More Details on NVIDIA's Tegra 4 & i500: 5th core is A15, 28nm HPL, UE Category 3 LTE
by Anand Lal Shimpi & Brian Klug on January 7, 2013 1:30 AM ESTWe just finished NVIDIA's CES press conference where it introduced the Tegra 4 SoC and Shield mobile gaming console. Immediately following the press event we snagged some more information about Tegra 4 and the NVIDIA i500 Baseband silicon:
- Tegra 4 is built on TSMC's 28nm HPm process (low power 28nm with High-K + Metal Gate)
Just confirmed that our initial information was incorrect, it's 28nm HPL (28nm low power with high-k + metal gates). The difference between HPL and HPM is a optimization for leakage vs. peak performance. This helps explain the 1.9GHz max frequency for the A15s in Tegra 4.
- The fifth/companion core is also a Cortex A15, but synthesized to run at lower frequencies/voltages/power. This isn't the same G in and island of LP process that was Tegra 2/3. NEW: the 5th core will run at between 700 and 800MHz depending on SKU.
- The fifth/companion core isn't visible to the OS, it's not big.LITTLE but it'll work similarly to how Tegra 3 worked. This probably means no companion core in Windows RT.
- The four Cortex A15s will run at up to 1.9GHz.
- NEW: die size is around 80mm^2, slightly bigger than Tegra 3 but on a much higher density process
- NEW: the shaders aren't unified, the majority are 20-bit pixel shader cores though. No idea on the ratio yet.
- dual-channel memory interface, LP-DDR3 is supported
- NVIDIA's i500 will launch with LTE UE Category 3 (100Mbps downlink) support, eventually we'll see an update to UE Category 4 (150Mbps downlink).
- i500 will launch with carrier aggregation for WCDMA, no idea what 3GPP release.
Shield
- As far as Shield goes, I wanted to correct one thing about how the PC display streaming works. The PC will stream to the display directly, not through Shield. Shield will pass controller commands to the PC.
- Shield will launch in Q2 at a price competitive with other mobile gaming systems and tablets.
- All of the games during the Shield demo were 720p, except for one which was 1080p.
- Miracast is supported, but something better will come later.
Facebook
Twitter
RSS
34 Comments
View All Comments
DanNeely - Monday, January 7, 2013 - link
I'm guessing T4 may have been too far into development to change when ARM announced big.little; and with nVidia's consistently slipping mobile timeline adding additional changes/delays would have been a bad thing. I suppose we'll find out what nvidia thinks about the concept more definitively when T5 is announced next year; but a 1:1 mapping between low and high performance cores seems like a waste of die space to me on a quadcore design since you only really need 4 cores for heavy loads to begin with.tipoo - Monday, January 7, 2013 - link
But then they could do that with any of the other four core. What's the point in a fifth low power core if it's still an A15? I guess it probably runs on lower power optimized silicon, but the power difference probably won't be as big as the older big.Little implementation with a much smaller fifth core.Exophase - Tuesday, January 8, 2013 - link
Almost every multicore ARM SoC out there, including quad cores like Tegra 3 and Exynos 4412, already lets you individually power gate each core. The only SoC I can think of that didn't have this feature was Tegra 2.If this works anything like it did on Tegra 3 you won't be able to run the 4 full frequency cores at the same time as the companion core, so there'd be nothing asynchronous about it.
The whole thing sounds weird to me, especially without any manufacturing differentiation for the fifth core this time (I hear that was expensive). Maybe a power optimized A15 is a lot different from a frequency optimized one at the same frequencies, I don't really know.. previous cases from ARM don't seem to be that incredibly dramatic..
mayankleoboy1 - Monday, January 7, 2013 - link
Whats the power usage and thermals ? I strongly suspect that T4 in a smartphone will run at lower frequencies to avoid major throttling.Ryan Smith - Monday, January 7, 2013 - link
All we know for sure right now is that Shield packs 38Wh of battery cells, which is on par with 10" tablets such as the iPad 4 and Nexus 10.Krysto - Monday, January 7, 2013 - link
Do we know anything about its API support? OpenGL ES 3.0? OpenCL? CUDA? Anything?I found it strange that he talked about "computational photography" but not even once mentioned OpenCL or CUDA. Nvidia is being a little odd not revealing much technical stuff about Tegra 4.
Ryan Smith - Monday, January 7, 2013 - link
The GPU is non-unified. This almost certainly excludes OpenCL; OpenGL ES 3.0 we still need confirmation on. However we've yet to see a non-unified OpenGL ES 3.0 (or OpenGL 3.0) GPU, as unification and that level of technology have previously gone hand-in-hand.Auzo - Monday, January 7, 2013 - link
the non-unified GPU is the single disappointment in an otherwise awesome SoC. Too bad.Krysto - Monday, January 7, 2013 - link
Yikes. Seriously? Still non-unified?I was pretty sure they were going to unify it now because they needed to do that! Everyone else is doing it, so it seems pretty insane not to do it.
But in the same time I knew from the beginning that if they change the architecture now, they will still need to change it again in 2014 - and that seems rather soon, to do it again after a year.
Why change again in 2014? Because I think the Tegra 5 might be part of Project Denver, with custom CPU design, and new GPU architecture, to make it all streamlined from smartphones to servers and super-computers.
But it seems changing it in 2 years was going to be too costly for them, so they decided to just skip the architectural change now, and just do the 2014 one with the 64 bit CPU's. But what does this mean now - that Tegra 4 won't have OpenGL ES 3.0? Or it won't be as efficient? This isn't very good news.
I also still don't think I like the idea of a single A15 core acting working in low-performance mode vs 4 Cortex A7 cores. That basically just sounds like some sort of Intel Turbo-Boost. But to me the simple chip base for low performance sounds like a better idea than using a complex single core chip for it. Hopefully we'll see some data when Anand reviews both Tegra 4 and the upcoming Exynos 5 Quad (which I assume is the one with 2 clusters of 4 A15 and 4 A7's).
tipoo - Monday, January 7, 2013 - link
Huh, I thought T4 was supposed to bring in Nvidias newer unified architectures from the desktop/laptop size. T5 or T4 Plus then? I'm a little dissapointed. SGX has had them for a while if I'm not mistaken. Can't wait for benchmarks though.