We have some good data coming from a Chinese reviewer;
12900K with a
DDR4-3600 CL16-16-16 XMP kit against a
DDR5-4800 CL40-40-40 XMP kit.
- Rainbow Six Siege: DDR4 +4.5% faster than DDR5
- Tomb Raider: DDR4 +4.3% faster than DDR5
- Far Cry 6: DDR4 +4.5% faster than DDR5
So about ~4.5% faster for the DDR4-3600 kit, and if we look back at DDR4, we can see that an increase of 600 MT/s (IMC +300 MHz) improves performance by roughly 5.5% (3000 CL16 to 3600 CL16), but that's with.. such a big IMC frequency increase, which we only see half of on DDR5 since it's running in Gear 2. But based on all of this information, old and new, with a 400MT/s (IMC +100MHz) and CL40 drop to CL38 (4800 CL40 to 5200 CL38), it would be safe to say, that DDR4-3600 CL16 XMP should be roughly (±1-2%) above/below of a DDR5-5200 CL38 kit,
basically the same.
So what does this mean? Well it doesn't tell us anything new, we've known this information for many weeks now, what we didn't know was how difficult it would be to overclock DDR5, because that's what it needs, and in the past few days we've learned a lot about just that, and it turns out to be harder than anticipated, and there are far fewer high frequency kits coming.
But let's just get past the value question, from this information above, it'd be unwise to recommend a DDR5 system for gamers, because the increased bandwidth from around 60GB/s to about ~90GB/s won't benefit you in any way (in gaming). So as of this moment, the cheapest 2x16GB DDR5-5200 CL38 kit I can find is $279 and the cheapest 2x16GB DDR4-3600 CL16 kit is $199, and they perform the same in gaming, so that's a 40% higher price for a ~30GB/s bandwidth increase on DDR5, that won't get you anything in gaming (it will in certain work loads). If we look at 8GB sticks things are getting dramatic, you get a 2x8GB DDR4-3600 CL16 kit for as little as $79, while for the cheapest 2x8GB DDR5-4800 CL40 kit I can find, you have to pay $140, so that's +75% and the DDR4 kit will be ~4.5% faster. And if you want something faster, we have the G.Skill 2x16GB DDR5-5600 CL36 for $380, if I had to take a guess that XMP would be around the performance of DDR4 at 3800 CL16, it's still uncertain how high DDR4 can go on the new IMC featured in Alder Lake, but rumours are as high as 4200, but I won't comment further than that, we'll have it confirmed very soon. And for DDR5 we have confirmation of between 8000 to 8300 MT/s, so about the same IMC frequency as for DDR4.
Conclusion on value? It's definitely more expensive for the same performance, in between 40 to 75%.
Now, a bit more info on the IMC, this will explain the challenge that DDR5 faces.
IMC (Integrated Memory Controller) on the Coffee Lake (9900K) and Comet Lake (10700K) processors was strong enough to get DDR4 frequencies all the way up to 4600 with ease, and on these architectures, Gear 1 and 2 wasn't a thing, so it always ran a 1:1 ratio, DRAM Frequency x IMC Frequency, so a 4000 MT/s speed meant a true frequency of 2000 MHz, and in a 1:1 ratio that meant IMC Frequency was also 2000 MHz, so the 9900K-10700K had no issues reaching a 4600 MT/s, that tells us the IMC was strong enough to go up to 2300 MHz and above.
But this changed on Rocket Lake (11700K), with a new IMC and Intel introducing Gears (Ratio Control between the DRAM and IMC Frequency), in preparation for DDR5, we saw a significant reduction in possible IMC frequencies, it was now difficult to reach even DDR4-3866 (1933 MHz IMC), and why does this matter? Well, it's significantly slower, from the above comparison it'd be a 16% reduction in IMC Frequency, and why does frequency matter on the IMC you might ask?
That leads us to what actually matters, for gamers.. speed! But there are different kinds of speed.
So, what sets games and work loads apart, a work load is typically storing or using a large amount of data, that gets chewed through over time by the processor. While a game, typically stores small amounts of data, that needs to be replaced and processed very fast, this can be demonstrated using storage drives very well;
Many, many years ago, almost a decade now, we had SSDs introduced to our gaming machines, and games loaded a lot faster, but why? So, old hard drive disks had a very slow access/seek time, because it had to use the actuator arm to find the data it needed, on spinning disks (large surface). As for the bandwidth, if you had a very large file, you could reach a sequential read MB/s of up to 150MB/s through the SATA interface, and if you put two disks in a striped RAID, you could double it to 300MB/s, but what did that mean for game loading times? It was faster, but it was still nowhere near the earliest SSDs on the market, that completely skipped the mechanical (moving) parts, because bandwidth was never the issue, being able to reach a high sequential read speed is next to completely irrelevant for gaming. This can be very easily seen through modern M.2 SSDs, like, on the SATA3 interface we had a theoretical bandwidth of up to 600MB/s, but after removing some overhead and such, we saw a typical max read speed of around 550MB/s, this is up to 4 times higher than a traditional HDD, but again, if you put two of these SSDs in a striped RAID, that doubled the sequential read speed to 1.1GB/s, games saw next to none, or no increase at all in loading times. Because again, bandwidth
doesn't matter, it's all about latency (accessing a file as fast as possible), that's how games operate, we're talking anything in between 5 to 300 MB large files usually. The true demonstration to really get my point across is M.2 SSDs as mentioned earlier, we went from SATA 2 300MB/s to SATA 3 600MB/s, to M.2 PCIe 3.0 3500MB/s, to 7000MB/s today on PCie 4.0 M.2s, and what can we observe? Well..
nothing, literally. Since the game files are so small, having a game installed on a 7000MB/s 1TB M.2 for $200 is barely going to improve loading times, we're talking between 0.1 to 1 second faster on a several second long loading screen,
compared to a decade old 256GB (now valued at $20) SATA3 SSD with a maximum read speed of 550MB/s, so again, access time (latency) is what matters here,
and the same exact principle applies to DRAM.
So by doubling the memory channels on DDR3, Single to Dual Channel, we went from 12.5 to 25 GB/s read, and using Quad Channel we went from 25 to 50 GB/s.
On DDR4 Dual Channel we went from 25 to 50 GB/s (same as Quad Channel DDR3), and DDR4 Quad Channel we went from 50 to 100 GB/s.
Now.. what actually changed when going Quad Channel on DDR4? The bandwidth doubled.. but did the FPS double? Eh.. no,
literally no change, because as you now know, bandwidth is irrelevant (for gaming), as long as you reach a threshold where the bandwidth is enough, which we passed long ago.
That leads us to DDR5, in Dual Channel we're looking at up to 100 GB/s (same as Quad Channel DDR4), but.. again, doesn't mean anything, we don't want bandwidth, we want access time/latency just like on SSDs!
So, if we try and forget bandwidth, how can you lower your latency? By increasing frequency (bandwidth) of course?.. No, actually, what is happening underneath the surface, is that we don't actually care about the DRAM frequency (4000 MT/s), what we're actually overclocking.. is the IMC (Memory Controller), the faster the controller frequency, the faster the access time/latency!
This is also why it's so important to lower your memory timings in BIOS, they are massively important in making your RAM faster (lowers the access time), you are making yourself a big disservice by running XMP!
As seen in one of my benchmarks, the only way to make your memory faster, is by increasing IMC frequency at the same time as lowering the timings. Framerate then rockets to the moon when increasing frequency
and lowering timings
as much as possible, and as seen in the 4133 MT/s 16-16-16 result, absolutely massive performance gain (very low latency)!
3466 MT/s 16-18-18 (1733 MHz IMC) is just as fast as 4133 MT/s 19-19-19 (2066 MHz IMC), that's something you might not have expected, since 4133 looks a lot faster than 3466, cost a lot more and required a stronger motherboard, which increased costs of the system even further. The 3200 MT/s 14-14-14 also shows that it challenges 4000 MT/s 18-19-19, and let me remind you, 800 MT/s faster is a decent amount of higher bandwidth, which as you now know, doesn't really matter, which this also shows.
So, back to DDR5, what do we actually want with DDR5? Do we want a high DRAM frequency? No. Do we want a high IMC frequency? Yes! How do you get higher IMC Frequency? By increasing DRAM frequency!
This is the current problem with DDR5, for DDR5 to be as fast as it can in gaming, it has to reach as low latency as possible, and DDR5 is limited to Gear 2, meaning that the Gear Ratio I mentioned earlier, is half, Gear 1 is 1:1, so if DDR4-4000 runs Gear 1, that means actual DRAM Frequency is 2000 MHz, and IMC is running at 2000 MHz, that is very fast! Not as fast as 2300 on Comet Lake and earlier processors, but fast nonetheless. So, with DDR5 you obviously can't run Gear 1 (1:1 Ratio), since running 5200 MT/s on DDR5 would mean a 2600 MHz IMC frequency, which is impossible, that leaves us with only one option.. we have to increase the IMC frequency through DRAM frequency! Since the ratio is now ½ (half), running 5200 MT/s (2600 MHz DRAM) on DDR5 results in an IMC frequency of only 1300 MHz (half the DRAM), that is 35% slower than DDR4-4000 in Gear 1, the goal is thus to go as high DRAM frequency as you possibly can, to reach the same 2000 MHz IMC as DDR4 on Gear 1, we'd then have to run DDR5 at 8000 MT/s! Now that is fast, way faster than most boards on the market today can reach, or even sticks, the fastest XMP kit we've seen coming out in the near future is 7000 MT/s with CL40, that's impressive for sure, that's an IMC frequency of 1750 (only 12.5% slower than 2000), combined with other DDR5 performance enhancements, that kind of speed will really put DDR4 to the test! But, to run that frequency, you can't just buy any board, or memory sticks, it'll be very expensive, and at best, even if DDR5 manages to beat DDR4 at a frequency of between 7000 to 8000 MT/s, it will only be by a few %, is that really worth spending that much more money? I'd obviously argue.. no, unless you either are an enthusiast/overclocker or competitive/professional gamer, but for the average person? Absolutely not.
So, where does this leave us? Is DDR4 faster than DDR5 in gaming? Yes.. but DDR5 can also be faster than DDR4! If certain criteria is met,
that happens to cost a lot of money, for example these new high performance kits are likely to cost $500 and above, pre-tax, maybe even $600, as the current 5600 CL36 is already at $380 and 7000 CL40 is
a lot faster!
DDR4 will without question be significantly cheaper, and faster than the majority of DDR5 systems, that's just the reality of it.. for now! Can the most expensive DDR5 system beat the fastest DDR4 system? Again, yes, but that doesn't change the fact that the majority of DDR4 systems will beat DDR5 with ease, it won't even be close really,
if the rumours are true that DDR4 IMC can reach up to 2100 MHz (4200 MT/s), that will put DDR4 ahead of DDR5 by anywhere between 5 to 15% depending on what DDR5 kit you'd go up against, and that's raw performance, 200 to 210-230 FPS for a competitive gamer in a game like Warzone, for a considerably lower price, and easier overclocking since we're all already familiar with DDR4.
Let me repeat myself, DDR5 is amazing, there's no doubt about that.. but DDR4 is also amazing, and it'll take quite some time for DDR5 to become as cheap as DDR4, and even more time for faster memory sticks to appear that beats DDR4 out of box. It's likely that DDR4 will remain the recommended memory for the next two years, until Meteor Lake (14th Gen) arrives in 2023, since Raptor Lake (12th Gen) coming in 2022 will still support DDR4, that's when DDR5 has a chance to really beat DDR4, new cheaper ($200-300) Z790 motherboards with XMP certification up to 8000MT/s and affordable 7000-8000 MT/s DDR5 kits (~$300), but that's
a lot of wishful thinking for just a year until 13th Gen processors drop.
