Chrispy_My biggest worry is that too much of the power and silicon budget has been moved away from the graphics cores and to the CPU cores. Both previous generations of APU have been perfectly adequate in the CPU department and sorely lacking in the IGP department, ranging from inadequate (Vega3 being as useless as low-end Intel HD) to acceptable-but-underwhelming - in that 720p30 might just be attainable in current titles.

In terms of CPU performance, 15W ultraportable customers really aren't clamouring for more cores. Those products are typically not multitasking mobile workstation powerhouses, lacking RAM, screen size, and storage for many serious workloads. To date, the vast majority of 3700U flagship APUs have been either 8GB or 16GB max, with 512GB NVMe 2x and 1080p displays. They are general-purpose consumption devices that could definitely use a little more GPU power but are typically beyond the point of diminishing returns in the CPU department already.

AMD keep saying that they have added 60% more performance per CU but then have stripped away 30-40% of the CUs The fact that AMD isn't singing praises about the 3D performance of its new 7nm APUs is a pretty bad sign of things to come, especially because the 4800U is likely to command a significant price premium. I would expect the 6CU option to sell at the same pricing tier as the previous 10CU 3700U models, which sucks because the 4800U will likely be price-competing against an MX250 or even faster dGPU options. I can only hope that LPDDR4X is used to its full advantage this generation. My 2700U shipped with a 13W TDP and single-channel DDR4 2400; The only reason I bought it was good cooling (for up to 25W) and an empty DIMM slot for dual-channel.

Is it too much to ask for a 15W ultraportable that provides a meaningful IGP upgrade? The bar is SO low, I'm honestly saddened by the lack of attempts to pass it.

I think you are not taking it the right way. They have stripped the CU's and it's 15W, but it doesn't mean this is the only one AMD is releasing. There might be more. On the other hand there is no need for more in this segment. It is a 15-25W laptop segment and stop thinking that there's only games that matter and you are worried that you won't be able to play games. You will but this laptop is not for playing games 2k 100FPS ultra.
Either way its performance is really good despite 15W. That is what matters at this point. If there's more CPUs for laptop 15-25W segment to come we will see.
Lisa Su on the CES keynote said. AMD brings "The Best mobile CPU" the 4000 Series. Let us see what it can do when it hits the market but for Christ don't expect to get discrete card like performance.

londistePretty sure that is not true. See @seronx post #16. The same "not tied" messaging has been used throughout AMD statements for Zen2. What AMD means by that phrasing is that IF frequency is not 1:1 to memory speed as it was for Zen/Zen+. They added a multiplier/divider and use that to facilitate faster memory speed.

Actually it is true. It says in the annandtech's article clearly that the:

"These CPUs all support LPDDR4X memory, up to 64 GB, and AMD says that the infinity fabric is not tied to this memory clock. This helps the chip reach even lower power in its idle states, and the company said that they have rearchitected a good portion of the power delivery in the APU in order to be able to power down and power gate more elements of the SoC than was previously possible"

Unless you doubt what the article says.

Pretty impressive when you compare it to the ~5500 and 31500 points of the 45W i9-10980HK.

londistePretty sure that is not true. See @seronx post #16. The same "not tied" messaging has been used throughout AMD statements for Zen2. What AMD means by that phrasing is that IF frequency is not 1:1 to memory speed as it was for Zen/Zen+. They added a multiplier/divider and use that to facilitate faster memory speed.

I sincerely doubt that. AnandTech specifically states that

AnandTechAMD says that the infinity fabric is not tied to this memory clock.

"Not tied to" is something quite different than "tied to, but variable based on a clock divider". "Not tied to" means no direct relation between. Otherwise AT is either misquoting AMD, or AMD is being more shady than usual with their wording. Neither sound likely. Also, @seronx's numbers are a bit off. How does 4266 with a 2:1 divider become 1600? Last I checked 4266:2 was 2133. so you'd need a 3:1 divider to keep IF clocks in a normally stable range with 4266MT/s RAM.

Chrispy_My biggest worry is that too much of the power and silicon budget has been moved away from the graphics cores and to the CPU cores. Both previous generations of APU have been perfectly adequate in the CPU department and sorely lacking in the IGP department, ranging from inadequate (Vega3 being as useless as low-end Intel HD) to acceptable-but-underwhelming - in that 720p30 might just be attainable in current titles.

In terms of CPU performance, 15W ultraportable customers really aren't clamouring for more cores. Those products are typically not multitasking mobile workstation powerhouses, lacking RAM, screen size, and storage for many serious workloads. To date, the vast majority of 3700U flagship APUs have been either 8GB or 16GB max, with 512GB NVMe 2x and 1080p displays. They are general-purpose consumption devices that could definitely use a little more GPU power but are typically beyond the point of diminishing returns in the CPU department already.

AMD keep saying that they have added 60% more performance per CU but then have stripped away 30-40% of the CUs The fact that AMD isn't singing praises about the 3D performance of its new 7nm APUs is a pretty bad sign of things to come, especially because the 4800U is likely to command a significant price premium. I would expect the 6CU option to sell at the same pricing tier as the previous 10CU 3700U models, which sucks because the 4800U will likely be price-competing against an MX250 or even faster dGPU options. I can only hope that LPDDR4X is used to its full advantage this generation. My 2700U shipped with a 13W TDP and single-channel DDR4 2400; The only reason I bought it was good cooling (for up to 25W) and an empty DIMM slot for dual-channel.

Is it too much to ask for a 15W ultraportable that provides a meaningful IGP upgrade? The bar is SO low, I'm honestly saddened by the lack of attempts to pass it.

I'm not worried at all about this. One of AMD's biggest points when presenting these is their new system for balancing power distribution between the CPU and iGPU (as well as any AMD dGPU installed in the system) using what they call "Infinity control fabric" (yeah, not very imaginative :P ). They're claiming significant performance increases due to this as power is supposed to be dynamically divided based on actual demands in real time rather than preset power levels. While this is mainly marketed for H-series chips, the silicon is the same, and the main difference between H and U in this regard is just that H-series chips will mostly have dGPUs attached and thus benefit more from the system also encompassing that. (Also, the lack of dGPU means thin-and-lights don't need the SmartShift setup to dynamically manage CPU/iGPU power.) They're still touting significant increases in mixed workloads like gaming despite there now being 8 CPU cores. Given the work AMD has put into improving the Windows scheduler for optimal function with desktop Ryzen I'm guessing they've brought that work with them so that the system keeps as few cores active as possible while gaming (and in similar loads) so that they can clock high while the rest sleep, freeing power for the GPU. If games suddenly start hammering 10-16 threads I guess the CPU might suffocate the GPU performance, but then you're talking about a game outside the capabilities of a system like this anyhow.

Given advertised performance levels (and leaked benchmarks) the perf/CU increase looks about right, meaning that we'll be getting overall gaming performance increases for all AMD APUs compared to previous generations. LPDDR4X is also going to be used in a lot of designs, especially thin-and-lights.

Still, the next generation, with RDNA on board, is likely to kick this one's butt. That's a given, especially if it's on 7nm+. But we'll still be getting a decent enough performance increase this go around.

Nonetheless, the 4800U is looking like it'll be both a beast for heavy CPU tasks like video editing or rendering for its class of laptop, while also being the 15W king of gaming. If it indeed delivers ~28% more performance than an Ice Lake 1056G7 like they say (normally I'd be wary of pointing to a canned benchmark, but in this case that's one of Intel's best-case scenarios in terms of optimization), that's a notable jump. According to NotebookCheck the 1065G7 on average scores ~967 in 3DMark Time Spy, so a 28% increase over that means ~1230 points. The 3700U scores about the same as the 1065G7.

ValantarI sincerely doubt that. AnandTech specifically states that
"Not tied to" is something quite different than "tied to, but variable based on a clock divider". "Not tied to" means no direct relation between. Otherwise AT is either misquoting AMD, or AMD is being more shady than usual with their wording. Neither sound likely.

I remembered incorrectly - AMD's wording was "decoupled". Which is not that much different.

ValantarI'm not worried at all about this. One of AMD's biggest points when presenting these is their new system for balancing power distribution between the CPU and iGPU (as well as any AMD dGPU installed in the system) using what they call "Infinity control fabric" (yeah, not very imaginative :P ). ... difference between H and U in this regard is just that H-series chips will mostly have dGPUs attached and thus benefit more from the system also encompassing that. (Also, the lack of dGPU means thin-and-lights don't need the SmartShift setup to dynamically manage CPU/iGPU power.)

From much of what I have seen, Zen/Zen+ APUs seem to err on the side of giving CPU more power than it needs and occasionally starving iGPU due to that. When they are able to balance the power better to how it is needed, iGPU might get a good boost out of that.

Geekbench is all over the place. A bit difficult to come to a conclusion when your scores range from 5800 to 1800 on the same chip.

seronxIt uses a low-power double-data rate four x(LPDDR4X) interface for 4266 MHz.
Fabric clock in Renoir is probably forced to asymmetrical 2:1, rather than 1:1.

.dram_speed_mts = 1600.0 => .fabricclk_mhz = 600.0 // DRAM @ 1600 MHz = IF @ 600 MHz
^-- probably a weird idle.
.dram_speed_mts = 2133.0 => .fabricclk_mhz = 1066.0 // DRAM @ 2133 MHz = IF @ 1066 MHz
.dram_speed_mts = 4266.0 => .fabricclk_mhz = 1600.0 // DRAM @ 4266 MHz = IF @ 1600 MHz
^-- If it was 3200 MHz, fabric clock would probably be also 1600 MHz.

Iiirc, it is not connected. These APUs are completely reworked. Why it's not connected is to reduce power consumption.

Doh, nvm since this has been pointed out above.

These would be especially fantastic CPUs if the TDP is user-adjustable. Something like a 45w power budget when plugged in and something lower when on-the-move would be the best of all worlds.

I suspect a lot of premium laptops will feature these CPUs.

You guys realise apu will be competing for the same LPDDR4X resource, right? Btw, that is a 16-bit datapath memory. Got to save on the muxes.

Considering that monolithic die has some serious power limits and thread handicaps compared to a 3700x...

It's kind of impressive that in single core it's only 20% slower in single threads when it's got a 50W TDP limit, and we aren't sure if the laptop APUs will basically clock and suck power until the cooling limits it like the desktop 3700x.

Give me a desktop version and just like I suspect the monolithic design will be faster.

mtcn77You guys realise apu will be competing for the same LPDDR4X resource, right? Btw, that is a 16-bit datapath memory. Got to save on the muxes.

16-bit, but four channels. Equal(-ish) to two 32-bit DDR4 channels, just significantly higher bandwidth at lower power, but also at the cost of latency. Chips paired with LPDDR4X will have significantly better graphics performance as that is largely bandwidth limited. Possibly slower in latency sensitive workloads, but which consumer workloads are truly RAM latency sensitive?

HugsNotDrugsThese would be especially fantastic CPUs if the TDP is user-adjustable. Something like a 45w power budget when plugged in and something lower when on-the-move would be the best of all worlds.

I suspect a lot of premium laptops will feature these CPUs.

... That's not how laptop designs work, nor how chip SKUs work. These are 15W chips, OEM-configurable to 25W if they choose to build a chassis to accommodate that amount of heat output. There are separate 45W SKUs with different binning and specs. For a laptop to have a 45W mode it would need the ability to dissipate 45W of heat energy and deliver a continuous 45W to the APU, which means it would be rather large. This bars it from being very thin and light, so the vast majority of laptop makers then go all in on performance instead (though you do get premium devices trying to deliver the best of both worlds).

While what you're saying is tempting on paper, the reality is that any thin-and-light built for portability would thermal throttle massively if the power limit was raised even halfway to 45W. Cooling is built to spec, and requires a lot of space, so making anything thin and light requires scaling cooling to fit a low power processor.

On the other hand, laptops do kind of do what you are saying already: most laptops limit max CPU and GPU power when running off battery below AC power levels. It's just that the ones with 45W CPUs are still large and still have disappointing battery life despite this. You pretty much need to build either for battery life or for high performance. Attempting to combine the two rarely works well.

gamefoo21Considering that monolithic die has some serious power limits and thread handicaps compared to a 3700x...

It's kind of impressive that in single core it's only 20% slower in single threads when it's got a 50W TDP limit, and we aren't sure if the laptop APUs will basically clock and suck power until the cooling limits it like the desktop 3700x.

Give me a desktop version and just like I suspect the monolithic design will be faster.

We'll see once the next round of desktop APUs roll out. It's possible, but the lower L3 cache might be a bit of a hindrance. Nonetheless I'm planning on getting one (though likely not a ...4800G? as it'll likely be far more than I need for my HTPC). We still need to remember that power doesn't scale linearly with clock speed increases, and laptops tend to boost above TDP for limited periods of time.

mtcn77You guys realise apu will be competing for the same LPDDR4X resource, right? Btw, that is a 16-bit datapath memory. Got to save on the muxes.

Valantar16-bit, but four channels. Equal(-ish) to two 32-bit DDR4 channels, just significantly higher bandwidth at lower power, but also at the cost of latency. Chips paired with LPDDR4X will have significantly better graphics performance as that is largely bandwidth limited.

AMD's Renoir uses 128-bit LPDDR4X with four 32-bit datapaths.
www.samsung.com/semiconductor/dram/lpddr4x/
2 options with 96 Gb x64 packages(2 chips w/ Renoir = 24 GB)
1 option with 80 Gb x64 packages(2 chips w/ Renoir = 20 GB)
4 options with 64 Gb x64 packages(2 chips w/ Renoir = 16 GB)
2 options with 64 Gb x32 packages(4 chips w/ Renoir = 32 GB)

128-bit LPDDR4X @ 4266 MHz = 256-bit DDR3/DDR4 @ 2133 MHz

seronxAMD's Renoir uses 128-bit LPDDR4X with four 32-bit datapaths.

You had me scared, there. Ram is still 16-bit dual-rank operated. Am I right, or am I right?

mtcn77You had me scared, there. Ram is still 16-bit dual-rank operated. Am I right, or am I right?

The channels are 16-bit wide, but the datapaths in Renoir are in 32-bit width.

128-bit => eight 16-bit channels of LPDDR4X if by specification.
128-bit => four 32-bit PHYs of LPDDR4X if by design.

Which equates to 2x 64-bit packages(four 16-bit channels per package) or 4x 32-bit packages(two 16-bit channels per package).

seronxThe channels are 16-bit wide, but the datapaths in Renoir are in 32-bit width.

128-bit => eight 16-bit channels of LPDDR4X if by specification.
128-bit => four 32-bit PHYs of LPDDR4X if by design.

Your facts cast terror to my self-confidence, at times.
I didn't suggest any inefficiency, to the contrary, I think this is the way for all future datapath architectures. Even AVX-512 instruction set MUX'es don't run natively, eventhough Intel went above and beyond in building the biggest and most efficient vias in its 14nm backend. No one can question Intel on that point. The surge is too rampant to feed.

mtcn77Your facts cast terror to my self-confidence, at times.
I didn't suggest any inefficiency, to the contrary, I think this is the way for all future datapath architectures. Even AVX-512 instruction set MUX'es don't run natively, eventhough Intel went above and beyond in building the biggest and most efficient vias in its 14nm backend. No one can question Intel on that point. The surge is too rampant to feed.

It is the same with Icelake:


4 channels w/ 32-bit width. I give this 5 out of 7 shrugs.

AVX512 at Intel is 2x256-bit w/ 1x512-bit. One unit runs full native(1x512-bit), the other two run AMD native(2x256-bit = 1x512-bit).

seronxIt is the same with Icelake:


4 channels w/ 32-bit width. I give this 5 out of 7 shrugs.

AVX512 at Intel is 2x256-bit w/ 1x512-bit. One unit runs full native, the other two run AMD native.

Icelake is a two bit cpu in so far as succeeding comet lake with its top notch huge power pillars that still cannot steamroll the MUX power surge without a clock stretch.

How have Intel given into this predicament? Running 512 is like nesting the big loop in a small 32 loop. It would take several ram cycles to feed the on die cache at speed. Good thing I'm not a weather forecaster, it would always go moody.

I see what you did there. XD
anyways, if the 4700U is performing this good with assuming a 15W TDP instead of 25W (as per spec from AMD to their OEM partners), then that's impressive for me. Imagine giving the 4800U a 25W TDP envelope... that I think will scare Intel's 10th gen U series mobile CPUs in a way that makes end users like us, encouraged to get an AMD powered slim laptop. Unless, Intel based laptops have MX3xx GPUs to make up for it..

Tsukiyomi91I see what you did there. XD
anyways, if the 4700U is performing this good with assuming a 15W TDP instead of 25W (as per spec from AMD to their OEM partners), then that's impressive for me. Imagine giving the 4800U a 25W TDP envelope... that I think will scare Intel's 10th gen U series mobile CPUs in a way that makes end users like us, encouraged to get an AMD powered slim laptop. Unless, Intel based laptops have MX3xx GPUs to make up for it..

The MX350 looks like it'll be interesting, likely a stripped-down GTX 1050 chip rather than yet another MX 150 refresh. A necessity, given that the advertised 28% performance boost of the 4800U over the i7-1065G7 indicates pretty much identical performance to the MX250. That's pretty darn impressive given that you're looking at nearly 1/3 the CPU+GPU power envelope.

The future of iGPUs looks pretty damn exciting right now. First Ice Lake saw Intel crawl out of the gutter, now this punches back, then Intel will bring out their Xe-based chips later this year, and then AMD will have 5000-series APUs with an RDNA-based iGPU (and possibly DDR5?). Things are looking up for sure. Could we get passable AAA 1080p high/ultra performance from a 15W APU in a couple of years?

ValantarCould we get passable AAA 1080p high/ultra performance from a 15W APU in a couple of years?

Oh that would be something :) Although considering what's happening in the APU's lately and what's to come this idea 1080p AAA games High/ultra, doesn't seem so crazy now.

Valantar... That's not how laptop designs work, nor how chip SKUs work. These are 15W chips, OEM-configurable to 25W if they choose to build a chassis to accommodate that amount of heat output. There are separate 45W SKUs with different binning and specs. For a laptop to have a 45W mode it would need the ability to dissipate 45W of heat energy and deliver a continuous 45W to the APU, which means it would be rather large. This bars it from being very thin and light, so the vast majority of laptop makers then go all in on performance instead (though you do get premium devices trying to deliver the best of both worlds).

While what you're saying is tempting on paper, the reality is that any thin-and-light built for portability would thermal throttle massively if the power limit was raised even halfway to 45W. Cooling is built to spec, and requires a lot of space, so making anything thin and light requires scaling cooling to fit a low power processor.

On the other hand, laptops do kind of do what you are saying already: most laptops limit max CPU and GPU power when running off battery below AC power levels. It's just that the ones with 45W CPUs are still large and still have disappointing battery life despite this. You pretty much need to build either for battery life or for high performance. Attempting to combine the two rarely works well.

We'll see once the next round of desktop APUs roll out. It's possible, but the lower L3 cache might be a bit of a hindrance. Nonetheless I'm planning on getting one (though likely not a ...4800G? as it'll likely be far more than I need for my HTPC). We still need to remember that power doesn't scale linearly with clock speed increases, and laptops tend to boost above TDP for limited periods of time.

The power limit is set by the motherboard, not the CPU. Users of the ThinkPad T480s and some Carbons can set TDP to 45w which provides a substantial increase in performance. Cooling, especially on the T480s without the dedicated GPU, is excellent while the laptop remains relatively slim.

Through university I ran an Asus UL30VT which was a SU7300 1.3ghz C2D that would factory overclock to 1.7ghz when plugged in (and further user overclock to 2.1ghz). It was the best of all worlds of portability and performance.

HugsNotDrugsThe power limit is set by the motherboard, not the CPU. Users of the ThinkPad T480s and some Carbons can set TDP to 45w which provides a substantial increase in performance. Cooling, especially on the T480s without the dedicated GPU, is excellent while the laptop remains relatively slim.

Through university I ran an Asus UL30VT which was a SU7300 1.3ghz C2D that would factory overclock to 1.7ghz when plugged in (and further user overclock to 2.1ghz). It was the best of all worlds of portability and performance.

OEMs do have the option to configure their own thermal limits (to some degree at least), with cTDPs being the most common. The latest generations of Thinkpad X1 Carbons have used 25W and 29W cTDP-up configurations to increase sustained performance. That of course also means they have overbuilt coolers to match. But they can't sustain that performance over time, unlike something built for an H-series CPU. To quote NotebookCheck's review of the 2018 X1 Carbon:

NotebookCheckWith last year's X1 Carbon, Lenovo raised the permanent thermal design power (TDP) to 25 W, while setting a rather conservative 75 °C temperature limit. This year, Lenovo has further increased the X1 Carbon's TDP and temperature limits, allowing the processor to boost up to 29 W for 28 seconds. This corresponds with a clock speed of between 3.1 and 3.2 GHz across all four cores. After this time, or once the laptop has reached 90 °C, the processor drops to 23 W and throttles down to 2.8 GHz across all cores.

With Intel having the XTU tool for end users to further configure things there's likely some room for further increases in short-term power (particularly with repasting and other optimizatoins to cooling), but they can't sustain 45W indefinitely. But yes, if you sacrifice some volume (like the 480s, which isn't particularly compact and has a cooling system built for adding a 25W GPU) there's of course room for higher power limits. But at that point, why not go for a H-series SKU which is binned and configured for sustained high performance, and simply limit it when battery powered (as Windows does anyway)? I absolutely agree that thin-ish-and-light-ish laptops with H-series APUs would be a very, very attractive proposition (particularly if they made an H-series APU with a fully enabled and high-clocking iGPU), but sadly it seems OEMs don't see things the same way. As such, I'd settle for a good 25W cTDP-up laptop which will likely deliver >90% of the performance of a 45W one in real world use.

On Ryzen 4000U FCLK works 2:1 with 3200Mhz RAM and 1:1 with 2666Mhz RAM. So:
RAM 3200Mhz - FCLK 800Mhz
RAM 2666Mhz - FCLK 1333Mhz

persiziOn Ryzen 4000U FCLK works 2:1 with 3200Mhz RAM and 1:1 with 2666Mhz RAM. So:
RAM 3200Mhz - FCLK 800Mhz
RAM 2666Mhz - FCLK 1333Mhz

Interesting. I wonder what frequency Renoir swaps from 1:1 to 1:2, and I'm also wondering if LPDDR4X-4266 drops the FCLK the same way as regular DDR4....

According to a Chinese review:

When you think that there is a frequency division of only 4266MHz LPDDR4x, AMD secretly penetrates the production version of the frequency division into all low-voltage U series, as long as the memory frequency is greater than 2666MHz, this frequency division will exist, surprise, surprise, unexpected ? !
But in order to prevent consumers from perceiving, on the production version of the CPU, AMD launched a new operation!
First, let’s talk about science. The frequency of the IF bus is called Infinity Fabric Clock (FCLK). Later, it is called FCLK.

AMD divides FCLK into multiple operating frequencies (800-1333MHz), and then automatically changes the frequency according to the needs of the scene, so that it is completely decoupled from the memory frequency and has been in a state of frequency division.

However, the standard H processor is still in a 1:1 running state (using DDR4-3200), so FCLK can maintain 1600MHz without frequency division.

AMD Ruilong 4000 Series Test & Analysis