One of the first memory modules to reach stable DDR500 performance levels was OCZ 3700 GOLD. Many of us at AnandTech have been impressed with the performance levels and timings we have seen with this memory. We were a bit surprised, therefore, when we learned that this memory was reportedly achieving these performance levels using Samsung TCB3 memory chips.

The surprise was not that OCZ was using Samsung chips in 3700 GOLD — it is common knowledge that almost every high-end memory manufacturer builds their modules with memory from a wide range of memory chip makers. They may order custom blanks, custom label, or do special "binning" to choose chips for their DIMMs, but virtually all of the custom memory makers use chips from huge memory makers like Samsung, Micron, and Hynix. The surprise in this case was the Samsung chips OCZ was using. We had seen the chips used in DDR333 and DDR400 DIMMs from Samsung and a few other manufacturers. In fact, we had a pair of Crucial DDR333 DIMMs that used Samsung chips with the same ID. We wondered how OCZ was able to get these DDR333/400 chips to perform at the stable DDR466/500 levels, which we had seen in our OCZ 3700 GOLD tests for the upcoming article, 'Searching for the Memory Holy Grail — Part 2'.

A Closer Look

We already knew that the Crucial DIMMs using Samsung chips were outstanding performers, but we decided to take a closer look at the DIMMs and their performance.

The Crucial single-sided 256Mb DIMM carries the Samsung label that identifies it as using CB3 chips.

The chip markings are difficult to read in the photograph, but the chip is marked as "K4H560838D-TCB3". A check with Samsung specifications at http://www.samsung.com/Products/Semiconductor/ show that these chips are used in modules of both DDR333 and DDR400 DIMMs. We have seen them used at DDR333 CAS2.5 and DDR400 CAS3.0.

Comparing Memory Bandwidth: UNBuffered Memory Performance

In our article, 'Searching for the Memory Holy Grail — Part 1', we demonstrated that the UNBuffered SiSoft Sandra Memory Test correlated very well with MemTest86 bandwidth results. With all of the Buffering techniques disabled, scores are much lower — sometimes a 50% drop — than the scores you are accustomed to seeing for Sandra Memory tests. However, the unbuffered scores have proven to be a very sensitive measure of memory bandwidth. Regarding memory performance, we will be looking at the maximum stable overclock for the memory and the SiSoft Sandra UNBuffered Memory Test.

The idea of the UNBuffered Memory Benchmark is very simple — you merely turn off all memory buffering techniques. Sandra makes this very easy to do. Select 'Memory Benchmark', right-click 'Module Options', and uncheck the nine boxes that are related to buffering.

Please keep in mind that the Crucial DDR333 modules using Samsung memory are single-sided DIMMs. As we illustrated in 'Searching for the Memory Holy Grail — Part 1', the performance of two SS modules is lower than the performance of two DS modules in 875/865 boards, and cannot be directly compared.

All memory tests were run on an Abit IC7, Intel 875 Canterwood motherboard with BIOS revision 1.5, and a 2.4C Pentium 4, 800FSB processor that has reached 290MHz (1160FSB) on this motherboard.

Crucial DDR333 Samsung SS Performance Intel 875 Chipset, Dual-Channel, Maximum Overclock
DDR Memory Speed	Memory Timings	Memory Voltage (vDIMM)	UNBuffered Sandra 2003 Memory Test (MB/Second)
450	Maximum SPD 2.5-7-4-4	2.7V	2472 INT 2460 FLT
460	3-8-4-4	2.8V	2436 INT 2527 FLT

Given the ability of this Samsung memory, rated at DDR333 or DDR400, to reach DDR460 at 2.8V, we can fully understand why OCZ chose this as their base memory chip. However, as good as this performance is, it does not match the specified performance of OCZ 3700 GOLD of DDR466, 2.5-7-3-3, at 2.65V. So how does a manufacturer of high-speed memory get from this base performance to the performance of 3700 GOLD?

Lasering for Heat Dissipation

When information about OCZ using Samsung TCB3 chips for 3700 GOLD first appeared, others were asking the same question we were asking: how did OCZ get a memory rated at DDR333/400 to the performance levels of 3700 GOLD — or did they?

All memory manufacturers (as well as CPU and chipset manufacturers) "bin" chips. This means that they sort chips by speed capabilities. The top-performing chips are then used in the highest-speed products — the rest going into lower rated products. OCZ explained that, in addition to speed binning, they laser the chips to improve heat dissipation before sorting, using the best chips for OCZ 3700 GOLD. They also released an internal document describing the lasering process and how it improves heat dissipation. The information below is from the released internal document:

"INTERNAL USE ONLY : ENGINEERING DEPT -TW-CA

Thermal modification of standard DDR DRAM TSOP packages , for increased thermal dissipation and speed reclassification.

Note #1 : Reference EL DDR packing specification is linked at /spec/elddr.html

Purpose : Increase speed of ( SEC packaged ) TSOP , through using EL DDR packaging theory. To fill product shortage during die revision change

EL DDR Packaging : by decreasing the thickness of encapsulate (low thermal transfer rate ) we are able to effectively increase heat dissipation in DRAM and thereby increase the effective speed of each IC . Using a third party fabrication and packaging plant this theory has been proven and the (B) die revision OCZ EL DDR die had great success. Average maximum speed increase during stage 2 testing has been verified over 23 MHz.

Proposition: While dies fabricated by 3rd parties for OCZ in 3rd party packaging have proven successful to this point . We feel that to insure our success we should supplement by speed binning large quantities of 3rd party IC's and by physically modifying the package decreased thermal resistance and re-speed grade the ICs using the standard Advantest ATE.

Industry Disclosure: See Ryan Petersen

Process Engineering : See XXXXXXXXX

Process Overview :

Samsung semi will provide die location information , we will be using the DP-7 CO 2 laser to remove the surface material from the package face only average laser cut depth is 7mill . After surface removal the package should be sent to ISE for environmental test and then re-speed graded after speed grade and classification chips will be marked ( standard process) and standard process thereafter."

Testing the Process

The Lasering explanation sounds good, but we at AnandTech are proud of being skeptics. Armed with the impressive base numbers above from a totally different source, we asked OCZ Technology for more info on their process. Furthermore, we asked OCZ for DIMMs made with unbinned, unlasered chips, as well as unbinned, lasered chips. The requested samples were received the next day.

As you can see in the unlasered samples, the chips are indeed Samsung TCB3. However, the Samsung labeling on the chips used by OCZ indicate a later revision number (E) than our earlier Samsung on the Crucial DIMMs (D).

A quick check at Samsung shows the later chips rated at the same specs to DDR400. We asked Ryan Peterson of OCZ about the significance, if any, of the later revision. He indicated that the later E rev chips do perform better at high speeds than D rev chips.

Pre-Laser Performance

Serial Presence Detect (SPD) timings for the unbinned, unlasered DIMMs were at the common settings that we see for Samsung DDR400, namely 3-8-3-3. To get a better idea of the performance capabilities that could be compared to a finished product, we tested maximum speed at SPD as well as maximum speed at the OCZ 3700 GOLD SPD ratings. We also tested for SPD minimum voltage at DDR466 to try to determine what impact, if any, the lasering and binning process had in comparison to the retail performance of OCZ 3700 Gold.

OCZ Raw Samsung DS Performance Intel 875 Chipset, Dual-Channel, Maximum Overclock
DDR Memory Speed	Memory Timings	Memory Voltage (vDIMM)	UNBuffered Sandra 2003 Memory Test (MB/Second)
466	Maximum SPD 3-8-4-4	2.7V	2894 INT 2916 FLT
466	2.5-7-3-3	2.8V	3042 INT 3084 FLT
476	3-8-4-4	2.8V	2864 INT 2992 FLT

The Revision E chips appear to perform better than the earlier Revision D. Compare the Revision D chips’ maximum speed of DDR460 at 3-8-4-4, memory voltage of 2.8V, to what we are now reaching: DDR476 with Revision E. Perhaps more significantly, at 2.5-7-3-3 timings, we increase to DDR466 (at 2.8V) compared to the DDR450 maximum speed at these same timings with Revision D chips.

Post-Laser Performance

The lasering process includes the removal of a small amount of surface from the Samsung chips, and OCZ remarking the chips.

The OCZ remark indicates a performance rating of 4.3ns, which is quite an improvement from the original Samsung rating.

We have already seen the Samsung chips perform well beyond their rated performance and timings, so what effect does the Laser process have on performance and timings? Is the claimed 4.3ns rating after laser treatment justified?

To offer a better comparison to the finished product, we looked at the HIGHEST overclock at SPD timings, the required voltage at OCZ 3700 GOLD rating of 2.5-7-3-3 and DDR466, and the lowest voltage required for SPD 3-8-4-4 rating at DDR466.

OCZ Lasered Samsung DS Performance Intel 875 Chipset, Dual-Channel, Maximum Overclock
DDR Memory Speed	Memory Timings	Memory Voltage (vDIMM)	UNBuffered Sandra 2003 Memory Test (MB/Second)
466	3-8-4-4	Minimum SPD voltage 2.5V	2877 INT 2923 FLT
466	2.5-7-3-3	Minimum GOLD-Spec Voltage 2.65V	3003 INT 3102 FLT
500	Maximum SPD 3-8-4-4	2.8V	3064 INT 3158 FLT

As OCZ claims, the laser process seems to have a significant effect on performance in two areas. First, it allows the memory chips to achieve a higher overclock. In our tests, the maximum overclock went from 476 to DDR500 – just with the laser process. Second, and more important, lasering appears to allow the use of lower voltages at the same memory speeds. In our test at DDR466 and 2.5-7-3-3, the laser process reduced the required memory voltage (vDIMM) for stable performance from 2.8V to 2.65V. This is a significant improvement in specifications.

Retail Performance

Both the Samsung stock DIMMs and the Lasered DIMMs were assembled from unbinned chips – chips pulled prior to sorting by performance. The final step is sorting the lasered chips and selecting the best-performing chips for the best-performing memory. Chips that do not meet the highest performance levels are used in lower-speed products.

The chips are also only one component in the final performance of the product. The PCB on which the chips are mounted is very important to performance. Six-layer boards are generally lower in noise and the quality of the PCB can significantly impact the performance of the memory DIMM. The programming of the SPD is also very important in determining the final performance of the memory module.

Using the best lasered chips, we finally get to the performance of the completed Retail OCZ 3700 GOLD. In this case, we looked at Maximum Speed at SPD timings, Maximum Speed at CAS2, which is faster than rated SPD CAS of 2.5, and performance at Specifications of DDR466, 2.5-7-3-3.

OCZ3700 GOLD DS 256Mb Performance Intel 875 Chipset, Dual-Channel, Maximum Overclock
DDR Memory Speed	Memory Timings	Memory Voltage (vDIMM)	UNBuffered Sandra 2003 Memory Test (MB/Second)
466	Specification 2.5-7-3-3	2.65V	3079 INT 3146 FLT
480	2.0-7-3-3	2.7V	3175 INT 3216 FLT
510	Maximum SPD 2.5-7-3-3	2.8V	3403 INT 3471 FLT

Speed-sorting of chips yields a retail product that now reaches to a speed of DDR510 at stock SPD timings with a voltage increase to 2.8V. We also see that speed-binning is producing even faster CAS 2.0 settings, up to DDR480. Probably the most significant performance improvement from speed-binning is the memory-bandwidth, as measured by the UNBuffered Sandra 2003 Memory Test.

DDR510 is the highest stable setting that we could achieve at SPD timings, but the memory will reach even higher speeds at more relaxed timings. OCZ 3700 GOLD is one of the many included in our evaluation of high-speed DDR500/DDR466 memory modules. Look for the results in 'Searching for the Memory Holy Grail — Part 2'. Our performance comparison of the fastest memory from Corsair, Kingston, Geil, Adata, and OCZ will be published soon.