How hot does DDR5 memory really get? Readout value from SPD hub vs. real measurement at the module | Investigative | Page 2
Disclaimer: The following article is machine translated from the original German, and has not been edited or checked for errors. Thank you for understanding!
DDR5 memory under full load with 1.1 Volt
Let’s first have a look at the measurements in delivery condition and with 1.1 volt voltage default. The practical thing about my camera is that it is a real 32 Hz camera and does not only produce still images. In addition, you can log pretty much all measured values in real time, even with a time stamp, so that the later comparison in a common graph does not cause any problems. To get enough load on the RAM, I use Karhu and log the data from the camera and the SPD hub (via HWInfo64) for a little over 30 minutes each.
We see an increasing difference between the value reported back by the SPD hub and the small hotspot in the middle of the modules. The value from HWInfo for the SPD hub is about 53 °C after 30 minutes, the hub IC itself is also hotter and also the board is a bit hotter than what is reported back here. Also within the chip there is a delta of about 5 degrees between the outer edge area and the narrow point in the middle of the single memory module, which I focused on to stay as accurate as possible.
In the end, we thus record a difference of a whopping 12 degrees between the readout value of the SPD hub and the hotspot in the middle of the memory module. That’s quite a small house number, but less than I feared. Since you also have to consider the thermal resistance between the surface and the interior, experience shows that the maximum temperature inside the volcano is up to 5 degrees. In this case, the final hot-spot temperature would already be just under 70 degrees, i.e. almost 17 degrees above the readout value. Let’s keep that in mind.
DDR5 memory under full load with 1.25 Volt
If we tighten up the timings, we also have to tighten up the tension a bit. Even if OCPC has stored a “benchmark profile” with 1.4 volts in its own settings, due to the lack of too large RAM reserves I leave it at 1.25 volts this time. But also there is already something going on, because the temperatures are rising quite visibly. The RAM goes up to almost 72 °C in the measured hotspot, the SPD stroke value is 60 °C. We already know the 12 degree difference, don’t we? I will make comparative measurements with other modules later to see if a simpler rule of thumb can be derived here. This is because the markup remains identical for the time being, only the final values change. At 1.35 volts, it would then get even hotter….
Here we also see the infrared image of the focused area, where the difference to the edges of the memory mode is almost identical. The only thing I noticed was the renewed, yet very different steep rise in both temperatures. The “hotspot” obviously gets up to speed faster than the somewhat sluggish value from the USB hub, which seems to be a rather averaged value. Even though the memory modules have 2 TS (Thermal Sensors) according to the specs, they are probably averaged a bit here.
Summary and interim conclusion
Yes, you could certainly cool the memory without much overclocking and with the 1.1 volts as voltage specification still without heatspreader, as we know it from the DDR4 server RAM. However, with overclocking and higher operating voltages, things get really tight and instabilities start to accumulate. If the RAM runs cooler, the operation definitely remains more stable. Here one should also think about the reported 60 to 65 °C for the SPD hub, above which it might become unstable.
DDR4 cooler: Better than nothing, but beautiful goes a little differently
Because even if the hotspot isn’t the sole criterion, with a delta of up to 17 degrees (including the losses in the substrate) one is then already with some bad luck and elegantly over 80 °C. Then it surely already becomes a bit more tippy and one should nevertheless think about a better RAM cooler. But we already had this very topic on Saturday. If you missed it or would like to read it again: Here you can read the article “DDR4-RAM-Cooler on DDR5? Not a good idea! The differences in detail including 3D scan and instructions“.
So much for that, I’ll stay tuned of course. I promise.
DDR5 memory from Micron comes without temperature sensors! Blind flight for the user and real measurements as an aid | Page 2
Disclaimer: The following article is machine translated from the original German, and has not been edited or checked for errors. Thank you for understanding!
What is actually already being read out and displayed?
Let’s look at the already known module first. As an example, I have marked two places where temperature sensors really exist, even if one has to make further restrictions here. It concerns on the one hand the SPD hub and on the other hand the PMIC, which also takes over the voltage regulation. With less than 5 watts of power dissipation in OC mode, one bar is certainly not yet a thermal time bomb, but 2 bars already reaches up to 10 watts. If you have all banks full, that’s up to 20 watts just for the RAM. Hallelujah.
The SPD hub as a basis for a possible rule of thumb
The IC for the SPD hub contains an (analog) sensor which returns readable decimal values and which also works quite accurately. In the JEDEC specifications, which are unfortunately hidden behind a paywall and for which you have to pay a lot, you will also find the temperature limits for this SPD hub. Since the IC itself produces hardly any power dissipation, but sticks directly to the PCB, the IC is a good indicator for the temperature of the entire PCB, which after about 30 minutes of full load has also warmed up in a fairly balanced way. Hotspots dissolve into averages over time, after all.
| High limit | 55 °C |
| Critical limit | 85 °C |
Despite OC and some tests with up to 1.4 volts in the “benchmark mode” of the OCPC Extreme, I did not succeed in bringing this value to the critical limit. With readout values of slightly over 70 °C at 1.35 volts and a maximum of 75 °C at 1.4 volts, the critical limit will therefore not be reached. So I can give the all-clear, everything remains in the green area.
But what can you use the value for the SPD hub for now? Since the individual memory modules also have a thermal effect on the circuit board, the SPD value can also be indirectly inferred from the memory after a longer period of heating! We will see later that here, depending on the make, there is a temperature difference of between 10 and 13 degrees between the SPD value and the hottest point on the surface of the memory module. What that means for the interior, we will clarify in a moment. But it is at least a rough indication and also a useful guideline for a possible rule of thumb.
The temperature values of the PMIC
The PMIC temperature is not a true sensor value, but again just a collection of flags that only very roughly outline the temperature ranges. Since hardly any power is dissipated in the IC itself and thus hardly any waste heat is generated, we can actually ignore this value calculation completely. You can find the flag in HWINFO under the messages for the respective RAM module as “High Temperature” (see picture above). For the sake of completeness, I have also listed the flags and the corresponding temperature ranges:
| 000 | < 85 °C |
| 001 | 85 °C |
| 010 | 95 °C |
| 011 | 105 °C |
| 100 | 115 °C |
| 101 | 125 °C |
| 110 | 135 °C |
| 111 | > 140 °C |
So we can completely skip the PMIC, unlike the SPD hub, because the value is of no use to us. Before a critical flag appears here, the memory modules have long since burned up. On the next page we will see how warm or even hot it all really gets.
2×8 vs 1×16 vs 4×8 vs 2×16 on Intel Alder Lake — i2HARD
Recently we looked at the significance of different DDR4 configurations on the 12th generation of intel processors. Today we will touch on another aspect of memory and already on DDR5. The first thing you need to know, since we have a two-channel in one module, is there a big difference with a four-channel in two modules? And the second aspect is no less important: is it worth considering 8-gigabyte modules. And we are not talking about the sufficiency of 16 GB of RAM, but about the number of chips in the module itself. nine0005
Contents:
- Test bench
- Stock tests
- Synthetic tests
- Tests in games
- Overclocking tests
- Synthetic tests
- Tests in games
- DDR5 temperature measurements
What is the salt? We are all used to seeing 8 Gigabyte chips on 8 GB DDR4 modules. But there are also modules with four 2 GB chips. And they are much worse.
For example, in Aida, the throughput of such modules differs by about 10%. nine0005
And if you go into Cyberpunk, then with full modules FPS is 17% higher.
In simple words: the more memory chips, the better. While the cells in one chip are being recharged, the other one is working. Recall the same two-rank memory, where ranks alternate in a similar way. Accordingly, it would be necessary to deal with this aspect on DDR5. Therefore, we compare 4 configurations: 2 modules of 8 GB, 4 of 8, 1 of 16 and 2 of 16.
Testbench
Since the 8 GB modules do not have their own XMP, JEDEC will be used in stock. An interesting point: all 4 modules were bought separately, and two of them have Samsung chips in the SPD, and the other two have Hynix chips. At the same time, if you look at the chips themselves, only Samsung is visible everywhere, which once again confirms the inexpediency of determining not only the revision, but also the chip manufacturer using the software method. nine0005
2 x 16 GB modules from TeamGroup, presumably on Hynix chips, will also be used in JEDEC with the same timings, so as not to introduce confusion. We’ll talk about overclocking in the second half of the review.
- Graphics Card: Palit GeForce RTX 3080 Ti GameRock OC
- Processor: Intel Core i7-12700K
- Motherboard: ASUS ROG Maximus Z690 Hero
- RAM #1: Patriot Signature LinePSD58G480041 RAM
- Cooling System: ARCTIC Liquid Freezer II-360
- Drive #1: Crucial MX500 2 TB
- Drive #2: ADATA XPG Gammix S50 Lite 1 TB Stock tests
Synthetic tests
So. A handy thing is JEDEC. Unlike XMP, you don’t even need to go into the BIOS for it to work; however, for 4 modules, a different profile is used. Not 4800 MHz CL40, but 4000 MHz CL36. And this could not but affect both the memory bandwidth and latency. 4 modules broke through the 100ns level, which we have not seen for a long time.
Pairs of modules of different sizes demonstrate the very consequences of using fewer chips in a module. 16-gigabyte modules are ahead in all respects. One 16 GB module has a similar latency to two, but its throughput is almost 2 times lower. nine0005
For GeekBench 5, 2 and 4 8 GB modules turned out to be equal, despite such different JEDECs. 2 x 16 GB modules outperform them by 7% and are 23% faster than a single module.
For Premiere Pro, memory bandwidth is incredibly important, so two-channel is an outsider here, and 4 modules with low-frequency JEDEC are a little behind.
Synthetics was warming up, we’ll make the main comparison in games.
Tests in games
Call of Duty: Warzone. The eSports settings, the tests were repeated and averaged many times, and the recording was made with a capture card from the second computer. The situation is close to that in GeekBench. 4 modules of 8 GB each compensate for the weaker JEDEC with the number of chips.
The other 32GB kit is ahead by 4%, which clearly reduces our concerns. Even 1 module is not so far behind. This is a noticeable 15%, but with such an FPS in general, without monitoring, and even with an emphasis on the GPU, it turns out not critical. nine0005
CyberPunk 2077. Graphics preset — ultra ray tracing, RT reflections disabled, DLSS — ultra performance, high crowd density. This game loves memory bandwidth much more and in this regard it is close to Premiere Pro. 4 modules of 8 GB are behind two. 32 GB with two modules is ahead by another 10%, and compared to a single dual-channel module, they give 29% more FPS. This difference is 2 times greater than in Warzone.
Shadow of Lara, highest graphics preset, resolution modifier — 20%. In the tomb raider again we see not such a big difference. Pairs of modules of different volumes differ by 6%, a four-channel is 17% faster than a two-channel, and 2 and 4 modules of the same volume, but with different frequencies and timings, are barely distinguishable.
On average FPS, you can scrape together only 3%. nine0005
StarCraft II, all settings maxed out. Who cares about memory bandwidth is the old games. One module quietly provides the necessary speed, and the rest already depends on the delays. And what’s funny, in one 16 GB module they are lower than in two 8 GB modules, and, accordingly, the FPS is higher. The four-channel gave a ridiculous 2% increase in average FPS. And in general, all memory options plus or minus are close.
Total War Saga: Troy, ultra preset, grass and units — extreme, resolution and anti-aliasing to a minimum. Completely opposite to the previous strategy, Troy leads to completely different results. Quad-channel memory mode increases FPS compared to dual-channel not by 2%, but by one and a half times. Even more chips within two modules gives a 10% increase in average FPS. But what has remained unchanged is the meager difference between 2 and 4 8 GB modules, despite a decent difference in frequency. nine0005
On average for a ward, it turns out that the purchase of two 8 GB DDR5 modules does not lead to large performance losses due to a smaller number of chips, as was the case with DDR4.
It comes out to about 7%, and it’s better to take them than one 16 GB module if your goal is to get as much FPS as possible here and now.
Naturally, buying DDR5 in this form makes little sense. We covered this in the i3-12100 review, where DDR5 at 4800 MHz often lost to DDR4 at 3200 MHz, so let’s look at overclocking. nine0005
Overclocking tests
So, for all configurations except 4 to 8, the same overclock will be used — again, to avoid confusion. It turned out 5800 MHz with the first timing of 30. This is much better than what we would have had on microns at one time.
Overclocking 4 DDR5 modules is currently a separate kind of pleasure that requires its own suit. Let’s start with the fact that after 4400 MHz, each frequency step requires a disproportionate increase in primary timings, and end up with the fact that high VDD voltage, which allows them to be reduced, leads to instability. Therefore, 4400 MHz came out with the first timing of 26.
Synthetic tests
Now that all timings are set identically, the latency is the same.
Although in JEDEC everything was the same for us. The gap in bandwidth also decreased between pairs of modules of different sizes. However, one module is still almost twice as far behind. A modest overclocking of 4 modules of 8 GB each, which many DDR4 modules will laugh at, is 10 ns inferior in latency, and the bandwidth is a quarter lower.
Thanks to the timings setting, and the higher frequency in general, the recharging time of the cells in the memory modules has been reduced, so a larger number of them gives less growth. If earlier there was a 7% difference in GeekBench between two modules of different sizes, now it has been reduced to 3%. nine0005
In Premiere, too, instead of 13%, we get only a 4% gap between 2 modules of 8 GB and 2 of 16. And the meager overclocking of 4 modules only led to a greater gap from two.
Tests in games
We return to Warzone and see that there is no difference between the extreme memory kits. One module lags behind them by 12%, and 4 to 8 are somewhere in the middle in terms of performance.
In Cyberpunk there is also practically no difference between pairs of modules of different sizes. 4 modules with a frequency of a whole gigahertz lower reduce the number of frames prepared by the processor by 8%, but a four-channel is 20% faster than a two-channel. Fortunately or unfortunately, in these, and in most other games, the emphasis on the video card and the lack of monitoring will not allow you to feel the difference. nine0005
In Lara, the large L3 cache partly compensates for slow memory, allowing fewer accesses to it. In this regard, the four-channel mode gives only an 11% advantage over the two-channel one, and weak overclocking of 4 modules leads to a loss of only 6% in average FPS.
In Starcraft, the difference between different channel configurations remained at the same level at around 2%. The rest of the results are even closer to each other. As in previous games, 2 overclocked 8 GB modules have almost identical performance to the same number of 16 GB modules.
And a higher latency of 4 modules leads to a decrease in FPS by 4%. nine0005
An interesting result came out in Troy. If in stock the four-channel had 51% dominance over the two-channel, now the difference has been reduced to 33%. That is, it looks like there is a certain threshold, after which the increase from overclocking gives a much smaller increase in FPS. If one 16 GB module received a 28% increase from overclocking and it still lags behind two modules in JEDEC, then they received only a 12% increase from the same setting.
On average, it turns out that if you have enough 16 GB of total RAM, then two overclocked modules will have almost the same performance as 32 GB with the same number of modules. In theory, the same can be applied if memory with 4 4 GB chips appears in the future. nine0005
In turn, it’s worth buying one module, just so you don’t have to suffer with overclocking four later, but keep in mind that if in the future the second module comes across on chips from another manufacturer or another revision, then overclocking can be problematic.
DDR5 temperature readings
DDR5 temperature readings without any artificial airflow are a bonus. Only convection.
Starting with two modules without heatsinks in JEDEC. One of them warmed up, judging by the internal sensor, to 50 ° C in 15 minutes of the TM5 stress test. What will the thermal imager show? 54°C. Higher, but not much. Go ahead. nine0005
With a maximum voltage of 1.435 Volts for such a memory in the BIOS, the temperature is quite acceptable — almost 62°C. True, without airflow, our overclocking became unstable and after 2 minutes rolled into a BSOD. That’s just if you look through the thermal imager, you can shudder in surprise. 85°C is no longer a joke. This is quite the operating temperature for the chips, but go even hotter inside.
What about modules with heatsinks? Despite its presence, twice as many chips did their job: the module warmed up higher than it was on uncovered modules with the same JEDEC. It is difficult to say how much the temperature of the chip and heatsink differs, but without airflow, the difference should be small.
nine0005
When using XMP, the heatsink temperature reached almost 70°C, causing errors. This is the turn…
In overclocking, we have 68 ° C on the sensor and 74 on the radiator.
Looks better than it was on four chips not covered by a heatsink, but, again, we don’t know the temperature delta between the heatsink, the chip case, and its inside. Therefore, a draft in the case is very desirable during overclocking.
And you shouldn’t draw parallels with the processor, saying «It’s hot in a stress test, it will be less in games.» Will not be. Not only will modern games tirelessly drive data through your RAM, but also the video card will throw hot air at it. Therefore, either a collective farm or a well-ventilated building. nine0005
Corsair stated that DDR5 memory modules will require powerful cooling systems
3DNews Technologies and IT market. News RAM modules, memory cards, flash drives.
.. Corsair announced that DDR5 memory modules…
08/06/2021 [16:24],
Nikolai Khizhnyak
Many leading gaming memory manufacturers, such as Corsair, have already announced that the new DDR5 RAM standard will be able to operate at an effective frequency of up to 12600 MHz. These speeds will be driven largely by power controllers (PMICs) and voltage regulation modules (VRMs), which will require a very high quality cooling system, says Corsair. nine0005
Image Source: Corsair
“Operating temperatures for DDR5 modules can be significantly higher than for DDR4 modules because power management has been moved from motherboards to the memory sticks themselves. As a result, they can generate more heat,” commented George Makris, advertising director for Corsair.
For many years, it was considered that the presence of a cooling system on DDR3 and DDR4 RAM modules was not mandatory, even when it comes to enthusiast RAM.
With the advent of the DDR5 standard, everything can change, since the memory modules in this case have become more complicated. The same Corsair noted that it is already developing cooling systems for the new RAM. nine0005
The standard for DDR5 operation is 1.1 V. However, the use of PMIC and VRM controllers in DDR5 modules will allow memory manufacturers to create more efficient solutions. For example, ADATA recently announced that it will also produce DDR5 memory modules operating at 1.6 V. At the moment, one can only guess how CPU memory controllers built in accordance with process standards 5 and 7 nm, because it will increase by 45% compared to the load created by memory with a voltage of 1.1 V. The PMIC and VRM controllers themselves also need cooling. As a result, DDR5 will require the use of more powerful cooling than DDR4. nine0005
By the way, Corsair has a lot of experience in developing efficient cooling systems. For example, as part of its Dominator series RAM, heatsinks are used that cool not only the DRAM chips, but also the printed circuit board of the memory module.
4×8 -Force Delta RGB FF3D532G6200HC38ADC01 2x16GB
