siliconbruh999
Well-known member
The analyst had a copy of Techinsight's analysis so we won't know how true is it cause it's paid.
Intel’s 18A rumors meet a thermal brick wall says SemiWiki
- Thread starter Daniel Nenni
- Start date
You should know that imec in itself is not a company that commercializes chip manufacturing; I've been there in the lithography department. It's a research institute that does research on the building blocks for scaling the microelectronics manufacturing. It would surprise me if they don't have research on new ways of cooling chips.
In the end it is up the IDMs and foundries to put together the building block in a commercial offering or product. It seems in the last 10 years TSMC has been better at this than anyone else.
Xebec
Well-known member
Is there any future possibility of running certain circuits without BSPD and the rest of the chip with BSPD to try to get better cooling in critical spots, or is that not possible due to yields, design rules, etc.?
For example, on 18A -- SRAM cells do not have BSPD because they would actually lose about 10% density. (I've also read claims 14A makes some changes to how BSPD is connected to the transistors that should allow more scaling with BSPD). In the 18A SRAM case, would there be some 'filler' material in place of the PowerVIAs that allows better heat conduction?
I'd have thought it's pretty much impossible, because the BSPD and FSPD regions have very different metal/substrate topography and the process flow is also different -- also FSPD has the power and I/O bumps on the topside and BSPD has them on the backside. This also makes improved cooling impossible since the chips are the other way up -- FSPD has the substrate at the top (good cooling) with all metal underneath the transistors and then bumps, BSPD has the thin metal/dielectric layers at the top (bad cooling), then the transistors and thin (<1um) substrate, then thick metal and bumps at the bottom. The two are fundamentally incompatible.
Having filler instead of powerVIAs in the SRAM region doesn't help with cooling, it's the wrong side of the chip (bump side not heatsink side). Having said that, SRAM is probably not such a problem for cooling anyway, at least in the memory array area, since power/current density here is very low.
jorgequinonez
Member
Backside Power Delivery Creates Fab Tool, Thermal Dissipation Barriers
Moving the power delivery network to the backside of a chip reduces congestion, but it introduces new challenges for fabs.
semiengineering.com
Pretty much word-for-word what I've been saying -- their detailed simulations predicted maximum on-chip temperature 23C hotter for BSPDN (1um to 3um square areas, Tj=80C instead of 57C -- but no details about heatsinking assumptions...) which is *very* close to the handwaving estimate of 20C for these sizes I gave earlier in the thread...Backside Power Delivery Creates Fab Tool, Thermal Dissipation Barriers
Moving the power delivery network to the backside of a chip reduces congestion, but it introduces new challenges for fabs.
All of which justifies TSMCs recommendation of BSPD for "HPC-class devices with dense power grids and active cooling", and shows the problem with using Intel 18A/14A for ASICs which don't fall into this category -- which is to say, most of them (by number of designs)... ;-)
Last edited:
How does Panther Lake being a a highly efficient mobile device with small cooling handle this properly? It seems to be a knock-out hit.Pretty much word-for-word what I've been saying -- their detailed simulations predicted maximum on-chip temperature 23C hotter for BSPDN (1um to 3um square areas, Tj=80C instead of 57C -- but no details about heatsinking assumptions...) which is *very* close to the handwaving estimate of 20C for these sizes I gave earlier in the thread...
All of which justifies TSMCs recommendation of BSPD for "HPC-class devices with dense power grids and active cooling", and shows the problem with using Intel 18A/14A for ASICs which don't fall into this category -- which is to say, most of them (by number of designs)... ;-)
Is it just that the computer chiplet is so small and low power that this issue doesn’t come up?
That's one possible reason, though not likely because it's local power density that matters not chip size.
Another is that it only gets hot for short periods at full load, so the fact that this might reduce lifetime doesn't matter -- or even that nobody will care if it dies after a few years. And let's face it, it's not like Intel have any choice, it's BSPD or nothing for them. Plus the fact that it bumps up maximum clock rate a bit which is what they care about, because -- well, data sheet and benchmarks, hence the "knock-out hit" comment.
Like I keep saying, for chips where the advantages of BSPD outweigh the disadvantages it's the right choice. For all the others -- which today is the majority of foundry chips -- FSPD is the better option.
siliconbruh999
Well-known member
Well on laptops i can't say that it only boosts for short period of time cause depending on what user do it has to maintain the clock or it looses performance like if user is playing video games i needs to maintain decent clock speed for like longer period of time otherwise you will notice a stutter or low FPS if the clock speed is not stable also some benchmarks are more demanding than the others
Xebec
Well-known member
To add - in the case of Panther Lake, 18A+BSPD didn't actually raise the max frequency over Lunar Lake or even Meteor Lake - all three (Intel 4, TSMC N3B, Intel 18A) max at 5.1 GHz. However, it looks like 18A+BSPD did pay some benefits on lower idle power and reduced energy usage under load. (Comparing CPU tile vs CPU tile). Real world usage testing confirms fantastic efficiency attributes for PTL, also ahead of AMD's N4 offerings.
(Also context - Raptor Lake refresh U/P parts topped out at 5.4 GHz as a comparison point).
You're misunderstanding what "short periods" means here (for hot-spots) -- it's not how long the "fast/hot/high-voltage" phase lasts (seconds or minutes), it's what percentage of the device lifetime is spent in this condition (because high temperatures and high voltages degrade reliability as well as performance).
Laptop/mobile CPUs spend most of their lifetime doing very little with short bursts of intense activity -- in this sense, an hour once a day (or week) is "short". Server CPUs/NPUs/switches spend most of their time 24/7/365 hammering away processing data, because anything else wastes money -- and also the DC owner had thousands of these, so will see any burnout failures. That's one reason why server CPUs run much slower and at lower voltages than consumer ones...
Last edited:
That may be partly due to the process, in a CPU like this BSPD will give more speed at the same power or lower power at the same speed than FSPD. And if you have a faster process you can always decide whether to increase the clock rate (keep similar gate structure) or reduce power while keeping the same clock rate (smaller gates) -- don't forget that all this also changes with voltage and transistor type, it's a *really* complex optimization problem...
But it can also be down to the CPU design/architecture or power/voltage control strategy, which changes for each new CPU. Without having the same CPU in both BSPD and FSPD processes (never gonna happen...) there's no way to tell how much is due to which -- but as an example BSPD will have the biggest effect on power in the high-power state (when processing), other CPU design aspects like sleep modes or performance/power core optimization will have the biggest power saving impact in the low-power states where typically most of the time is spent.
Xebec
Well-known member
Agreed - I just primarily wanted to point out that Intel did not actually use BSPD to raise clocks further. Though I would think given how many cores are idling these days even when doing "multiple tasks" -- the voltage stability opportunities provided by BSPD at idle/sleep should help increasingly too?
The efficiency piece is interesting too as I've seen some conjecture that TSMC N3 is a more efficient process than 18A, but the 18A Panther Lake is definitely exceeding the efficiency of previous-gen Intel products on N3B. (Definitely not apples to apples here, but worth noting).
Not sure what you mean by this -- estimates are that just the lower power mesh impedance / access resistance from BSPD gives a speedup of a few percent, but mostly at high voltages/currents i.e. maximum clock rate. Once you lower the voltage/current/clock rate this gain decreases because the voltage drops become smaller.
At the low voltages typically used for idle/sleep the benefit is negligible -- the limit here is that if you drop the voltage too close to transistor threshold voltages the timing mismatches due to random transistor variation go through the roof, so timing closure becomes more and more difficult. The tools then force you to add extra ultra-low-voltage timing margins which increase power consumption, negating some of the saving and making things worse (lower speed, higher power) at high voltages/clock rates.
So there's a complex tradeoff to do depending on the use case of the chip, things you do to save idle/sleep power tend to increase active power and vice versa -- so you also need to know how much time the chip is expected to spend in which mode.
For a CPU with lots of different modes and dynamic clock rate/voltage control this is a horrendously complex optimization problem, and it's a bit like whack-a-mole -- something you do to push down power in one mode is likely to push it up in a different mode. And if you compare different CPUs with different design choices, which one wins will depend on what the chip is doing...
Thanks,really helpful.
Miners are most likely to be liquid cooling too.
See also this recent article which says the same as I did but gives more background information... ;-)
Backside Power Delivery Creates Fab Tool, Thermal Dissipation Barriers
Moving the power delivery network to the backside of a chip reduces congestion, but it introduces new challenges for fabs.
semiengineering.com