Do you have any more information on saturating vmax/vmin instructions? I can only find reference to FMA and half precision conversion extensions (http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0473c/CIHJEBCE.html)
Tegra 3 officially announced; in tablets by August, smartphones by Christmas
- Thread starter Mike11
- Start date
Do you have any more information on saturating vmax/vmin instructions? I can only find reference to FMA and half precision conversion extensions (http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0473c/CIHJEBCE.html)
It's more accurate to say that the implementation of VMAX and VMIN on A5 doesn't quite match the documentation (or rather, what one would expect). It doesn't say what happens to the source register (the smaller of the two in VMIN, larger of the two in VMAX). In A5 and I believe A15 as well, IIRC, the sources are saturated to max/min values.
In Yamato's case that may not be so obvious. Actually, I'd say it's far from obvious there. IMO, the architecture does need a balance between crunching power and tiling constraints - buffer resolves are not exactly free there. Increasing GMEM is a very low-hanging fruit performance-wise, so there's will always be the temptation to pick it, by whoever designs the next Yamato iteration.
At the end of the day, Yamato is a spawn of an architecture that took tiling as a second priority. Many of the characteristics of today's Yamato iterations stem from the original design priorities of that precursor.
The architecture manual looks clear to me, there's no reason for the source register(s) to be modified (unless of course it is also the dest reg). If A5 or A15 changes it then it looks like a bug.
D
Deleted member 13524
Guest
Is that frequency scaling or core-scaling for the "CPU" portion? I had no idea T3's memory system was clocked so high; it's still single-channel, right?
Wild guessing as I don't know that benchmark...
The integer part result is probably both freq and core scaling though it should be higher if there was perfect scaling.
The memory part might be due to either better bandwidth utilization thanks to the 4 cores (if the part of the benchmark isn't able to defeat L2 caching) or perhaps they just widened the DDRn memory interface.
The 3D graphics seems to be the same as before, that's odd.
The big difference in 3D scores between Xoom and Optimus 2X suggest that it's resolution dependent. Hence it's better to compare these results, most likely from a tablet, to Xoom's, where the 3D score is much higher. But could still influenced by vsync.
Kal-El doesn't support a larger DRAM width, and the supported frequency is only somewhat higher.
Kal-El doesn't support a larger DRAM width, and the supported frequency is only somewhat higher.
Ha yes indeed; according to Anandtech it's still 32-bit wide and frequency went from 600 to 800 MHz for LPDDR2. So I guess that's some benchmark artefact.
I think the 2D, 3D and database scores are most telling about how the overall experience will feel relative to other devices. I'd like to see some javascript / browsing benchmarks.
I'm surprised to see memory bandwidth improved considering that I read it would be very similar to Tegra 2.
I'm surprised to see memory bandwidth improved considering that I read it would be very similar to Tegra 2.
As I wrote this might be a benchmark issue: it might be unable to saturate T2 memory interface when running on only two cores. Hard to say given it seems to only provide a score and not MB or GB/s figure.
Incidentally, you've been able to saturate said Tegra 2 interface, right?
I've been thinking about memory bandwidth vs latency on these devices more. I think on something like OMAP3 it's not possible to come very close to saturating bandwidth with just the CPU, because the store queue is not big enough to hide latency for stores, and the preload queue is not big enough to hide it for loads. The latencies are just too high for the CPU/cache design. But you can get a lot closer using the pre-load engine on Cortex-A8, since it involves programmed sequences instead of explicit requests.
In a situation like this more cores would be an obvious help, like you say.
A9 should bring in a similar advantage by having automatic prefetching built into the core. Then it's a matter of whether or not the benchmark triggers it. But quality could still vary with Tegra 2/Kal-El having a potential advantage if the L2 cache and/or memory controller has its own additional prefetching or deeper issue queues.
D
Deleted member 13524
Guest
Anand's Tegra 3 article from February actually says:
They're using Elpida LPDDR2 chips, which exist in either 800 or 1066MHz variants.
So where did the 800MHz assumption came from?
I assume most people are just speculating, but I looked into it and they're actually right. The OMAP4 Pandaboard uses Elpida chips with the same performance class (end of the marking) and OMAP4 doesn't support more than 800MHz so it seems unlikely they paid for something faster.
D
Deleted member 13524
Guest
First Tegra 3 smartphone, HTC Edge:
http://pocketnow.com/android/htc-edge-first-quad-core-phone-revealed-images
Tegra 3 AP30 @ 1.5GHz
4.7" 720p screen
1GB RAM
32GB internal memory
8MP F2.2, 28mm camera
1080p video recording
LTE isn't confirmed.
Expected Q1 2012.
If confirmed, isn't this HTC's first Android smartphone without a Qualcomm SoC?
http://pocketnow.com/android/htc-edge-first-quad-core-phone-revealed-images
Tegra 3 AP30 @ 1.5GHz
4.7" 720p screen
1GB RAM
32GB internal memory
8MP F2.2, 28mm camera
1080p video recording
LTE isn't confirmed.
Expected Q1 2012.
If confirmed, isn't this HTC's first Android smartphone without a Qualcomm SoC?
Weren't there some OMAP3 based HTC smart-phones in the past also? I'm not in the least surprised if phone manufacturers don't want to depend on just one SoC or technology source if you prefer.
SoC manufacturers lose in one spot and gain in another; the map is consistently changing and stakes seem to be higher than ever before.
What remains now to be seen is what a quad core CPU will accomplish in a smart-phone.
SoC manufacturers lose in one spot and gain in another; the map is consistently changing and stakes seem to be higher than ever before.
What remains now to be seen is what a quad core CPU will accomplish in a smart-phone.
I'm far more optimistic on the CPU side of Tegra 3 than I was before from the standpoint of efficiency (but not performance), after listening better to their explanation of how processing will be balanced.
Tegra 3 will still be a let-down, though, due to the underwhelming graphics performance. Tegra 3+ is kind of funny when you think about it in context.
But, wow... those HTC Edge rumors are fascinating in a number of ways in its specs.
Tegra 3 will still be a let-down, though, due to the underwhelming graphics performance. Tegra 3+ is kind of funny when you think about it in context.
But, wow... those HTC Edge rumors are fascinating in a number of ways in its specs.
D
Deleted member 13524
Guest
Google tells me there was this HTC Qilin for chinamobile back in 2009, but it was a WM6.x device, not android (many of the old windows mobile HTC smartphones used OMAP SoCs, BTW).
Nonetheless, it's been at least two years since HTC used anything other than a Qualcomm SoC for any of their smartphones.
