G.Skill Shows Off Two New DDR4 RAM Kits Clocked At Over 4000 MHz

[IInuyasha74]

G.Skill demos two DDR4 RAM kits clocked at over 4000 MHz at Intel Developer Forum 2015.

G.Skill Shows Off Two New DDR4 RAM Kits Clocked At Over 4000 MHz : Read more

 

[Rookie_MIB]

Unfortunately with those CAS timings they're really no better than 2133 CAS 9 DDR3..

4266/CAS19 = 217
2133/CAS9 = 237 which means more operations/second

 

[IInuyasha74]

Unfortunately with those CAS timings they're really no better than 2133 CAS 9 DDR3..

4266/CAS19 = 217
2133/CAS9 = 237 which means more operations/second

It is still improvement though. The bandwidth is much greater 68,256 MB/s vs 34,128 MB/s. The access time is a little slower than than a DDR3 2133/CAS9 kit, 8.9ns for the 4266 MHz kit vs 8.4ns for the DDR3, but you are transferring twice the amount of data each time. The end result is that the DDR3 kit mentioned here would achieve 18 cycles in the time it takes the DDR4 kit here to complete 17 cycles. The DDR4 is still better, however, because in 17 cycles it will provide 1,160,352 MB/s of bandwidth, while the DDR3 only provides 614,304.

 

[darkokills]

I'm sitting pretty with 2400Mhz DDR3 and will be for a while. Don't need anything more than my 4790k for now.

 

[Rookie_MIB]

Unfortunately with those CAS timings they're really no better than 2133 CAS 9 DDR3..

4266/CAS19 = 217
2133/CAS9 = 237 which means more operations/second

It is still improvement though. The bandwidth is much greater 68,256 MB/s vs 34,128 MB/s. The access time is a little slower than than a DDR3 2133/CAS9 kit, 8.9ns for the 4266 MHz kit vs 8.4ns for the DDR3, but you are transferring twice the amount of data each time. The end result is that the DDR3 kit mentioned here would achieve 18 cycles in the time it takes the DDR4 kit here to complete 17 cycles. The DDR4 is still better, however, because in 17 cycles it will provide 1,160,352 MB/s of bandwidth, while the DDR3 only provides 614,304.

Well, that's only true if the memory controller is a quad channel controller as in Ivy Bridge-E or Haswell-E. For those moving to Skylake, that only has a dual channel memory controller and as such will only be operating at the same relative speeds as mainstream Haswell products. And because this particular kit is ONLY a 2x4GB kit, it cannot achieve those speeds that you mention.

 

[IInuyasha74]

Unfortunately with those CAS timings they're really no better than 2133 CAS 9 DDR3..

4266/CAS19 = 217
2133/CAS9 = 237 which means more operations/second

It is still improvement though. The bandwidth is much greater 68,256 MB/s vs 34,128 MB/s. The access time is a little slower than than a DDR3 2133/CAS9 kit, 8.9ns for the 4266 MHz kit vs 8.4ns for the DDR3, but you are transferring twice the amount of data each time. The end result is that the DDR3 kit mentioned here would achieve 18 cycles in the time it takes the DDR4 kit here to complete 17 cycles. The DDR4 is still better, however, because in 17 cycles it will provide 1,160,352 MB/s of bandwidth, while the DDR3 only provides 614,304.

Well, that's only true if the memory controller is a quad channel controller as in Ivy Bridge-E or Haswell-E. For those moving to Skylake, that only has a dual channel memory controller and as such will only be operating at the same relative speeds as mainstream Haswell products. And because this particular kit is ONLY a 2x4GB kit, it cannot achieve those speeds that you mention.

No. First, Ivy Bridge-E can only use DDR3 memory, not DDR4. Second, the calculations I did were assuming the use of a dual-channel memory controller.

Bandwidth = Memory clock speed x data width / 8

so

68,256 = 4266 x 128 / 8 Or more simply 68,256 = 4266 x 16

If it was using a quad-channel memory interface like Haswell-E then it would again double the total bandwidth again. In that case you are right you would need a second kit, or a 4 x 4 GB kit at least.

 

[Joe Black]

It's likely got something to do with the processor (as in it supports it), but mostly with the manufacturing tech that leaks less current and generates less heat. Let's be realistic here.

 

[Jerahmia Wood]

Im not really impressed, its well proven (especially by maximum PC) that memory speed has diminishing returns and little performance increase even with modern software with the acception of using on-die GPUs and APUs (witch for the forseeable future will not rival the performance of a discreet lower-mid to high end GPU) memory speed matters little above DDR3-1600 speeds, id rather spend the money on higher capacity cheaper kits, and put that money saved toward a faster GPU, CPU or SSD where it really matters.

 

[gilbadon]

I love how CPU's changed to GHz as soon as they hit 1, but memory is like "screw it, our number looks bigger anyway" even though they are breaking 4.

 

[mapesdhs]

I'd much rather they just yank the prices down on the higher capacity, upper end DDR3 kits. There's no reason for a 32GB/2400 kit to continue to be so expensive.
Meanwhile, the 8GB and 16GB kits should drop too. Only this week have I finally seen an 8GB/2133 kit for the same low price that they reached several years ago before everything went through the roof (blamed of course of floods & supplier factory-damage, but that was never a satisfactory explanation for all of the price rises; much of it was just price gouging because they could IMO).

 

[mapesdhs]

Jerahmia, that's too much of a generalisation. Some apps very much do benefit from faster RAM speeds, eg. After Effects. Friend of mine tested his 3930K/4.7 system with 2133 vs. 2400, render times increased several percent with the lower speed.

For consumer apps though, yes indeed, a lot of programs don't gain much. Games vary, but generally also not affected much. However, supply/demand issues often mean that the pricing sweet spot for DDR3 is around the 2133 speed anyway, ie. buying a slower kit is often more expensive. That was certainly the case when I bought my 16GB 2400 kit, it was cheaper than the 2133 kit.

 

[IInuyasha74]

Im not really impressed, its well proven (especially by maximum PC) that memory speed has diminishing returns and little performance increase even with modern software with the acception of using on-die GPUs and APUs (witch for the forseeable future will not rival the performance of a discreet lower-mid to high end GPU) memory speed matters little above DDR3-1600 speeds, id rather spend the money on higher capacity cheaper kits, and put that money saved toward a faster GPU, CPU or SSD where it really matters.

It seems you are looking at it from a "Don't fix what isn't broke" mentality. Testing under Ivy Bridge and Haswell has shown that some applications do experience significant performance improvements from faster RAM. Some only get a small increase like 1 percent or so, and some don't get any measurable improvement. However, technology is continuing to evolve. Wouldn't you rather the industry move to a faster memory standard before it becomes a serious issue rather than after?

 

[velocityg4]

I love how CPU's changed to GHz as soon as they hit 1, but memory is like "screw it, our number looks bigger anyway" even though they are breaking 4.

A competitor should come out and say their memory works at over 4,000,000,000 Hz.

 

[InvalidError]

Unfortunately with those CAS timings they're really no better than 2133 CAS 9 DDR3..

4266/CAS19 = 217
2133/CAS9 = 237 which means more operations/second

This is only the first word latency. A complete memory operation also needs to actually transfer data. On top of that, all SDRAM since the original PC60 supports pipelining: setting up the next read/write (from/to a different DRAM rank or bank) while data is being transferred. With pipelined access, 4266MT/s RAM offers roughly twice the potential transactions per second.

Memory transfers are usually 64-256 bytes at a time (cache line fills, 2-8 cycles) on top of that, so you have (19+3)/2133 = 10.3ns completion time for 4266MT/s vs (9+3)/1066 = 11.3ns for 2133MT/s.

This means the 4266MT/s memory is able to complete 9% more memory operations per second even without pipelining.

 

[InvalidError]

I love how CPU's changed to GHz as soon as they hit 1, but memory is like "screw it, our number looks bigger anyway" even though they are breaking 4.

The only thing that runs at "4GHz" (should be 4Gbps or GT/s) is the IO drivers and the multiplexers between the row register array and IO pads. The actual DRAM array with its analog bits (sense amplifiers) which comprise ~95% of the die area runs at something closer to 150MHz.

If you look at PCIe, GDDR5, DisplayPort, HDMI and other interfaces, 4Gbps per pin is barely par for the course.

 

[Rookie_MIB]

Memory transfers are usually 64-256 bytes at a time (cache line fills, 2-8 cycles) on top of that, so you have (19+3)/2133 = 10.3ns completion time for 4266MT/s vs (9+3)/1066 = 11.3ns for 2133MT/s.

This means the 4266MT/s memory is able to complete 9% more memory operations per second even without pipelining.

Yeah - after I started digging more into it, I started seeing some of the optimizations that can be made into reducing latency. From a pure numbers point of view, the raw math is still a 'best case' scenario, actual results depend on exactly what operations are going on as there still can wind up being some overhead.

 

[IInuyasha74]

Memory transfers are usually 64-256 bytes at a time (cache line fills, 2-8 cycles) on top of that, so you have (19+3)/2133 = 10.3ns completion time for 4266MT/s vs (9+3)/1066 = 11.3ns for 2133MT/s.

This means the 4266MT/s memory is able to complete 9% more memory operations per second even without pipelining.

Yeah - after I started digging more into it, I started seeing some of the optimizations that can be made into reducing latency. From a pure numbers point of view, the raw math is still a 'best case' scenario, actual results depend on exactly what operations are going on as there still can wind up being some overhead.

That is certainly true. All of this is the best theoretical performance. The efficiency of the memory controller and various other aspects will cause this to dip from time to time.

 

[Gillerer]

The probable reason they only have 2 x 4 GB kit and not 4 x 4 GB or 2 x 8 GB one is the memory can't maintain stability at the same timings if you increase module count or size.

 

[IInuyasha74]

The probable reason they only have 2 x 4 GB kit and not 4 x 4 GB or 2 x 8 GB one is the memory can't maintain stability at the same timings if you increase module count or size.

Not quiet. RAM makers use lower capacity chips on the DIMMs because it results in less heat, less power demands, and requires less silicon per chip, which increases the yield. So it is treu they don't create 2 x 8 GB chips for this reason. However, creating a 4 x 4 GB kit wouldn't change the actually chips on board. They just need to pull 4 modules at a time instead of two and test them to ensure they work together.

They may not be doing it currently because of supply issues, or they simply don't think customers are interested in it at the moment. Could be a number of reasons, but the RAM itself isn't constraining it.