Lenovo ThinkPad T16 G1 AMD Laptop-Test: Sparsamer & schneller als Intel

[Redaktion]

Das Lenovo ThinkPad T16 AMD ist unter den 16-Zoll-Business-Laptops etwas allein, besitzt mit einer Ryzen-6000-CPU jedoch Ambitionen für eine schlanke Workstation, ähnlich dem ThinkPad P16s, von dem es sich kaum unterscheidet. Dank WQXGA-Display und RDNA-2-GPU ist es auch für einfache Multimedia-Tasks geeignet.

https://www.notebookcheck.com/Lenovo-ThinkPad-T16-G1-AMD-Laptop-Test-Sparsamer-schneller-als-Intel.667833.0.html

[RobertJasiek]

Marvin, deine Tests sind immer eine Freude!

Wie erklären sich die schlechten RAM-Messwerte trotz nominell sehr gutem RAM?

[Marvin Gollor]

Marvin, deine Tests sind immer eine Freude!

Wie erklären sich die schlechten RAM-Messwerte trotz nominell sehr gutem RAM?

Danke! <3

Die Messwerte entsprechen in etwa der advertised Bandwidth: "The cutting-edge speed enables huge transfers to be made at 51.2 GB/s" https://semiconductor.samsung.com/dram/lpddr/lpddr5/

DDR5-4800 hat hingegen eine Bandwidth von 76,8 GB/s. Denke, das liegt daran, dass LPDDR5-RAM meines Wissens auf zwei 16-bit-Channel setzt und regulärer DDR5-RAM auf zwei 32-bit-Channel. Sprich halbe Anzahl an bits + 1/3 höhere Frequenz entspricht genau 51,2 GB/s vs 76,8 GB/s.

Übrigens: AMDs iGPU profitiert ja eher von der rohen Frequenz, weshalb LPDDR5-RAM da schon mehr bringt.

[RobertJasiek]

Auch die Latenz ist hoch. Ebenfalls wegen LPDDR5 statt DDR5?

[NikoB]

No, LPDDR4/5, like DDR4/5, always work on a 128-bit bus. But DDR has 2x64bit, and LPDDR has 4x32bit - so it is always displayed as quad channel in cpu-z.

The problem is not in the memory, but in the shamefully slow AMD memory controllers, which has already been proven by dozens of reviews.

Intel memory controllers are at least 1.5 times faster and in top models like the Yoga Pro, the 12700H with LPDDR5 5200 is almost 2 times faster than AMD with 6900HX and LPDDR5 6400.

This is a complete shame for AMD.

[Marvin Gollor]

DDR5 has 2x 32 bit per channel while DDR4 has 1x 64 bit per channel. So both have 64 bit per DIMM and 128 bit in dual-DIMM-configuration, yeah. But LPDDR5 has 2x 16 bit per channel, so 32 bit per DIMM and since in reality it's also in dual DIMM mode, is the reason why it is 64 bit instead of 128 bit.

But sometimes it has 2x 32 bit per channel, too, which is 64 bit per DIMM and 128 bit in dual-DIMM-config, just like regular DDR5-RAM, but with lower voltages. But just Intel uses it and it doesn't exceed LPDDR5-5500 in actual devices, yet. So LPDDR5-6400 is always half bit-width, while the only AMD machine with 32 bit per channel LPDDR5-RAM I can find so far, is the steam Deck (LPDDR5-5500 with 88 GB/s).

The RAM in the mentioned Lenovo Yoga Slim 7i Pro X has 32 bit-per-channel LPDDR5-5200-RAM. So its ~82GB/s instead of regular 41 GB/s bandwidth and therefore matches regular non-LP DDR5-5200-RAM. We could say that it's just regular DDR5-5200-RAM but with lower voltage. LPDDR5-6400-bandwidth on the other hand is just 51 GB/s for which I provided the source (Samsung). Just Apple connects LPDDR5-6400-RAM to a 512-bit-controller instead of 64 bit, which results in 8 times higher bandwidth (~400 GB/s).

dual-channel DDR5-4800 on AMD systems is always ~31% faster in read operations than LPDDR5-6400-RAM with the same AMD-APU (10% faster in write operations), which our benchmarked systems show. There you have proof that LPDDR5-6400-RAM is slower than DDR5-4800-RAM in general, even on AMD-Systems.

But since the speed is higher, the bit-width has to be less (half) to generate such a significant lower bandwidth with the same APU. There is no LPDDR5-6400-RAM with 32 bit per channel, yet. Why do u think Intel doesn't use LPDDR5-6400 if it would be that fast on their systems? Because in reality it's slow, slower than the LPDDR5-5200 they use, which features double the bit-width.

Soon LPDDR5x-8533-RAM will be provided and it will feature just 68,26 GB/s of bandwidth. So less then LPDDR5-5200 and even DDR5-4800 and just ~10% more than DDR4-3200. U still think the have the same bit-width? I don't think an Intel system with LPDDR5x-8533 will get ~136 GB/s of bandwidth.

Auch die Latenz ist hoch. Ebenfalls wegen LPDDR5 statt DDR5?

Ja, LPDDR-RAM ist bekannt für hohe Latenzen. Kann man beim Acer Swift X etwa im BIOS deutlich tweaken, gibt für einige Laptops Mod-BIOSe, bin da aber noch nicht eingelesen.

Man mus sich ja nur AMD-Geräte mit regulären DDR5-RAM anschauen, wie das Legion 5 Pro oder XMG Core 15. Gleiche Latenz wie Intel-Systeme mit regulärem DDR5-RAM. Siehe angehängtes Bild:
Clipboard02.jpg

[RobertJasiek]

Sehr interessant! RAM-Kauf ist zur Wissenschaft geworden, o o...

[NikoB]

But LPDDR5 has 2x 16 bit per channel, so 32 bit per DIMM and since in reality it's also in dual DIMM mode, is the reason why it is 64 bit instead of 128 bit.

Can you prove this statement with a screenshot from AIDA64 for the Chipset>Memory Controller>Active Mode tab, where the active mode of the memory is explicitly indicated in bits?

Sehr interessant! RAM-Kauf ist zur Wissenschaft geworden, o o...

NB does not draw readers' attention to this key point. As well as on the mode of operation of the memory relative to the frequency of the memory controller - Gear1 or Gear2.

Just Apple connects LPDDR5-6400-RAM to a 512-bit-controller instead of 64 bit, which results in 8 times higher bandwidth (~400 GB/s).

Can you prove this statement with some official links and screenshots from the test software? It turns out that in Macbooks with ARM memory is 8 times faster than in x86? It just can't be.

Moreover, your statement about Apple directly contradicts your statements about Intel and the inability to make a controller with a wide bus at a high frequency. If they can't make a 128-bit bus on the 6400, but only the 5200, how can Apple make a 512-bit bus even on the 4800? Do you see a logical contradiction here? If Apple can, then Intel and AMD (and they certainly do everything on TSMC) can do at least 256 bits and certainly 128 by 6400. But this is not.

I have never seen benchmarks anywhere where Apple laptops have memory bandwidth 7-8x faster than x86. It just can't be. Although it is I who constantly advocate and you know this, that the lack of memory bandwidth on the x86 platform has reached a monstrous 10 times relative to what is necessary, which is why both Intel and AMD are rapidly increasing the L3 cache - and this is a pointless crutch, because. outside the cache (as well as on SSDs outside of SLC), the speed immediately drops by an order of magnitude.

And all the tests of Intel and AMD laptops on this and other sites clearly prove that the AMD Zen3+ memory controller is on average 1.5 times slower than the Intel memory controller in Alder Lake. And in peak cases, as in Yoga, it is almost 2 times slower.

Why is there not a single AMD solution with memory that pumps around 68-70GB/s and there is not a single solution with 80GB/s, like on Yoga Pro, where not even a top-end i9 is used, but a regular i7 12700H. Therefore, AMD completely disgraced the memory controller.

---
I already wrote earlier under reviews and news that the new 7945HX from AMD has 28 free pci-e 5.0 lanes, which is at least 112 GB/s need for work it, despite the fact that the memory speed in dual-channel mode with DDR5 5600  and does not come close to this, but really there you need 10 times more, i.e. 1TByte/s. And this can only be done with HBM memory, just with 512 or 1024 memory bus. At the same time, as I already suggested, in view of the fact that video memory has become huge for laptops onboard and is often not used most of the time when working from igpu, it would be a reasonable decision to switch video memory directly to the processor cores through its memory controller with the same bus width as used by dgpu chip. All the same, 16-24GB of such memory is idle in vain in normal work.

[RobertJasiek]

"the mode of operation of the memory relative to the frequency of the memory controller - Gear1 or Gear2."

Please explain!

As to HBM, don't you think that it consumes too much power?

[Marvin Gollor]

I didn't state that (M1 Max) MacBooks feature 8 times faster memory, I just stated that it's connected to a 512 bit controller and therefore has 8 times higher bandwidth on paper than if conncted to 64 bit. In reality the bandwidth isn't always maxed out, also the case with x86 systems.

Morevover, I didn't say Intel is unable to make a wider mem-controller, I just stated they're just not using 64bit LPDDR5-6400-RAM. I guess 5200 is fast enough for intel for now, considering the additional costs. Yes, AMD doesn't use wider LPDDR5-6400-RAM on most machines to reduce costs or it's too difficult to justify the work, yet. Just because Apple finds a way for its high-end macbooks, doesn't mean AMD/Intel is able to copy Apple in this regard from a business standpoint.

For me the most interesting thing would be if the Steam Deck, which uses LPDDR5-5500 RAM with 32 bit per channel, reaches the advertised bandwidth of ~80-88 GB/s or if it stays below LPDDR5-6400-RAM at ~40-50 GB/s.

If the last is the case, AMDs mem controller really is BS but if the Steam Deck reaches ~80 GB/s, then AMD either uses a different mem controller for the steam deck or wider LPDDR5-RAM. Couldn't find any RAM benchmarks of a Steam Deck, though. We could derive it by the rdna-2-performance while knowing how this gpu scales with additional bandwidth, which is: ~20% more performance with ~30% added mem-bandwidth. And it turns out that the Steam Deck isn't faster in Games than an AMD 6800U+680M Laptop. It may indicate that the LPDDR5-5500 RAM is wider than LPDDR5-6400 on AMD Systems since rdna2 profits from the higher clocks of the last, which could means that the first could feature more bandwidth to compensate and reach same fps in games. But  could be to vague without actual mem-benchmarks.

AIDA says 128 bit but I am not necessarly trusting any software in this regard, as CPU-Z states wrong things as well. Would be extremly odd imho if the actual bandwidth would be 102 GB/s but we're reaching just the exact half of it. From what I read the mem-controller may be 128 bit but the memory itself just features 64 bit in total. But I am not sure at this, because I get too less info out of the net so far.

there is sitll the point, that regular DDR5-4800-RAM ist up to 30% faster in AMD machines than LPDDR5-6400-RAM which would be very odd, too, if both RAM is as wide as the other.

[NikoB]

I didn't state that (M1 Max) MacBooks feature 8 times faster memory, I just stated that it's connected to a 512 bit controller and therefore has 8 times higher bandwidth on paper than if conncted to 64 bit. In reality the bandwidth isn't always maxed out, also the case with x86 systems.

But still...You have already done your own work by finding this data and know the sources of this information. Could you provide links to this?

I guess 5200 is fast enough for intel for now, considering the additional costs.

I know for sure that it isn't. But they just can't seem to get better.

Just because Apple finds a way for its high-end macbooks, doesn't mean AMD/Intel is able to copy Apple in this regard from a business standpoint.

Don't you find your argument illogical and absurd considering the prices of expensive x86 solutions? Why is there no HBM memory? Or a 512 bit controller? I have already given an example of the new 7945HX, which is simply catastrophically short of memory bandwidth, at least 3 times (and this is purely bad based on my empirical experience and data).

AIDA says 128 bit

Thank you. Q.E.D. AMD has a shameful memory controller in Zen3+, I can't say anything about Zen4, but as I already noted, for some reason, the younger Phoenix "4nm" chip has 256GB of memory available, and the top 7945HX with 28 pci-e 5.0 lines has only 64GB.

Please explain!

In Gear1 mode, the frequency of the memory controller is equal to the frequency of the memory, i.e. 1:1 ratio in Cpu-Z. In Gear2 mode - the memory controller runs at half the memory frequency, the ratio is 1:2.

The main gain is not in memory bandwidth, but in memory latency. It speeds up by about 20-25% in atomic (no longer divisible into separate operations) operations. This has a significant effect in some software.

DDR4 works well in Gear1 up to 3733. But most laptop manufacturers deliberately include Gear2 because Gear1 consumes more. It doesn't matter from the outlet. But slightly increases the processor heat, enough to make the coolers noisier. For these reasons, Gear2 is used most often (except for workstations). On desktops, this is the first thing on which you can overclock the memory, with good stability.

As to HBM, don't you think that it consumes too much power?

Marvin assures that Macbooks use an analogue of HBM with a 512-bit bus. What do you think - is there a strong growth in consumption or not? And is the battery life of the laptop less or not?

[RobertJasiek]

Thanks for your Gear explanations!

About Mac, I know too little about unified memory to judge.

[Marvin Gollor]

But still...You have already done your own work by finding this data and know the sources of this information. Could you provide links to this?

Sure, we have the info in our database, too. It says "fast LPDDR5-6400 unified memory with a 512 bit bus (up to 400 GBit/s)." https://www.notebookcheck.net/Apple-M1-Max-32-Core-GPU-Benchmarks-and-Specs.579797.0.html

I'll link straight to a deeper analysis, which states that the full bandwidth could be reached if CPU, GPU and media engines stress the system at the same time. for cpu loads only the full bandwidth isn't used: "While 243GB/s is massive, and overshadows any other design in the industry, it's still quite far from the 409GB/s the chip is capable of. More importantly for the M1 Max, it's only slightly higher than the 204GB/s limit of the M1 Pro, so from a CPU-only workload perspective, it doesn't appear to make sense to get the Max if one is focused just on CPU bandwidth."
Source: https://www.anandtech.com/show/17024/apple-m1-max-performance-review/2

Wikipedia: "The M1 Pro has 256-bit LPDDR5 SDRAM memory, and the M1 Max has 512-bit LPDDR5 SDRAM memory. While the M1 SoC has 66.67GB/s memory bandwidth, the M1 Pro has 200GB/s bandwidth and the M1 Max has a 400GB/s bandwidth." (https://en.wikipedia.org/wiki/Apple_M1)

[NikoB]

So 400Gbit/s or 400Gbyte/s? If 400 Gbit/s, then this is a shameful 50Gbyte/s. And it turns out that Apple has slow memory and Intel is king in the market in terms of its speed with 80GByte/s (640Gbit/s) in Yoga Pro with 12700H and LPDDR5 5200 128 bit bus.

You have some data confusion.

[Marvin Gollor]

it's ofc GByte/s. Our data base has it wrong, wikipedia and the other source have it right. The original Apple screenshot, which is shown in the first link, also says GB/s, not GBit/s.

should be fixed in our database very soon, btw.