Review 'The Microsoft Surface Laptop 3 Showdown: Amd Picasso vs Intel Ice Lake' - Anandtech

[RetroZombie]

I expect Renoir to be quad core. Until the CCX grows to 8 in Zen 3, I don't see AMD going past 4 cores.

If Amd was smart, I expect an complete SOC like RR but with the ability to connect one chiplet to it.
They need to have high end APUs for the mobile, all-in-one systems and desktop.
If they really want to increase their apus margins, because with R3-3200G and R5-3400G (and respective mobile versions) they are going no where.

With 4C SOC + 1 chiplet they would get:

SOC with CCX0	CHIPLET CCX1	CHIPLET CCX2	TOTAL # CORES
4C	n/a	n/a	4C
4C	4C	0C	8C
4C	4C	4C	12C
3C	3C	0C	6C

With 8C SOC they would get:

SOC CCX0+1	CHIPLET CCX2	CHIPLET CCX3	TOTAL # CORES
4C+4C	n/a	n/a	8C
4C+4C	4C	0C	12C
4C+4C	4C	4C	16C
3C+3C	3C	3C	12C (alt)
2C+2C	2C	2C	8C (alt)
3C+3C	n/a	n/a	6C
2C+2C	n/a	n/a	4C

 

[moinmoin]

It's not terrible, but for the $ charged, it makes no sense to me in the context of the SP.

Branding-wise, I *hate* that they call non-Zen2 SKUs 3000 series such as the 3200G and this mobile series. It makes things murky and is a bite off Intel's playbook. Should have stuck with :

All 1000 series = Zen derived
All 2000 series = Zen+ derived
All 3000 series = Zen2 derived

Beh. Agree with Mark, things will get interesting with 7nm Zen2 mobile products. They're really giving Intel some breathing room here (in mobile) with this half baked last gen offering.

Keeping using the next gen Ryzen numbers for APUs certainly messes with the impression of the chip people have. But the APUs do use the microcode of the respective gen, even if the silicon design is still from the previous gen. AMD obviously thought this is enough to warrant branding the APUs as part of the next gens respectively.
Renoir potentially finally having an IMC capable of LPDDR may indicate the APUs turning into something own.

So my all point is, for battery test one suite is used, and for benchmarks another suite is used and then all the results are all screwed because the cpus behave differently in each of the suites scenarios!

For laptops/mobile devices all benchmarks ideally would also show the power consumption required for said results.

 

[RetroZombie]

I do know that the AMD version's battery life is shorter than the Intel from personal experience. But the Intel has hardware advantages and possibly lower power consumption.

I'm not denying that. But you understand that the anandtech (and others) didn't test the battery life with the applications which they tested the performance nor they did any performance/benchmarks tests with the laptop only on battery?

 

[LikeLinus]

I'm not denying that. But you understand that the anandtech (and others) didn't test the battery life with the applications which they tested the performance nor they did any performance/benchmarks tests with the laptop only on battery?

I've done some test personally and had a thread on it, but it became a cluster and was pointless.

That said, if you have to wonder why the sites aren't providing AMD testing for battery life, there may be a good reason for it.

AMD vs. Intel in the 15-Inch Surface Laptop 3: Both Models Benchmarked

Intel and AMD fight on the Surface battlefield.

www.tomshardware.com

The Surface Laptop 3 with the Ryzen 5 CPU ran for 8 hours flat. The Intel Core i7, however, lasted 9:31.

The AMD Ryzen 5 3580U (15w) loses against the i7 (15-25w). Pretty sure the higher clocked Ryzen 7 isn't going to increase battery life.

 

[UsandThem]

That said, if you have to wonder why the sites aren't providing AMD testing for battery life, there may be a good reason for it.

Both the Anandtech and Tom's review did test the battery life of both the units.

Anandtech uses a web test, and Tom's uses slightly different parameters (dimmer screen, OpenGL tests. etc).

Anandtech did more power tests with their original Ryzen 3 review, but they stated the battery life was "in the same range":

https://www.anandtech.com/show/15008/the-microsoft-surface-laptop-3-15-inch-review/6

Regardless of the tests (or which mobile Ryzen CPU they used), Intel is much more efficient with their mobile CPUs (especially with Ice Lake).

 

[LikeLinus]

Yes, both of them have done test with the Surface Laptop 3. That original Anandtech article is with the Ryzen 5 as well. It is odd that all the test are with the Ryzen 5 and not the 7.

I'm just saying, using the Ryzen 5 as a basis, I don't expect the 7 to be better. It's just odd that no reviews will publish actual Ryzen 7 numbers....

 

[UsandThem]

Yes, both of them have done test with the Surface Laptop 3. That original Anandtech article is with the Ryzen 5 as well. It is odd that all the test are with the Ryzen 5 and not the 7.

I'm just saying, using the Ryzen 5 as a basis, I don't expect the 7 to be better. It's just odd that no reviews will publish actual Ryzen 7 numbers....

Didn't Anandtech publish Ryzen 7 numbers?

I see the battery life and the minimum idle power draw listed for both the Intel Ice Lake and AMD Ryzen 7 "Picasso" listed.

The Tom's article you linked to does use a Ryzen 5 however (as did the original Anandtech Surface 3).

I guess I just don't know what's missing from the article I linked to in the OP?

 

[Thunder 57]

Two CCX would be pretty small if you cut it to 4 MB L3 each. Even if you end up using 6+ only for H parts it might be worth it.

Perhaps so, but I think a lot of the reason they were able to get away with such a small L3 compared to desktop/server models was because of the lack of the cross-CCX penalty. I'm not so sure two CCX's with 4MB of L3 each would perform as well as needed.

If Amd was smart, I expect an complete SOC like RR but with the ability to connect one chiplet to it.
They need to have high end APUs for the mobile, all-in-one systems and desktop.
If they really want to increase their apus margins, because with R3-3200G and R5-3400G (and respective mobile versions) they are going no where.

With 4C SOC + 1 chiplet they would get:

SOC with CCX0	CHIPLET CCX1	CHIPLET CCX2	TOTAL # CORES
4C	n/a	n/a	4C
4C	4C	0C	8C
4C	4C	4C	12C
3C	3C	0C	6C

With 8C SOC they would get:

SOC CCX0+1	CHIPLET CCX2	CHIPLET CCX3	TOTAL # CORES
4C+4C	n/a	n/a	8C
4C+4C	4C	0C	12C
4C+4C	4C	4C	16C
3C+3C	3C	3C	12C (alt)
2C+2C	2C	2C	8C (alt)
3C+3C	n/a	n/a	6C
2C+2C	n/a	n/a	4C

Sorry, you've lost me. I'm trying to picture what you posted but I'm not understanding just yet. How does the 4C SoC then one chiplet equal two chiplet CCX's? Could just be me, been a long day.

 

[jpiniero]

Perhaps so, but I think a lot of the reason they were able to get away with such a small L3 compared to desktop/server models was because of the lack of the cross-CCX penalty. I'm not so sure two CCX's with 4MB of L3 each would perform as well as needed.

16 total is an option too but doing 8 would save a decent amount of space.

 

[RetroZombie]

Sorry, you've lost me. I'm trying to picture what you posted but I'm not understanding just yet. How does the 4C SoC then one chiplet equal two chiplet CCX's? Could just be me, been a long day.

The 'first' CCX would be inside the SOC like it is right now on the Raven Ridge/Picasso. The other 'two' extra CCXs would be on the chiplet on the models with higher core count.
I would draw a design but it wouldn't be much different of what Matisse looks like (Ryzen 7 3700 for example).

Resuming what I'm trying to say they will keep the design has an single full SOC like before, but with the ability to add a chiplet to it.

 

[Atari2600]

However, their mobile Ryzen 'Picasso' CPUs have a ways to go.

Yep - to be fair to Intel they have a significantly superior product in the mobile space right now.

Zen2 *might* change that - but I doubt it'll bring enough power saving to the table to be superior.

 

[IntelUser2000]

Resuming what I'm trying to say they will keep the design has an single full SOC like before, but with the ability to add a chiplet to it.

That'll have its own problems. Zen 2 puts I/O die on the center, and the CPU core chiplets spread out from it like a tree branch. Meaning all cores more or less have the same latency.

The way you are suggesting would mean the main CCX would have the monolithic latency, while others will have to hop to that one and have different latencies.

They won't do it. It makes no sense.

 

[RetroZombie]

The way you are suggesting would mean the main CCX would have the monolithic latency, while others will have to hop to that one and have different latencies.

Yeah I thought on that too, but even with zen1 everything go through infinity fabric anyway, so the latency would't get worst because of that.
And remember with zen1 the many different latency's of the various designs (Ryzen/TR/Epyc) doesn't make it unworkable.

 

[IntelUser2000]

A more applicable sustained usage case would be something like video playback battery life which is also a test often done for mobile devices but not in this particular article.

Video playback is even lower power. Most of the compute is offloaded to the dedicated media playback units in the GPU, and the CPU can also burst in between frames. Unlike web browsing, video playback frame times are consistent and even easier to plan for.

(eg. I/O moves back to on die) could that result in Ryzen 4000 APUs being faster (at least on a per clock/per thread basis) vs Ryzen 3000 CPUs (at least for certain tasks like gaming)?

I don't think so? The same question was posed when Raven Ridge came out. Raven Ridge has 1 CCX so it doesn't need to jump.

- So the Intel chip at WALL POWERED is using up to 44 Watts and 'goes down' to 25 Watts when it gets 'hot', so all the BENCHMARKS results are a little flawed since they are against the amd part with only 15 Watts?

The AMD part is also bursting to higher frequencies, before it ramps down. Otherwise there's no way it can be even remotely competitive in the multi-threaded workloads.

https://www.notebookcheck.net/Surface-Laptop-3-15-Ryzen-5-Review-Microsoft-Can-Do-Better.440183.0.html#toc-energy-management

The above review shows the system power peaks but drops after running intensive tests such as Prime95. The Load battery tests also show the Ryzen Surface gets similar 100 min battery life meaning its using similar power in average.

Here's also a quote from the Anandtech review:

The Intel power numbers are for the SoC package, but the AMD power numbers appear to be just the CPU cores, which is an unfortunate byproduct of testing two different platforms.

Yeah I thought on that too, but even with zen1 everything go through infinity fabric anyway, so the latency would't get worst because of that.
And remember with zen1 the many different latency's of the various designs (Ryzen/TR/Epyc) doesn't make it unworkable.

I've heard of people talking about Infinity Fabric as its all the same thing. No its not. Its a protocol. Internal IF is different from CPU-CPU IF, and different from GPU IF and all are different from GPU to I/O IF.

There's also an important distinction between the latencies in the aforementioned CPUs versus what you are suggesting. The existing dies are set so they are mostly similar. If you want to make one die disproportionately faster than the others, that's very very different.

Imagine from a programmer's point of view. With Threadripper and Ryzen you can just code it so it stays within the CCX and not have to jump to the other one, so between CCX set it performs identically. What will they do with your configuration? One CCX will always be faster than others.

 

[RetroZombie]

The AMD part is also bursting to higher frequencies, before it ramps down. Otherwise there's no way it can be even remotely competitive in the multi-threaded workloads.

But on the intel system they got 63.5 Watts and with the amd system 52 Watts. This what they said about the new intel surface:
"This spike is higher than what we recorded on last year's Surface Pro 6 by over 20 W to suggest that the Surface Pro 7 is able to offer higher boost clock rates for short periods."

I've heard of people talking about Infinity Fabric as its all the same thing. No its not. Its a protocol. Internal IF is different from CPU-CPU IF, and different from GPU IF and all are different from GPU to I/O IF.

They will concur for the same integrated memory controller, latency will be the same, it's not much different of what already happens with the R9 3900X/3950X.

Imagine from a programmer's point of view. With Threadripper and Ryzen you can just code it so it stays within the CCX and not have to jump to the other one, so between CCX set it performs identically. What will they do with your configuration? One CCX will always be faster than others.

If Amd already have made the proper firmware adjustments, now it first 'fills' threads in the same CCX before jumping into another CCX, what you typed already fixed the problem

 

[DrMrLordX]

@IntelUser2000

One of the other possible problems with a Renoir die connected to an external CCX would be memory write behavior. With Matisse, a 1xCCD configuration produces 1/2 memory write bandwidth, while a 2xCCD configuration produces full memory write bandwidth (see 3700x vs 3900x memory write). We don't know what the Renoir memory controller will be like, but memory write performance in other Zen2 products links memory controller behavior to the number of linked CCDs . . . so it could introduce weird latency issues trying to hybridize Renoir and Matisse like that.

 

[mikk]

The AMD part is also bursting to higher frequencies, before it ramps down. Otherwise there's no way it can be even remotely competitive in the multi-threaded workloads.

https://www.notebookcheck.net/Surface-Laptop-3-15-Ryzen-5-Review-Microsoft-Can-Do-Better.440183.0.html#toc-energy-management

The above review shows the system power peaks but drops after running intensive tests such as Prime95. The Load battery tests also show the Ryzen Surface gets similar 100 min battery life meaning its using similar power in average.

I have to add that some tools like Hwinfo didn't recognize the power of AMD mobile Ryzen CPUs correctly and the recommendation is AMDuprof. I have seen several Ryzen notebook tests where HWinfo recorded 12W when in reality the device was configured for 35W. Pretty sure the 15W power graph from anandtech is plain wrong, it has nothing to do with "appear to be just the CPU cores"....it's pretty clear from the notebookcheck test.

 

[IntelUser2000]

But on the intel system they got 63.5 Watts and with the amd system 52 Watts. This what they said about the new intel surface:
"This spike is higher than what we recorded on last year's Surface Pro 6 by over 20 W to suggest that the Surface Pro 7 is able to offer higher boost clock rates for short periods."

The peak numbers don't matter when we're talking about battery life. If you go back to the Notebookcheck review and search for 1065G7 in the search bar below, you can add the Intel SL3 result. The Intel and AMD system uses nearly identical amount at 44W. Yes, the Intel system peaks higher but it happens so infrequently that the average is the same.

You can even see the "load average" numbers are different from battery life numbers. To get 100 mins using a 46WHr battery means its actually using only 28W. Unless you get inside the system and probe while on the battery, plugged in results will not accurately reflect battery life results.

They will concur for the same integrated memory controller, latency will be the same, it's not much different of what already happens with the R9 3900X/3950X.

@DrMrLordX has some interesting points.

I have seen several Ryzen notebook tests where HWinfo recorded 12W when in reality the device was configured for 35W. Pretty sure the 15W power graph from anandtech is plain wrong, it has nothing to do with "appear to be just the CPU cores"....it's pretty clear from the notebookcheck test.

Yea for monitoring the Core-based systems are top notch. Their Atom-based platforms don't seem accurate either.

 

[RetroZombie]

@DrMrLordX has some interesting points.

I'm not saying you or DrMrLordX are wrong, I agree 100%. But what you guys say isn't reasons that makes it impossible to implement.
For example in the case of the 3900X and 3950X because of having two chiplets with the memory write, did any test showed any weird performance behavior because of that? (outside of synthetics tests).

Besides Amd needs much better apus, even intel 'apus' are already at eight cores for (ultra) high end mobile/desktop chips. Amd the company with the much better gpu sells us low end to mid range chips.

 

[Gideon]

Comparing the cost of adding an additional CCX on the die vs what additional markets that same CPU can address, it would be very strange to leave Renoir at 4 Cores. Especially as power draw at optimal frequency is the biggest improvement of 7nm over 12nm (almost 2x improvment). AMD could do 6 cores at 15W and 8 cores at 25W.

Especially considering that Intel can put a 8-cores and an IGP (on a much bigger 14nm die) and sell it in almost all kinds of laptops.

With 6-8 cores they could address high-end laptops, 99% of OEM Desktops, etc ... And as with raven-ridge another 4core Die could appear later.

 

[DrMrLordX]

For example in the case of the 3900X and 3950X because of having two chiplets with the memory write, did any test showed any weird performance behavior because of that? (outside of synthetics tests).

I don't think you grasp exactly what may be the issue here. In the case of the 3900x and 3950x, the two-chiplet solutions simply doubled the memory write speed of the CPU vs. single-chiplet Matisse running the same memory speed/timings.

Have you ever stopped to wonder why the memory controller on Matisse works that way? What if the memory controller on Renoir doesn't work that way?

 

[majord]

Yep - to be fair to Intel they have a significantly superior product in the mobile space right now.

Zen2 *might* change that - but I doubt it'll bring enough power saving to the table to be superior.

1.7w seperates the two during the anandtech battery test. That's all out takes to post crap run time numbers on these light use tests

 

[IntelUser2000]

I'm not saying you or DrMrLordX are wrong, I agree 100%. But what you guys say isn't reasons that makes it impossible to implement.
For example in the case of the 3900X and 3950X because of having two chiplets with the memory write, did any test showed any weird performance behavior because of that? (outside of synthetics tests).

Since its a redesign of the layout, if anything I expect them to have a 6 or 8 core monolithic die.

 

[uzzi38]

Yep - to be fair to Intel they have a significantly superior product in the mobile space right now.

Zen2 *might* change that - but I doubt it'll bring enough power saving to the table to be superior.

Oh deary me how wrong you are.

AMD loses in battery life tests because of two things.

Number 1: Idle power draw is notably higher.
Number 2: Race to idle. Picasso is straight up inferior to Whiskey, Comet and Ice Lake in terms of raw performance, so when something needs to be done, it spends more time in a higher power state, leading to greater overall power draw.

Renoir fixes both of these, and more. And I'm like 90% sure there will be an increase in core count too.

I've said it before, I'll say it again. Tiger Lake is the only thing that can compete, and even then it'll do so in ST and iGPU workloads only. Not like the latter matters when AMD's throwing out near free Navi 14's to any mobile OEM that wants GPU performance.

 

[uzzi38]

Actually, I'll just leave this here:



This is from a document taken from AMD's site (though it's been taken down now, it was found by Komachi). Page 553.

product-master.pdf

drive.google.com

Personally, I don't think AMD are dumb enough to even consider an r9 sku without being able to offer more cores. Does anyone want to suggest otherwise?

Note: this document is actually form September, by the time it was found, it was out of date. This can be told by the 16 core Threadrippers mentionned in the spec sheet, which are no longer planned to hit the market.