DDR-SDRAM Has Finally Arrived — THG.RU

Introduction

The first time we heard about the future successor of the popular SDRAM memory type ‘DDR’-SDRAM is several years ago now. Back then it simply sounded like the logical consequence to prolong the life of synchronous memory and the anticipation was as high as the expectation to soon welcome this speed-improved new memory type in the market. It took surprisingly long until now finally the first DDR SDRAM-equipped systems started to become available. The main reason for the long delay was the once so strong semi-monopolist Intel, which had announced about two years ago, that the future performance memory type would not be DDR-SDRAM memory, but Rambus RDRAM. At this time it almost looked as if DDR-SDRAM would never make it in the PC as main memory, because products that weren’t supported by Intel’s processors happened to die within a short period of time.

Intel And Rambus Weren’t Able To Stop DDR-SDRAM

Things worked out differently though. The marriage with Rambus started an unfortunately long line of failures for the Santa Clara-based chipmaker. At the same time Intel’s arch rival AMD started to come out with the K7-architecture, today known as Athlon and Duron, which happened to be rightfully received very positively. AMD gained market share while Intel lost it. The Taiwanese chipset maker VIA benefited from Intel’s Rambus-policy as well and finally all the support was there that was needed for the successful launch of the DDR-SDRAM technology. Intel had to learn that products could be successful even without the blessing from Santa Clara. Today, Intel has finally announced the break-up with Rambus and we know that there will be Intel-chipsets with DDR-SDRAM support soon too.

DDR Standard In High-End 3D For A While

Of course, DDR-SDRAM is really not that new at all. It was pretty much introduced in the mainstream-PC arena by NVIDIA almost a year ago. NVIDIA’s GeForce 3D-chip was one of the first PC-components that worked with DDR-SDRAM and since this time there is hardly any high-end 3D-card that could do without this memory type anymore. However, it took a while longer until all the specifications for DDR-SDRAM as main memory were finalized.

The Memory Industry Gives Broad Support

Due to its free and open specification, DDR-SDRAM enjoys a broad support in the memory industry. Right now we know of eight different memory makers that are already producing DDR-SDRAM DIMMs in all kinds of sizes.

DDR-SDRAM gets manufactured in a very similar way as normal SDRAM, which is why there won’t be much of a price premium for the new memory type.

DDR-SDRAM Specs

Although the technical realization may be a lot more difficult, the basic principle of DDR-SDRAM is very simple. While the new memory module is clocked at the same speed as normal SDRAM, it is able to transport double the amount of data by using the rising as well as falling edge of the clock signal for data transfers. We are all aware of this technology since AGP2x and JEDEC is already working on the DDR II spec, which will double the data transfer once more, using the quad-pumped technology known from AGP4x or the upcoming Pentium 4 bus. DDR-SDRAM has another important improvement over PC133 SDRAM. Its voltage supply is using only 2.5 V, instead of 3.3 V. This and the lower capacities inside the memory chips lead to a significantly reduced power consumption, which makes DDR-SDRAM also very attractive for notebooks.

Unfortunately DDR-SDRAM DIMMs are not compatible with the SDRAM DIMMs we are using now. The new DDR-DIMMs come with 184 instead of the 168 pins used by SDRAM-DIMMs. The module itself looks almost identical to the older SDRAM, but it has only got one notch instead of the two notches found in SDRAM-DIMMs.

The official naming is also a bit confusing. Originally the naming was PC200 for DDR-SDRAM that operates at 100 MHz memory bus and PC266 for the 133 MHz bus. After Rambus however used PC600, PC700 and PC800 for their RDRAM modules, which sounds a lot faster than PC200 or PC266 although it isn’t, the memory industry came up with ‘PC1600’ and ‘PC2100’ instead. While PC200 and PC266 are only using the effective clock of the data transfer for their numbering, PC1600 and PC2100 use the actual peak data transfer rate in MB/s. Thus PC200 is the same as PC1600 (64 bit * 2 * 100 MHz = 1600 MB/s) and PC266 is equal to PC 2100 (64 bit * 2 * 133 MHz = 2133 MB/s).

DDR-SDRAM Specs, Continued

DDR-SDRAM comes in CAS-latencies of 2 and 2.5 ns. Obviously the CAS 2 DDR-SDRAM is the faster and thus more expensive of the two. JEDEC once defined that CAS 2/PC266 DDR-SDRAM should be called ‘DDR-266A’ and the CAS 2.5/PC266 DDR-SDRAM ‘DDR-266B’. Here’s the part out of the official JEDEC-spec:

DDR-Chipsets

What would this new memory type be if there weren’t any chipsets to support it? Today AMD released its AMD760 chipset, which main new feature is the DDR-SDRAM support. VIA is also close to its release of the Apollo Pro266 DDR-chipset for Pentium III and Celeron. We had the chance to look at motherboards with both chipsets and were able to put them through a series of tests.

Let’s have a quick look at a comparison table with the two and Intel’s i815:

AMD’s 760 Chipset

Today AMD announced the single-processor version of its AMD760 chipset. The SMP version is expected early next year.

It is pretty well known that AMD isn’t really interested in the chipset business, but there wasn’t much of a choice for the Sunnyvale-based chipmaker if Athlon and Duron were supposed to get a DDR-platform anytime soon. Although VIA is supposed to have a strong partnership with AMD, the Taiwanese chipset manufacturer is known to prefer releasing new technologies for their Intel-platforms first, simply because there are still more customers with Celeron and Pentium processors out there than Athlon or Duron owners.

Except for the old-fashioned PCI bus connection between the AMD761 Northbridge and the AMD766 Southbridge, AMD put a lot of the latest technologies into AMD760, so it’s not at all as obsolete as AMD750 was at its release date last year.

One of the most important new features of AMD760 is the support of 133 MHz front side bus. This increases the data transfer bandwidth of Athlon’s bus from 1.6 GB/s to 2.1 GB/s, which makes perfect sense, because PC266/PC2100 DDR-SDRAM can supply data at 2.1 GB/s as well. The only other noteworthy new feature of AMD760 is the ATA100-support. All the other specs are listed in the table above.

A New FSB Clock Requires A New Athlon Processor

A faster bus clock requires smaller clock multipliers to reach the same processor clock. While the recently released Athlon 1200 ran at 12x 100 MHz, an Athlon 1200 for the new AMD760 comes with a multiplier of only 9x. On the other hand, the new front side bus could prove extremely helpful for the overclocking of Duron processors.

You can tell an Athlon for 133 MHz FSB from the letter ‘C’. The ‘old’ Athlon for 100 MHz FSB comes with a ‘B’. Besides that there isn’t much of a difference between the 100 MHz-Athlon and the 133 MHz Athlon processor. Let’s have a look at the power requirements:

A New FSB Clock Requires A New Athlon Processor, Continued

Surprisingly, even the future Athlon 1400 is not supposed to require more than 1.75 V. This shows how well designed the Athlon-core is. It doesn’t take any ‘overclocking’ tricks to get the clock speed up.

Here’s the thermal power spec:

Athlon 1400 will radiate a whopping 76 W of heat! I am sure that the heat sink manufacturers will love to hear that.

VIA VT8633 — Apollo Pro266

The first DDR-chipset for Intel processors will obviously not come from Intel, but most likely from VIA. The upcoming Apollo Pro266 chipset is sounding as if it’s just a new version of the Apollo Pro133 chipset, beefed up with DDR-support. Surprisingly however, VIA put a whole lot of improvements into VT8633.

The most noticeable new feature besides the obvious DDR SDRAM-support is VIA’s new ‘Vlink’. Following Intel’s funky ‘hub architecture’, VIA is now also using a dedicated connection rather than the PCI-bus to communicate between north and south bridge. It is 32-bit wide, clocked at 66 MHz, its peak data transfer rate is 266 MB/s and it’s called ‘Vlink’. I do admit that I have yet to see a device that takes real advantage of this ‘hub architecture or ‘Vlink’, but I am sure that at least the marketing people will love it. Besides that the VT8233 south bridge comes with ATA100-support and an integrated networking MAC controller with standard MII interface to an external PHY, which basically is a cost effective network adapter..

The Test Platforms

AMD’s Corona Reference Platform

AMD supplied us, as many others, with a test system sporting AMD’s own ‘Corona’ 760-refernce board and a 1200/133 MHz Athlon processor. Unfortunately the setting for 100 MHz FSB would not work on this board and in the shortness of the time we were also not able to find a way of changing the CPU’s clock multiplier. This left us with only one test setting, the predefined 1200/133 MHz setting. The DDR-memory that came with the board was CL 2.5 DDR SDRAM from Micron, which we exchanged with Micron CL2 DDR SDRAM. This had a major performance impact. We also tested several other DDR-SDRAM DIMMs in this platform, but Micron’s memory turned out to be the fastest. The ‘Corona’-board worked reliably even without a lime and you will see that the performance was rather impressive in most of our tests.

Gigabyte’s GA-7DX AMD760 Platform

As an alternative we had also received an early version of Gigabyte’s GA-7DX motherboard. Unlike AMD’s Corona it comes with VIA’s 686A south bridge, which will most likely be found on most of the commercial AMD760 platforms. The GA-7DX comes with only 2 DIMM sockets, but this will probably still satisfy most users. With a secret BIOS-feature and some dipswitches we were able to adjust the FSB on this board, but unfortunately it was a too early version to run reliably at 133 MHz FSB. Thus we tested an Athlon 1000/100 in this board, as AMD was unwilling or unable to supply our German lab with an Athlon 1200/100.

VIA’s VT8633 Reference Platform

VIA’s Reference motherboard for the Apollo Pro266 chipset was used as our first DDR-platform for Intel processors. It was tested with a Pentium III 933 as this time Intel was unable or unwilling to supply us with a 1 GHz FCPGA Pentium III CPU. We don’t know if the new driver or the early stage of this motherboard were the reason that 3D-applications crashed consistently, but we could at least run several benchmarks that didn’t require the AGP.

Expectations

Due to unfortunate constraints that were partly due to unsatisfying support by AMD we weren’t able to run a full-blown scientific test suite, as I would have liked. Still we should see a clear advantage of the DDR-platforms in applications that need a lot of memory bandwidth, such as e.g. 3D-games. However, we should also not forget the days of the BX-introduction, when the 50%-increase in memory performance (from 66 to 100 MHz FSB) hardly showed in any benchmark.

Test Setup

Benchmark Results

MEMTIME

I ran this low-level memory benchmark from Intel just to show the theoretical numbers. Many memory performance numbers don’t translate into real world applications, or only when seen in context with others. I picked four different memory operations and displayed for each platform the numbers with ‘cold cache’, source and target aligned, source unaligned, target aligned and source aligned, target unaligned.

Here is a typical example showing the two-sided blade of RDRAM. If the source and the target are aligned (green bars), RDRAM is in this byte-move operation even faster than DDR-SDRAM. However, once the target is unaligned (red bars), RDRAM drops down to the worst performance and DDR SDRAM on AMD760 leads. All in all this test shouldn’t have a lot of importance, since byte moves should by now be pretty much out of fashion in the software developer community.

Moving double-words (32 bit of data) is a lot more common. Here VIA’s DDR-chipset is very close behind the dual-channel RDRAM of i840. However, VIA’s early VT8633-motherboard doesn’t look very good once the source is unaligned. AMD’s 760 doesn’t make a good figure here at all, neither does VIA’s Apollo KT133 plus SDRAM.

MEMTIME, Continued

The fill-double-words operation ‘rep stosd’ shows AMD’s760 plus DDR in a completely different light. Rambus can’t keep up with it, but neither can VIA’s VT8633.

This test shows the AMD760 plus DDR-SDRAM as the clear performance leader, while VIA’s DDR-chipset looks really bad.

It’s difficult to summarize those results. For what it’s worth one could say that DDR-SDRAM seems pretty close behind RDRAM, sometimes even ahead of the dual-channel RDRAM platform i840. What you cannot tell from those results is the real-world performance. You will soon see that VIA’s early VT8633 didn’t fare too well in the application level benchmarks. Still it gave a pretty good impression in this low-level test. The same is valid for i820 and i840. Both disappointed in each but a few of the application tests.

Sysmark 2000

Good old Sysmark2000 had to show how much gain DDR-SDRAM is bringing users of office applications.

Those results are rather interesting and I asked for several re-runs until I could completely trust them. The AMD760-platform with Athlon 1200/133 MHz is performing extremely well, scoring almost 11% better than an Athlon 1200/100 on a Asus A7V motherboard. However, the Athlon 1000/100 in Gigabyte’s GA-7DX couldn’t take much advantage of its PC266/PC2100 DDR SDRAM. Pentium III 933 on VIA’s VT8633 reference platform was also able to gain a few points due to the DDR SDRAM.

Quake 3 Arena

Quake 3 Arena is definitely able to benefit from DDR SDRAM quite nicely. 177.6 fps is the highest score I have ever seen at 640x480x16. The Athlon 1000/100 in Gigabyte’s GA-7DX is looking good too and so is the P3 933 in VIA’s VT8633 reference platform.

Unreal Tournament

The results in UT are similar to what we saw under Quake. 3D-games are able to benefit from DDR SDRAM very nicely. This stands in strong opposition to RDRAM on i820/i840. Neither of the two chipsets is able to beat i815 with its PC133 SDRAM, although RDRAM scored so well in the low-level memory tests.

Flask / DivX — MPEG4 Encoding

I have added this new benchmark mainly out of own curiosity. The ones of you who are familiar with MPEG4 or who have read our articles about it will remember that there is hardly any other software right now that has such high requirements of the system performance. If you are using Word and Excel a lot, you will still not be able to notice a major system performance increase, because most of the time you are not waiting for your system, your system is waiting for you. I also doubt that a Quake 3 player would notice a difference between 150 and 180 fps. The story is very different for people who do a lot of MPEG4 encoding. Here each performance change is noticeable. Thus I consider this benchmark as one of the most meaningful. The procedure is very simple. I encode a short VOB-file with the latest Flask and DivX. To keep it simple I omit the audio encoding completely. Flask reports the time. Dividing the number of frames by the time it took to render the short film brings me the rendering frame rate I am looking for.

Again AMD760 with DDR is having quite a lead ahead of the KT133-system. The most surprising result of this benchmark however is the bad performance of the RDRAM-systems. Rambus and Intel were always preaching that RDRAM is the perfect ‘streaming-memory’. Well, MPEG4-encoding is using a permanent memory stream and still RDRAM comes out worst. VIA’s early VT8633 plus DDR SDRAM doesn’t give a very good impression as well though. It’s slower than i815 with PC133 SDRAM.

Linux

A kernel compilation under Linux doesn’t seem to benefit much from DDR SDRAM. In case of VIA’s early VT8633-system the result is even sobering.

Viewperf

I only picked a few Viewperf results, because I know that the majority of you don’t care too much for this high-end OpenGL-test.

In each of the three benchmarks AMD760 plus DDR SDRAM scores way ahead of KT133 plus PC133 SDRAM. Working on 3D-models with a huge number of polygons requires a huge amount of memory load and stores. I didn’t include RDRAM-results here, but you can believe me that neither i820 nor i840 plus RDRAM were able to touch the performance of Athlon, AMD760 and PC266/PC2100 DDR SDRAM.

Summary

You’ve seen the results, now what do you think? I have heard from several people that they were disappointed by the small improvement seen with DDR SDRAM. They don’t consider it as a particularly amazing thing. Well, I have to heavily disagree with that.

I am doing this crazy hardware-reviewing job now for over 4 1/2 years. In this time I’ve seen a lot of new memory types coming up and each was hyped more than the one before. The first new memory I came across was EDO-RAM. Do you still remember 430VX and the first SDRAM-tests? PC66 SDRAM that was! Or what about the introduction of 440LX and the new SPD-requirement? Then there was BX and PC100 SDRAM. Finally we had PC133 SDRAM and the dreaded RDRAM. Can you remember a performance increase of over 2-4% with any of those memory types when they came out? I certainly can’t. That’s why my expectations weren’t exactly high when we received our first DDR SDRAM platform.

The introduction of DDR SDRAM is indeed an impressive thing. Performance gains of over 10% are worth a lot of respect. We have never seen anything like that ever before with any other new memory type. As the DDR-chipsets get more mature, BIOSes get tweaked and drivers become more stable we might see even more benefit from this new memory standard. At the same time DDR SDRAM isn’t ridiculously overpriced. RDRAM was never able to live up to Rambus claims. On top of that it was badly overhyped and even forced to the customers. DDR SDRAM is faster, cheaper and its technology is not held by a so called ‘IP’-company that has just lost its very last friend in the business.

Looking at e.g. the MPEG4-benchmark results should also give you an idea of what you can expect from Pentium 4 and i850. Intel’s upcoming Pentium 4 platform will use the very same dual-channel Rambus architecture as i840. Intel has learned that it has made a bad mistake. Pentium 4 will be inhibited by a platform that can only use a poorly performing and expensive memory type. Intel will have to hurry up with its DDR SDRAM chipsets unless it wants to lose even more market share to AMD.

Finally I’d like to commend AMD on a well performing chipset as well as processor. Now we are only waiting for 760MP, the multi-processor version of the 760 chipset. VIA will need to do some more work with VT8633, but it has nothing to worry. For the time being VIA will be the only provider of a DDR266-chipset for Pentium III and Celeron. This is just the same situation as last year, when VIA started its success story as the only provider of PC133-chipsets.