could stop getting freezes while gaming for hours... You got to have the product first and then talk about it ;)Impulses - Thursday, July 30, 2015 - link
At this point, heat spreaders are probably more about style and keeping people from peeping at the ICs than about actually spreading heat... :p