NVIDIA Reveals First Details about Project Denver CPU Core
by Anand Lal Shimpi on January 6, 2014 12:08 AM EST- Posted in
- SoCs
- CPUs
- Project Denver
- NVIDIA
- Tegra K1
During its CES press conference, NVIDIA revealed its new Tegra K1 SoC will be available in two versions. One version will ship with four ARM Cortex A15s, while the other will ship with two NVIDIA designed Denver CPU cores. We don't know much about Denver other than it's a custom 7-wide superscalar 64-bit ARMv8 design.
NVIDIA claims very high single and multithreaded CPU performance. It's important to note that the Denver version only comes with two cores, something we've argued is likely the more optimal use of die area in mobile. Max frequencies top out at 2.5GHz. Given that NVIDIA has silicon back already I suspect this is a 28nm design. The L1 cache is 128KB + 64KB (I$ and D$?).
The Denver CPU cores will be paired with a 192 CUDA core Kepler implementation (1 SMX). No word on differences in power or performance. The Cortex A15 version of Tegra K1 will be available in 1H 2014, while the Denver version will be available in 2H. The two SoCs are pin-compatible, indicating that there won't be any difference in memory interface width.
Facebook
Twitter
RSS
36 Comments
View All Comments
jasonelmore - Monday, January 6, 2014 - link
just look at the cores in the picture.I know these are not die shots, but it looks like Denver cores are roughly double the size of A15 Cores.
Krysto - Monday, January 6, 2014 - link
They're 7-way vs 3-way superscalar so I'm not surprised. It's interesting that they're moving to dual-cores, though. So far Denver looks more like a competitor to Haswell/Broadwell than any other ARM chip. I'm actually starting to lose hope that it will be very efficient.iwod - Monday, January 6, 2014 - link
Since Nvidia are now licensing their Ge force Tech out, I wonder what's the implication of that.And no Maxwell?
Jeffrey Bosboom - Monday, January 6, 2014 - link
Why would you want a larger instruction cache than data cache? Is this something specific to mobile?Exophase - Monday, January 6, 2014 - link
Enticing article. Not the news I was expecting today.One thing I don't agree with - I don't see why the Denver K1 has to be 28nm even if there's silicon already. TSMC 20nm silicon is easily possible at this point, it entered risk production forever ago and volume production is due to happen very shortly if not already underway. This could also explain why the Denver variant is coming later, although to be fair it's really coming earlier than any of us expected, unless this K1v2 is actually Parker..
The only thing that makes me question 20nm is nVidia's more or less open criticism of following leading edge TSMC processes. And that on paper 20nm doesn't offer terribly great power consumption reduction.
errorr - Monday, January 6, 2014 - link
Just to update, the 20nm FPGA's are already available as of November. Supposedly "volume" production should start in the next few weeks and ramp up through Q4.Krysto - Monday, January 6, 2014 - link
I think they want to make both K1 and K1v2 (Denver) on 28nm HPM, and then just jump straight to 16nm FinFET in 2015 (hopefully in first half of 2015). They are probably one of the very first customers of TSMC for 16nm FinFET, because they mentioned it since last year.I wish AMD would do the same and jump to FinFET as soon as it's available, and forget about extra costs. It would give them a much needed boost to shrink the gap between Intel in CPUs.
jospoortvliet - Monday, January 6, 2014 - link
I feel the word 'details' is a bit of an exaggeration... ;-)MrSpadge - Monday, January 6, 2014 - link
At least it's more detailed than anything we officially knew before ;)errorr - Monday, January 6, 2014 - link
I'd be more worried about the cost of these. Even if they are 'faster" than A7 they are still going to be significantly hotter and suck more battery.Until tablet android becomes a viable ecosystem where there are actual apps optimized for those screens I don't see the point of Denver. Android tablets are about media consumption and other use cases are theoretical for now.