Together with the "evolution" of DDR memory chips and modules (increasing chip clock frequencies or decreasing latencies), the most important event of the expiring year is that JEDEC accepted a new DDR2 standard (JESD79-2A) and there appeared memory chips and modules based on this standard. DDR2-SDRAM memory is an evolutionary replacement of the current generation of memory – DDR. It's based on similar functional principles – data transfer along the 64-bit bus on both clock edges, which insures double efficient data transfer rate relative to the module frequency.
It goes without saying that DDR2 wouldn't have been a principally new standard,
if it hadn't had a number of innovations, which allowed a leap to
much higher frequencies (larger throughput) and larger chip array
capacities, as well as reduced power consumption of modules. The main
features of the new memory standard are reviewed in detail in our
article, so let's just dwell
on the most important of these features – modification of the
data prefetch scheme. As is known, the DDR standard uses
the 2n-prefetch scheme. It means fetching
2 bits at a cycle via the internal data bus (by using multiplexer/demultiplexer
of the 2-1 type), which allows memory chips to operate at the frequency
twice as low relative to the effective frequency of the external data
bus. So, DDR-400 module chips operate at 200 MHz. In the DDR2
standard this scheme is replaced by 4n-prefetch
(i.e. a more complex multiplexer of the 4-1 type is introduced), which
has both advantages and disadvantages. The advantages are obvious
– fetching 4 bits per cycle allows to reach memory bandwidth,
equivalent to the memory bandwidth of DDR of the same frequency at
twice as low memory chip frequency (DDR2-400 module chips operate
at 100 MHz), which in its turn contributes to the reduction of
power consumption in memory module chips. But the disadvantages are
also obvious – complication of the data transformation scheme
is always accompanied by increasing latencies, we shall cover this
topic later.
DDR-400 memory, 2n-prefetch
DDR2-400 memory, 4n-prefetch
Standard is standard, but it wouldn't be that interesting in itself, but for the largest manufacturers of memory modules and chips who started to produce chips and modules of the new standard. That's why we should dwell on the modules. So, DDR2 memory modules are manufactured in a new form factor, as 240-pin DIMM modules, electrically incompatible with DDR slots (by the number of pins, pin distance, and module pinout). Thus, the DDR2 standard does not provide back compatibility with DDR. DDR2 chips are manufactured in a compact FBGA package (at least, it's recommended in the standard), which allows larger chip capacities at smaller dimensions and better electric and thermal characteristics. Voltage of chips/modules has also changed – it was decreased from 2.5 to 1.8V (due to the new process technology – upgrade to 100-nm and, later, 90-nm process), which also contributes to the reduction of power consumption of memory modules – this feature is important for notebooks as well as for large workstations and servers.
The first DDR2 memory modules to see the light were DDR2-400 and DDR2-533 modules. A number of largest manufacturers (Kingston, Corsair, Samsung) started to produce them a couple of months before the official announcement (to say nothing of the real release) of the first Intel 915/925 series chipsets supporting memory of this type. Other manufacturers acted somewhat differently – they timed the announcement of their DDR2 modules to the announcement of the above mentioned Intel chipsets (this was done by GeIL, KINGMAX, OCZ and... the same Kingston with DDR2-533 modules and Corsair with XMS2 series modules – DDR2-667, which mark cracking the 533 MHz barrier). By the end of the year, 667 MHz DDR2 modules also appeared in products from Kingston (marked 675 MHz) and Samsung (with a funny mark – 711 MHz :) ). Moreover, there take shape the first steps towards 400 MHz external data bus (DDR2-800) – successful experiments to overclock DDR2-667 OCZ and Corsair modules to begin with, and then – readiness to manufacture "sterling" DDR2-800 chips, this time from Elpida.
Component speed |
Module speed |
Bus frequency |
Data transfer rate |
Module throughput |
single-channel mode |
dual-channel mode |
DDR2-400 |
PC2-3200
|
200 MHz
|
400 MT/s
|
3.2 GB/s
|
6.4 GB/s
|
DDR2-533 |
PC2-4300
|
266 MHz
|
533 MT/s
|
4.3 GB/s
|
8.6 GB/s
|
DDR2-667 |
PC2-5300
|
333 MHz
|
667 MT/s
|
5.3 GB/s
|
10.6 GB/s
|
DDR2-800 |
PC2-6400
|
400 MHz
|
800 MT/s
|
6.4 GB/s
|
12.8 GB/s
|
Thus, manufacturers of memory chips and modules do not have serious
problems with higher frequencies in memory modules and consequently
with higher throughput. Frankly speaking, this is not surprising –
taking into account that DDR2 chips operate at the frequency four
times as lower than the effective frequency of the external data bus.
How justified this frequency increase is –
that's another story. Let's see. Our first tests
of DDR2-533 modules in dual-channel mode with the first Intel chipsets
915P/G and 925X demonstrated... that there was completely no point
in increasing the memory bus frequency to 266 MHz (533 MHz
DDR). Why? Because the peak throughput of a CPU bus operating at 200 MHz
in Quad-Pumped Bus mode is just 6.4 GB/s. While the peak throughput
of DDR2-533 in dual channel mode is 8.6 GB/s. It's obvious that
the throughput of any system is always limited by the throughput of
the slowest component – that's axiomatic. In this case it's
the FSB throughput. So the real throughput of DDR2-533 in dual channel
mode is no better than that... of a regular DDR-400! And what about
memory latency? It's not difficult to guess that it's higher in DDR2-533
than in DDR-400. Firstly, due to memory timings – typical timings
for DDR2-533 are 4-4-4. DDR-400 timings are usually no more than 2.5-3-3.
Secondly, due to its asynchronous operating mode. What conclusion
can we draw? At present, for the current generation
of chipsets with 200 MHz FSB (the majority) there is just no
need in DDR2 memory. In general, DDR2-400 is no better, and
even worse than DDR-400. DDR2-533 can do better, but we managed to
reveal its full potential with the first 915/925 series chipsets only
in single channel mode,
which is hardly justified (you will hardly buy a system with single
channel DDR2-533 as an alternative of dual channel DDR-400 system).
To say nothing of DDR2-667 (800).
DDR2 positions could have remained poor but for the announcement
of the new Intel 925XE chipset, which was presented by the company
last year in November – the first chipset supporting 266 MHz
FSB (in fact it's just a reworked 925X – there is nothing new
in the 925XE, except for the 266 MHz bus). It was accompanied
by a slightly upgraded version of Pentium 4 Extreme Edition (Gallatin
core aka Northwood with L3 cache) at 3.46 GHz designed right
for the 266 MHz FSB. Nevertheless, the first tests
of this combo of "two XE" (P4XE + i925XE) were not comforting
– memory bandwidth gain turned out far from that considerable.
A good thing – the latency dropped due to the memory system
operating in synchronous mode with the FSB. This time the culprit
was... the processor, its core, to be more exact. That's because FSB
frequency alone is not enough to squeeze maximum from the memory system.
What's also important is the implementation efficiency of the CPU
system, responsible for data exchange with RAM – Bus Interface
Unit (BIU), as well as of the hardware/software prefetch algorithms.
Which are noticeably worse in Northwood/Gallatin than in Prescott.
But the trouble is there are no Pentium 4 processors with this core
on the 266 MHz bus so far, and none is expected. There is only
one solution and we certainly found it – to overclock the existing
Prescott with the 200 MHz bus. As a result we got sort of a future
processor – Pentium
4 3.73 GHz. Tests
of DDR2-533 memory with this processor at last allowed to reveal its
potential completely – first of all the intended
memory bandwidth. But: only DDR2-533 (DDR2-667 is
out of the question – it will require a 333 MHz (!) FSB
to reveal its potential) and only with overclocked
processor. Thus, top DDR2 models – 667 MHz and 800 MHz
(to be released) models – still have dubious prospects, to be
more exact – no prospects at all. Of course, they can also be
used in single channel mode, but who will do it when we have DDR-400,
which operates perfectly in dual channel mode and has much better
characteristics compared to "single-channel" DDR2-667/800?
Memory type, operating mode |
Frequency
FSB/DRAM,
MHz |
Maximum
real memory read
bandwidth, MB/s |
Minimum
latency1,
ns |
DDR-400, synchronous, single-channel2 |
200/200
|
3290
|
46.6
|
DDR-400, synchronous, dual-channel1 |
200/200
|
6278
|
47.7
|
DDR2-533, asynchronous, single-channel2 |
200/266
|
4287
|
51.9
|
DDR2-533, asynchronous, dual-channel1 |
200/266
|
6327
|
55.0
|
DDR2-533, asynchronous, dual-channel3 |
200/266
|
6412
|
47.5
|
DDR2-533, synchronous, dual-channel3 |
266/266
|
8247
|
40.8
|
1Read about the measurement procedure in the article "DDR2
– future replacement of DDR. Theoretic background and first
results of low level tests"
2Testbed configuration is given in the article "DDR2
vs DDR. Test results in single-channel mode"
3Testbed configuration is given in the article "Discovering
DDR2-533 potential. Part 2: 266 MHz FSB, Intel Pentium 4 Prescott
processor"
That's actually all we wanted to say about the present state of DDR2 (the
summary table with test results is published above). In conclusion,
we shall not skirt the DDR industry, which is still developing and
prosperous. Certainly, we cannot speak of standards (like in DDR2)
here – standards for overclockers' memory were not accepted
this year and they will hardly be accepted by JEDEC at all. But how
wide is the range of these products! There are such modules as DDR-433,
-466 (PC3700), -500 (PC4000), -533 (PC4300, sometimes PC4200), and
even -550 (PC4400). This year, among the largest manufacturers of
overclockers' modules we can note OCZ Technology, Kingston, Corsair,
and GeIL. We are also glad to mention Russian products under the DIGMA
brand. KINGMAX was very original by announcing a series of colored
memory modules. One can note two DDR tendencies – on the one
hand, it's the increase of memory chip frequencies (that is the reduction
of access times); on the other hand it's the reduction of latencies
(new models with lower timings, which are often designed for really
extreme voltage – like 3.2 V!). It's not difficult to guess
that these two tendencies are closely related. At present high frequency
DDR modules are interesting only to "overclockers at any cost"
– first of all by raising FSB frequency far beyond the values
noted in CPU/chipset specifications. The release of chipsets/motherboards
with standard FSB frequency higher than 200 MHz
only for the sake of non-standard overclockers' memory
seems highly improbable. Although some manufacturers may decide to
launch motherboards based on Intel 925XE supporting both DDR2 and
DDR memory types (similar to "hybrid-like" motherboards
on 915 series chipsets). Only in this case DDR-533 modules may become
a good alternative to DDR2-533 modules.
Write a comment below. No registration needed!
|
|
|
|
|