EETimes: TSMC starts FinFETs in 2013, tries EUV at 10 nm

[FwFred]

Kabini at 28nm will beat 32nm and 22nm ATOM and Celeron/Pentium, not to mention Kaveri at 28nm and GCN/HSA.

Beat Atom at what power levels? Please tell me we aren't comparing 18/25W Kabini to 2W Atom.

 

[Abwx]

What is the max power usage (w/ Turbo) for the GPU in 45W Ivy? How about for GPU in 35W Trinity

Power at the wall can be > 80W for a i7 QM , actualy
the CPU max TDP is not related to the claimed TDP
but to the thermal available headroom so the GPU
can be allowed to consume a lot if necessary....

http://www.notebookcheck.biz/Critique-des-processeurs-Intel-Ivy-Bridge-Quad-Core.77081.0.html

 

[Exophase]

Also, Kabini at 28nm will beat 32nm and 22nm ATOM and Celeron/Pentium, not to mention Kaveri at 28nm and GCN/HSA.

Silvermont will clock up to 2.7GHz and come in quad-core variants. It's way too early to say that the highest clocked Kabini we know of, the 2GHz quad-core, will be able to beat it. We know too little about Silvermont's IPC.

 

[Idontcare]

Ok, but do they already have to know which process they are going to use?

Process has to be known up-front and center. In fact knowing the specific electrical parameters of a yet-to-be-developed node is one of the critical parts of the iterative process of developing a process node.

Process guys make commitments - "we will hit Idrive = 1.5mA/mu, IDDQ of 1pA/mu, etc" that all get put into a spice model and preliminary design rules. Mind you the process node doesn't exist yet, but the specs for it must exist.

Why must the specs exist? Because the design guys must know exactly (more or less) what they are designing for. They have to take it on faith that the process development guys will deliver (4yrs from now) on their stated commits.

Which is why the whole thing can fall into a bloody mess so easily if the process guys fail to deliver their electrical targets. If the leakage is too high, or the drive current is too cool, or the reliability is so poor that the chips must be kept below rather low max operating temperature if they are going to last 10yrs and so on.

That is why Wichita had to be scrapped, utterly scrapped, 3yrs worth of R&D just tossed into the dustbin when 28nm failed to materialize in the form it had been promised (electrical-wise). Its what kills parametric yields throughout the production lifetime of the IC.

If Excavator is going to production in 3yrs then AMD must know precisely what process node (and who's) it is going to be produced on as of today (well, last year really).

 

[Sleepingforest]

Power at the wall can be > 80W for a i7 QM , actualy
the CPU max TDP is not related to the claimed TDP but to the thermal available headroom so the GPU can be allowed to consume a lot if necessary....

http://www.notebookcheck.biz/Critique-des-processeurs-Intel-Ivy-Bridge-Quad-Core.77081.0.html

Exactly. TDP and power consumption are related at best. Look at Trinity, for a second example:

It's the whole system, but so was the other test.

EDIT: I embarrassingly put up desktop Trinity power usage originally. What you see above is an edited post.

 

[FwFred]

Power at the wall can be > 80W for a i7 QM , actualy
the CPU max TDP is not related to the claimed TDP
but to the thermal available headroom so the GPU
can be allowed to consume a lot if necessary....

35W TDP Trinity is 32nm and has better performance in Graphics than 45W TDP 22nm Intel FinFet Core i7. Also, Kabini at 28nm will beat 32nm and 22nm ATOM and Celeron/Pentium, not to mention Kaveri at 28nm and GCN/HSA

My point is that AtenRa is trying to say that Trinity @ 35W vs. Ivy @ 45W is somehow indicative of future Haswell 35/45W chances of competing with Kaveri. I mentioned power usage, not TDP because the Intel GPU is hitting the frequency wall, not the power wall @ 45W.

See Jared's article about Ivy's 17W vs 45W GPU. Only 20% difference in clocks, and Ivy @ 45W keeps the GPU clocks pegged at max turbo frequency.

HD4000 17w vs 45w

I have no idea, but Trinity's GPU @ 35W could use more power than Ivy @ 45W (given the die area of each GPU). Haswell with 40EU instead of 16EU may be very competitive with Kaveri @ 35W.

 

[AtenRa]

Beat Atom at what power levels? Please tell me we aren't comparing 18/25W Kabini to 2W Atom.

We can compare low power Temash (sub 5W up to 6W), we can also compare higher power variants (9W or more) and Desktop APUs.

 

[AtenRa]

Silvermont will clock up to 2.7GHz and come in quad-core variants. It's way too early to say that the highest clocked Kabini we know of, the 2GHz quad-core, will be able to beat it. We know too little about Silvermont's IPC.

As of now,40nm Brazos 2.0 E2-1800 Dual Core (1.7GHz) is almost 70% faster in single thread (Cinebench R10) than 32nm Atom N2800 Dual Core + HT(1.86GHz). In Cinebench R10 MT, Atom closes the gap due to the HyperThreading(4 threads) but still falls behind.

Kabini will raise 15% the IPC over Brazos 2.0 and it will double the Core count to 4 cores. Silvermont will have Out of Order (OoO) but do you actually believe that it will catch Kabinis IPC ??

Im not expecting more than 20-30% IPC increase from Silvermont and that may be too optimistic.

 

[AtenRa]

Exactly. TDP and power consumption are related at best. Look at Trinity, for a second example:

That's just the CPU, since x264 HD is only CPU.

That is for the entire system not the CPU alone.

 

[Exophase]

As of now,40nm Brazos 2.0 E2-1800 Dual Core (1.7GHz) is almost 70% faster in single thread (Cinebench R10) than 32nm Atom N2800 Dual Core + HT(1.86GHz). In Cinebench R10 MT, Atom closes the gap due to the HyperThreading(4 threads) but still falls behind.

Using results from only one benchmark isn't enough to draw broad conclusions from.. Unfortunately there are very few reviews for Brazos 2.0 but if we just want IPC going with E-350 vs its contemporary Atoms should be fair.

Even more unfortunately HT makes it impossible to compare single-threaded IPC if the benchmark can load more than two cores :/

Looking at the following, trying to weed out stuff that's heavily influenced by GPU or other system stuff, and trying to pick out stuff that I think is single-threaded..

http://www.anandtech.com/show/4134/the-brazos-review-amds-e350-supplants-ion-for-miniitx/6
http://www.xbitlabs.com/articles/cpu/display/core-i3-2100t_7.html#sect1
http://www.lostcircuits.com/mambo//...ask=view&id=100&Itemid=1&limit=1&limitstart=8
http://www.tomshardware.com/reviews/asrock-e350m1-amd-brazos-zacate-apu,2840-11.html

Some IPC ratios:

AT Browsermark: 47.1%
AT Sunspider: 87.7%
XBit iTunes: 56.7%
LC Cinebench R9.5: 111.0%
LC Cinebench 10: 89.2%
THG Lame: 104.6%
THG Winzip: 47.5%

Average is 77.7%, but there's so much variation that I feel a little uncomfortable giving a representation IPC difference.. your number for Cinebench (86%) isn't that far off though.

Kabini will raise 15% the IPC over Brazos 2.0 and it will double the Core count to 4 cores. Silvermont will have Out of Order (OoO) but do you actually believe that it will catch Kabinis IPC ??

I don't have to believe Silvermont will catch Kabini's IPC, I said that it will be offered with up to 35% faster clocks.

If it increases IPC by 30% and clock by 35% the overall increase will be 75.5%, which will probably still keep it behind Kabini. But I still think it's too early to say for sure.

 

[Khato]

Im not expecting more than 20-30% IPC increase from Silvermont and that may be too optimistic.

At this point all we can do is make educated guesses based on the information available for both Kabini and Silvermont. There have been supposed leaks of Kabini performance figures, but who's to say that they're credible?

With respect to Silvermont the only tidbit that I'm aware of is the projected Atom 2013 tablet CPU performance according to SPEC2000int_rate being ~3.2x that of the 2012 platform. If the slides from end of last year regarding Valleyview are correct, then that performance figure would be for a 2.4 GHz quad core with no hyperthreading. I've emphasized that point because the current in-order Atom core receives a pretty hefty benefit from hyperthreading - on the order of 33% for spec2000int_rate and 50% for spec2000fp_rate. Such is quite important since that by itself effectively cancels out the clock speed difference in the single-threaded equation: <2012 Atom> * 1.33(hyperthreading) * 2(# cores) * 3.2(performance difference according to chart = <2013 Atom> * 1.33(clock speed advantage) * 4(# cores).

But yeah, none of that matters too much because it's nothing more than the logical conclusion to the information presented in PR slide.

 

[Ajay]

That is why Wichita had to be scrapped, utterly scrapped, 3yrs worth of R&D just tossed into the dustbin when 28nm failed to materialize in the form it had been promised (electrical-wise). Its what kills parametric yields throughout the production lifetime of the IC.

If Excavator is going to production in 3yrs then AMD must know precisely what process node (and who's) it is going to be produced on as of today (well, last year really).

Good stuffs IDC. Based on some Linked-In info (forgot where) Excavator (XV) was being worked on till the end of 2012. Now AMD hasn't mentioned XV for a while, so I wonder if it is still being worked on, but it maybe targeted 20nm and will be delayed or cancelled. Now the guy who posted this was a lead designer for some important CPU blocks, and he switched to AMD's ARM effort (don't know it it was voluntary or not). He is looking for a job while still at AMD. People change jobs for all sorts of reasons, so speculation will get me no where. Considering that I've heard that many good engineers are leaving AMD b/4 they have to endure the shame and aggravation of being downsized, this sort of thing just adds fuel to the fire.

Working on a high end CPU and finding out that it's been stuffed up by GFL not being able to deliver on the promised electrical parameters and that the project is being canceled or put on a back burner is the kind of thing that can make a good engineer bolt (here, now just implement this prefab ARM design, don't think too hard and get it done fast... meh). If I was working on some cutting edge stuff and then asked to work on some software that bored me, I would seriously be thinking of moving on (if it was a long term assignment).

 

[Idontcare]

Working on a high end CPU and finding out that it's been stuffed up by GFL not being able to deliver on the promised electrical parameters and that the project is being canceled or put on a back burner is the kind of thing that can make a good engineer bolt (here, now just implement this prefab ARM design, don't think too hard and get it done fast... meh). If I was working on some cutting edge stuff and then asked to work on some software that bored me, I would seriously be thinking of moving on (if it was a long term assignment).

You are absolutely right there.

When TI cut its CMOS R&D at 45nm I was already working on 32nm (and some 22nm pathfinding stuff including Finfets). Out of hundreds (thousands?) they laid off all the development engineers save for a "lucky" six (6).

Those six engineers were picked based on their rankings and breadth of skillset. They were re-assigned R&D jobs in analog, to be more accurate they were assigned the only R&D jobs in analog that were fully devoted to nothing but analog process R&D (up until then analog R&D was actually handled by the production engineers in a dual-purpose job role manner).

It was meant to be an honor, for the chosen six, to be hand-selected by TI's Sr VP's for this new and exciting adventure to go work on 130nm analog...:| yeah, right

I lasted about 2 months before handing in my 2wk resignation. It was the uncoolest, unsexiest, most unexciting job I had worked on since working at Burger King in college. Sure the pay was great, but the job satisfaction was simply not there at all. It was like stepping back a decade in time and carrying on with one's professional life as if sub-100nm never happened.

And you can bet a fair amount of reshuffled and reassigned engineers at AMD are experiencing the same "I did NOT spend 4yrs (10yrs) in college getting a BS (PhD) to do this for the rest of my life " epiphanies.

 

[Sleepingforest]

That is for the entire system not the CPU alone.

Actually, (and this is really embarassing) it's even worse than that: that was the desktop Trinity. If you don't mind, I'm going to go back and fix that.

His test was full system as well, observe the methodology (in French, but understandable):

• Processeurs Intel Core i7 Ivy Bridge
• Chipset Intel HM77
• 8 Go de RAM DDR3 (1333MHz)
• Nvidia GeForce GTX 670M
• Intel SSD 320 Series (80 Go)
• Écran 17.3" Full-HD LED
• Windows 7 Home Premium 64Bit

And the Trinity test can be found here (it's in a table, so not easily copy/paste).