AMD Ryzen 9 7940HS analysis - Zen4 Phoenix is ideally as efficient as Apple

[locksleee]

I don't get the reported idle power consumption numbers.

On my 7940HS flow x13 (received 3 days ago), the CPU power package (as reported by HWinfo) drops below 1W at idle.

Is the 3W reported here a G14 specific issue?

When you're measuring the 1W, is your laptop plugged in or running on battery power?

On my G14 I'm seeing 3W as well and the CPU bottoms out at about 3200MHz.  But if I unplug and enable Battery Saver mode in Windows, the CPU speed drops to 1600MHz and pulls 0.8-1.2W while idle, so this is how the G14 can get its 7940 down to those lower power levels, it's just not something that's possible to enable when the laptop is plugged in.

[tipoo]

Well this is quite whelming tbh. I suppose one could have guessed at this, since TSMC N4 is not really a new node, but a derivative of N5, where Apple again will likely get the first run of 3nm, the true new gen node. 

It looks like Intel may have a very good shot here with Meteor Lake also, the 128 Xe core Tile GPU at 4Tflops would probably be close to the 680M, and I can surmise that like the desktop side of things, the 780M doubling the tflops on paper didn't amount to much performance uplift, just counted different.

[ArsLoginName]

I can't believe these numbers. It's the biggest letdown of 2023. Going from TSCM 7 to TSCM 4 gives absolutely nothing. 190 points/W for 4nm versus 182 points/W for 7nm. I was waiting for Zen 4 notebooks, but it's not worth it.

Efficiency does appear to be underwhelming in the manner the results are presented. But the results are in line with Zen 4 IPC gains from AMD's presentation of 11% in CB R23 coupled with TSMC's N4 to N7 guidance on process node of 15-18% for a total improvement of 25-30%.

Efficiency is not really a plot of CBR23 score vs processor TDP. That only shows the trade-off in performance vs power. Total power consumption (W) to do a task is a much better metric to show efficiency. Since CB score is based on how many frames can be rendered in a fixed quantity of time, all you have to do is some simple ratios to analyze the 'true' efficiency gains.

Let's start with the 80 W TDP with total power consumption of 108 W and a CB R23 score of 18044 pts. Since score = # frames/time, the fixed quantity of work (# of frames) = time*CBR23 score. Since the highest # of frames is generated at the highest power settings, everything can be ratioed to these values (total power consumption of 108 W for 'z' frames and 18044 pts). When the processor is set to a TDP of 35 W, it should take (18044/13723) = 1.315 longer time to generate the same number of frames during which the system consumes 1.315*54 W = 71.0 W of total power instead of 108 W. Over the power range investigated for the 7940HS for this task, 71.0 W is the least amount of power consumed but at the trade-off of the longest time to complete the task. So it takes 52.1% more power (108 W/ 71 W) to save 31.5% in time.

Doing the same at 45 W TDP for the 7940HS (15625 pts, 66 W consumption) yields 1.15x longer time with a total power consumption of 76.2 W (41.7% more power to save 15% in time compared to 80 W settings). When compared to the 35 W values, the 45 W setting is consumes only 7.3% more power (76.2 W/71 W) while gaining back over 50% of the 'extra' time it took to complete the task.

Similarly 55 W (1.086x extra time, 85.8 W total) gains a little bit more than 75% of the extra time back for 21% more consumed total power (85.8 W/71 W) compared to 35 W. Compared to 80 W, 55 W saves you 20.5% power for only 8.6% longer in time. But it is all just a trade off of total energy vs time over this power range.

The 28 W 7840U processors will have a higher efficiency due to the 20% reduction in power compared to the 35 W settings for a 5-10% reduction in clock speeds/CBR 23 scores. But leak says 14789 pts so they may be on a slightly different N4 process than these higher power H/HS parts. Or just better binned. Maybe the HS parts are 'worse' binned than U and H are worse binned than HS.

Let's look at the N6 based 7735HS/6900HS scores of about 11200 and 12200 pts at 35 W and 45 W respectively. Assuming similar total power consumptions for RAM since both DDR5, similar refresh rates, SSDs, and taking off an additional 1 W for the difference between the mini-LED screen vs LCD, this would give a total energy consumption of 85.4 W and 96.1 W for the 35 W and 45 W 7735HS/6900HS respectively. These powers are 20.3% and 25.8% higher than the 7940HS at equivalent TDPs. These increases in power also correspond to time increases of 1.61x and 1.479x compared to the 80 W 7940HS and are substantially longer than the 7940HS at same TDPs. A 45 W 7735HS/6900HS consumes almost exactly the same total power as the 65 W 7940HS (94.4 W) to complete this task but the 7940HS completes the task 30% faster.

So yes. The 7940HS is more efficient than the 7735/6900HS. But the mini LED screen really seems to kill the idle power. Wait for reviews with standard LCDs to make final judgements.

[Abc]

Thanks for doing the power limit test. Efficiency is tied to how high of frequency you run your processor (the faster above the sweet spot, the less efficient), so it makes no sense to use the default limits. We all know too well many manufacturers will sacrifice 50% power draw for a 5% benchmark gain (and Intel and AMD happily let them).

In fact, a lot of efficiency has to do with how much power each core gets (which determines how high of frequency the core can reach). Which is why Intel U and P series processors have different efficiency curves because they have different core counts, with the U processor reaching its sweet spot at lower power levels.

There is still a gap between AMD and Apple at similar power levels. I think it is mostly due to architectural differences. X86 needs to move to 64 bit only like ARM has, to get rid of useless silicon dedicated to 32 and 16 bit code which leaks current.

The gap between Intel and AMD has a lot to do with TSMC's 7nm and 5nm process being so good. We see that from Exynos and Snapdragon as well that Intel and Samsung foundry is simply lagging behind TSMC in leakage, which translates to higher power draw and heat production. If Intel used TSMC for some models we should see the gap close considerably.

[Crear]

I sincerely wish it's just a flaw of ROG G14 only...I respect the power efficiency that M2 Pro reached, but my work flow requires x64 instead of ARM64 (well done, Nvidia SDK Manager). Still waiting for the Razer Blade 14 2023. Hopefully the idle power can be optimized, along with 32GB DDR5 and 16:10. Mercury edition will be a bonus. If not the lack of CUDA and x64 support, I would most likely just get a 14-inch MacBook Pro.

[S.Yu]

I can't believe these numbers. It's the biggest letdown of 2023. Going from TSCM 7 to TSCM 4 gives absolutely nothing. 190 points/W for 4nm versus 182 points/W for 7nm. I was waiting for Zen 4 notebooks, but it's not worth it.

Efficiency does appear to be underwhelming in the manner the results are presented. But the results are in line with Zen 4 IPC gains from AMD's presentation of 11% in CB R23 coupled with TSMC's N4 to N7 guidance on process node of 15-18% for a total improvement of 25-30%.

Efficiency is not really a plot of CBR23 score vs processor TDP. That only shows the trade-off in performance vs power. Total power consumption (W) to do a task is a much better metric to show efficiency. Since CB score is based on how many frames can be rendered in a fixed quantity of time, all you have to do is some simple ratios to analyze the 'true' efficiency gains.

Let's start with the 80 W TDP with total power consumption of 108 W and a CB R23 score of 18044 pts. Since score = # frames/time, the fixed quantity of work (# of frames) = time*CBR23 score. Since the highest # of frames is generated at the highest power settings, everything can be ratioed to these values (total power consumption of 108 W for 'z' frames and 18044 pts). When the processor is set to a TDP of 35 W, it should take (18044/13723) = 1.315 longer time to generate the same number of frames during which the system consumes 1.315*54 W = 71.0 W of total power instead of 108 W. Over the power range investigated for the 7940HS for this task, 71.0 W is the least amount of power consumed but at the trade-off of the longest time to complete the task. So it takes 52.1% more power (108 W/ 71 W) to save 31.5% in time.

Doing the same at 45 W TDP for the 7940HS (15625 pts, 66 W consumption) yields 1.15x longer time with a total power consumption of 76.2 W (41.7% more power to save 15% in time compared to 80 W settings). When compared to the 35 W values, the 45 W setting is consumes only 7.3% more power (76.2 W/71 W) while gaining back over 50% of the 'extra' time it took to complete the task.

Similarly 55 W (1.086x extra time, 85.8 W total) gains a little bit more than 75% of the extra time back for 21% more consumed total power (85.8 W/71 W) compared to 35 W. Compared to 80 W, 55 W saves you 20.5% power for only 8.6% longer in time. But it is all just a trade off of total energy vs time over this power range.

The 28 W 7840U processors will have a higher efficiency due to the 20% reduction in power compared to the 35 W settings for a 5-10% reduction in clock speeds/CBR 23 scores. But leak says 14789 pts so they may be on a slightly different N4 process than these higher power H/HS parts. Or just better binned. Maybe the HS parts are 'worse' binned than U and H are worse binned than HS.

Let's look at the N6 based 7735HS/6900HS scores of about 11200 and 12200 pts at 35 W and 45 W respectively. Assuming similar total power consumptions for RAM since both DDR5, similar refresh rates, SSDs, and taking off an additional 1 W for the difference between the mini-LED screen vs LCD, this would give a total energy consumption of 85.4 W and 96.1 W for the 35 W and 45 W 7735HS/6900HS respectively. These powers are 20.3% and 25.8% higher than the 7940HS at equivalent TDPs. These increases in power also correspond to time increases of 1.61x and 1.479x compared to the 80 W 7940HS and are substantially longer than the 7940HS at same TDPs. A 45 W 7735HS/6900HS consumes almost exactly the same total power as the 65 W 7940HS (94.4 W) to complete this task but the 7940HS completes the task 30% faster.

So yes. The 7940HS is more efficient than the 7735/6900HS. But the mini LED screen really seems to kill the idle power. Wait for reviews with standard LCDs to make final judgements.

On the spot. Andrei at Anandtech once made many efficiency calculations based on total power, he left quite some time ago though...I don't think there's a valid replacement for his analyses, at least not in public.

[James21342132]

The whole cool thing about these processors is the new AI cores, supposedly better than mac mini's M2, and you didn't test them AT ALL. Run an LLM, or stable diffusion or test it's TOPs for INT8. Do something, lol.