CES 2015: Nvidia Announces the Tegra X1 Processor

During this year's Consumer Electronics Show (CES), Nvidia has given us a glimpse of their next generation mobile processor. And it's fantastic. For the first time implementing Nvidia's new mobile first strategy, they were able to quickly port their latest graphics architecture on to their Tegra line. As a result, the new Tegra X1 processor has a GPU based on Nvidia's brand new architecture, dubbed Maxwell. Thanks to its advanced power efficiency, Nvidia was able to build a very, very powerful GPU for the Tegra X1 without exceeding the power budgets that define the ultra-mobile market. The Tegra X1 will be most likely destined for high-performance and gaming tablets, and maybe even high-end Chromebooks.

Lithography

The truth is that 28nm is getting old. This generation, we're starting to see manufacturers moving to the smaller 20nm process node. Apple has done it with its A8 and A8X SoCs, and Samsung has done it with its Exynos 5433 processor. Nvidia has also now jumped on the 20nm wagon with the Tegra X1. Built by TSMC, the Tegra X1 is Nvidia's first SoC to benefit from a 20nm process node, and that should really help to keep the power consumption in check, which is necessary for Nvidia, especially considering the extremely beefy GPU.

CPU

Unlike with the Tegra K1, which had a 32-bit Cortex-A15 version and a 64-bit Denver version announced, for the Tegra X1 Nvidia has so far only mentioned one version, which ditches their own Denver core in favor of complete ARM-designed cores. It's still 64-bit, luckily. The Tegra X1 features a big.LITTLE CPU configuration, with four high-performance Cortex-A57 cores, and four Cortex-A53 cores designed for low-power operation. At this point, clock speeds are not specified, however. It must be pointed out that, unlike with Samsung's Exynos 5433, which also uses big.LITTLE quad-core Cortex-A57s and A53s, the Tegra X1's CPU cannot use both CPU clusters at the same time, so it can't be considered an actual octa-core CPU. Which is a good choice by Nvidia, considering that the Cortex-A57 is already a very powerful core, and considering how most applications don't scale well beyond four cores. At this point, performance estimates cannot be made due to the unannounced clock speeds, but given that the Tegra X1 is meant for high-end tablets, I imagine that the clocks will be pretty high, and it's pretty safe to say that the Tegra X1 will be no slouch when it comes to CPU performance.

Of course, another variant of the Tegra X1 with a Denver CPU is definitely very possible, but like with the Tegra K1, such a variant would only be released later into the year. In truth, I'm surprised that Nvidia didn't make Denver the default CPU configuration for Tegra X1. It showed very good performance in the Nexus 9, and I would like to have seen the Tegra X1 launch with it.

GPU

This is, of course, the spotlight of Nvidia's announcement. Nvidia's mobile first development strategy has enabled them to adapt their latest GPU architecture for mobile very quickly, and that will represent a huge advantage for Nvidia over the competition. With a 2x performance/watt advantage compared to Kepler, the new Maxwell architecture is extremely power efficient, delivering much more performance than Kepler, without using any more power.

For the Tegra K1, Nvidia had a single Kepler SMX (192 CUDA cores) running at up to 950MHz (although the devices that launched with it usually kept the clock speed at 850MHz). A single SMX had four ROPs (Render Output units) and 8 TMUs (Texture Mapping Units). However, with the Tegra X1, Nvidia is moving to two of Maxwell's basic graphics units, named SMM. Each contains 128 CUDA cores, therefore, the Tegra X1 has a total of 256 CUDA cores. These are accompanied by 16 ROPs and 16 TMUs, all this at, according to Nvidia, a max clock speed of 1GHz. This clockspeed sounds even a bit preposterous, and it is very possible that tablets running on Tegra X1 will keep the GPU clock at a bit less than that, for thermal and power budget purposes.

Wrapping up the technical stuff, here's a table comparing Nvidia's last few SoCs:

	Tegra X1	Tegra K1	Tegra 4	Tegra 3
CPU	64-bit Quad-core Cortex-A57 + Quad-core Cortex-A53	32-bit Quad-core Cortex-A15 @ 2.3GHz + Single "companion" Cortex-A15 core  or 64-bit Dual-core Denver @ 2.5GHz	32-bit Quad-core Cortex-A15 @ 1.9GHz + Single "companion" Cortex-A15 core @ ~800MHz	Quad-core Cortex-A9 @ 1.6GHz + Single "companion" core @ ~500MHz
Lithography	20nm	28nm	28nm	40nm
GPU core configuration	256 CUDA cores  16 ROPs  16 TMUs	192 CUDA cores  4 ROPs  8 TMUs	48 Pixel shaders  24 Vertex shaders	8 Pixel shaders  4 Vertex shaders
GPU clock	1,000MHz	950MHz	672MHz	520MHz
FP32 Peak Compute power (GFLOPS)	512	365	97	12.5
Pixel Fill Rate (MP/s)	16,000	3,800	?	?
Texture Fill Rate (MT/s)	16,000	7,600	?	?
Memory Interface	Dual-channel 64-bit LPDDR4-1600 (25.6GB/s)	Dual-channel 64-bit LPDDR3-1066 (17GB/s)	Dual-channel 32-bit LPDDR3-1866 (15GB/s)	Single channel 32-bit LPDDR3-1600 (6.4GB/s)

The table clearly shows how far Nvidia has come since the Tegra 3. The Tegra X1 is a huge leap forward compared to the K1, in every aspect, especially in the graphics department. The Tegra X1's GPU is far beyond what previous-gen consoles like the Xbox 360 and PS3 could achieve, and even some current low-end dedicated laptop GPUs are less powerful than the Tegra X1's GPU. Kudos to Nvidia for this impressive achievement.

Conclusion

Nvidia's new focus on bringing their latest GPU architectures to mobile is doing them a lot of good. The new Tegra X1 has a very, very large graphics processor, but with the benefit of the Maxwell architecture's astounding power efficiency.

In general, Nvidia's new processor is a great package overall, showing off excellent specs and top notch future proofing. Everything that's necessary for a new high-end SoC is there: 64-bit processing, 20nm process, for instance. While in most aspects Nvidia is playing in equal ground with other flagship mobile processors, its GPU sets it apart from anything else on the market now. The Tegra K1 was already ahead of pretty much every other SoC, except for the Apple A8X, in graphics benchmarks. Now the Tegra X1 will help Nvidia extend their lead even further. I was only a bit disappointed that Nvidia, at least for now, is not making use of its Denver CPU cores for the Tegra X1 (and considering how well they performed in the Nexus 9, one might wonder why Nvidia made this decision).

In terms of actual products that might eventually carry the Tegra X1, I believe that smartphones are still out of the picture. Despite the 20nm process and Maxwell's efficiency, a 256-core GPU might still be too much for a smartphone's battery size and thermal dissipation capacity. However, I can speculate that maybe a Tegra X1 with a much lower clocked GPU could make its way into a high-end phablet. 

Overall a great package, with the added benefit of a GPU that rivals even some lower-end dedicated laptop GPUs. Nvidia did a great job with its new SoC, and while it may still not be fit for smartphones, the Tegra X1 is just perfect for high-end tablets and any form of compact gaming devices, and might just turn out to be this year's most powerful SoC. 

CES 2014: NVIDIA Introduces Tegra K1 SoC: 64-bit Denver CPU and 192-core Kepler GPU

The Tegra line has always seemed like a second-thought product for NVIDIA due to lack of the innovation we've come to expect from NVIDIA. Well, this may be because they were busy working on something extraordinary, and it's finally here. NVIDIA's latest addition to the Tegra line, the Tegra K1, was announced today at its CES 2014 event, and it's pretty impressive. Tegra K1 brings NVIDIA's custom CPU core named Denver as well as a GPU built on the Kepler architecture, which according to NVIDIA can even outperform the Xbox 360 and the PS3 and DX11 compatibility, all the while keeping a 5W TDP. The PowerVR GPUs Apple always uses in its SoCs were always the pinnacle of mobile GPU performance, but if NVIDIA's performance claims about Tegra K1 pan out, Apple's GPUs will be utterly blown out of the water. 

Tegra K1 is, just like the Tegra 4, built on a 28nm process, which is pretty much the standard for modern SoCs, save for Intel's latest Atoms, which have already moved to 22nm. Hopefully the efficient 28nm process will keep the TDP at 5W or below, despite that beefy Kepler GPU. 

NVIDIA's latest SoC will actually come out in two variants. One will have a Quad-core Cortex-A15 CPU with a 2.3GHz clock speed and the other, which will only be available later this year, will feature a dual-core configuration of NVIDIA's own Denver CPU core. 2.3GHz is actually the highest clock speed we've ever seen a Cortex-A15 run on, so performance should be superb. The dual-core Denver-toting variant has an unknown clock speed, but what's really important is that Denver is a) NVIDIA's first custom ARM CPU and b) one of the first CPUs that use the ARMv8 architecture and therefore support 64-bit processing. I'm very excited to see how Denver performs when it comes out, and the Quad-core Cortex-A15 @ 2.3GHz will be very impressive too. Also, NVIDIA says Tegra K1 will, like its predecessors, use the 4-PLUS-1 architecture, so there's going to be a single "shadow" CPU core for handling light tasks while using very little power. Whether it's going to be used with both the Quad-core A15s and the Dual-core Denvers, I don't know, but I suspect the dual-core Denver won't need the extra shadow core. 

Perhaps the most interesting GPU we've ever seen on mobile is the Tegra K1's GPU. Considering how every previous Tegra GPU was based on a very old architecture and seldom topped benchmark charts, a jump to Kepler in one generation is quite satisfying. NVIDIA's Kepler GPU architecture was introduced last year and brought high performance and much better power efficiency to notebook and desktop GPUs, and even supercomputers, but NVIDIA has now achieved the impressive feat of bringing this architecture to the ultra-mobile space. The Tegra K1's GPU uses one full Kepler SMX, which is 192 unified shader units (or as NVIDIA calls it, cores). That's much more shading units that any mobile GPU has ever packed (for instance, Apple's A7's GPU had 128 shader units). NVIDIA claims that this GPU can even outperform both the Xbox 360 and the PS3. According to our calculations, it can, at 950MHz at least. At this clock speed, this GPU would have 365 GFLOPS of power, which is much more than the Xbox 360's 240 GFLOPS GPU and the PS3's 230 GFLOPS. I don't know whether NVIDIA's 5W TDP claim account for the GPU at 950MHz, but if it does (and it might, given how power efficient Kepler is), I pity NVIDIA's SoC competitors. For the record, to match the Xbox 360's performance, the Tegra K1's GPU would have to be clocked at 625MHz, which is actually lower than the Tegra 4's GPU clock. At the rather standard GPU clock speed for many mobile GPUs, 500MHz, the Tegra K1 can achieve 192 GFLOPS of peak theoretical performance, which is more than all of its competitors have reached. Of course, there are theoretical calculations, and we'll have to wait for a device running the Tegra K1 to be released to test whether its performance (vs its power consumption) is as good as it sounds.

The Tegra K1 GPU also touts DirectX 11 compatibility, and will probably also support OpenGL ES 3.0. NVIDIA showed us a demo of a Tegra K1 running a game simulation with the DX11-based Unreal Engine 4, and it just looked fantastic. Far ahead of anything we've ever seen on a mobile device. This is probably the first time when a mobile GPU's capability can really be called console-quality (almost every mobile GPU vendor makes that claim every year). While the GPUs on Tegras 2, 3 and 4 were a bit disappointing, Tegra K1 is exactly the innovation I was always expecting from NVIDIA in the ultra-mobile space. 

NVIDIA has, for the first time, come up with a mobile SoC that really pushes the boundaries of mobile processing. Its Denver cores will probably rival, if not outperform, the Apple A7's performance, and its 192-core Kepler GPU is downright amazing. Wrap that up with a 5W TDP, and you have just about the most impressive SoC to date. Now all NVIDIA has to do is ensure it can get OEMs to release devices using the Tegra K1, and before the competition catches up. Tegra 4's time-to-market wasn't bad, but the adoption of its last SoC wasn't very widespread, and Qualcomm's SoCs simply trumped the Tegra 4 in terms of OEM adoption. Hopefully NVIDIA will try to change that with the Tegra K1, maybe by releasing a Tegra K1 with an integrated Icera modem to attempt to find its way into LTE-enabled smartphones. For the first time NVIDIA has industry leading performance (previously Qualcomm held that title), so now it only needs to attract OEMs to use this fine silicon on their devices.

NVIDIA Tegra 4 makes an appearance on DXBenchmark; results are unimpressive

Today, some DXBenchmark results appeared on their website, referring to a device named NVidia Wayne. While there is not much information available, it looks a lot like a development board containing the much anticipated NVIDIA Tegra 4 SoC. The results were posted by an anonymous user.

The results point to a device named Covington, which could be the name of the development board. The GPU is named Wayne, which is just the codename for the Tegra 4. The screen resolution is 1371 x 771, and the device runs on four CPU cores, further indicating it's the Tegra 4, or the Tegra 4i.

As these results seem to be far below what NVIDIA promised for the Tegra 4, and as these results would point to the Tegra 4 being much weaker compared to the competition, it could be that these are benchmark results for the watered-down Tegra 4i. However, Tegra 4i`s codename is 'Grey', and not 'Wayne', despite the results having the name 'Wayne'. But still, these results do sound more like Tegra 4i than Tegra 4 from a performance standpoint.

Ok, so let's break these results down and compare them with current competitors and older Tegra iterations.

Fill rate looks particularly low in these results. In fact, Tegra 4 is supposed to have a fill rate close to the Nexus 10 or the iPad 4, since they all offer peak theoretical memory bandwidths of 12.8 GB/s. Even so, if this really is the Tegra 4 at final clocks, fill rate is extremely low, especially considering that current hi-res smartphones and tablets need high fill rate to satisfy the immense amount of pixels. It would make a lot of sense if these results pertained to Tegra 4i, since the lower end SoC has a single-channel memory controller (as opposed to Tegra 4's dual-channel memory interface) of 6.4 GB/s. If this was Tegra 4i, it would really explain the low fill rate.

In this fill test, Tegra 4 (or 4i, hopefully) appears far below the competition from Apple/ImgTech, and Google/ARM. Also, it appears to be only a marginal improvement over the last-gen Tegra 3, and even offers less fill rate than the Tegra T33 inside the ASUS Transformer Pad Infinity, which has HALF the theoretical memory bandwidth of the Tegra 4 and the same bandwidth of the Tegra 4i. The fact that the fill rate is lower than Tegra 3 is the main reason for me to believe that this Tegra 4/4i is either running at lower than final clocks or on incomplete drivers. 

Triangle throughput has always been the forte of the Tegra series, but the Tegra 4/4i doesn't seem to live up to that tradition in those results. While the triangle throughput is definitely an improvement over what is seen in Tegra 3, both of which outperform the ARM Mali-T604. it lags far behind Apple's latest iPad. 

The Tegra 4/4i performs very well in this test. In fact, it outperforms the beastly iPad 4's graphics, but that, of course, is because of the very low resolution at which Wayne is running in this test. Wayne is running this test at the lowest resolution of all the devices compared here, together with the Nexus 7. Nevertheless, whatever Tegra 4 iteration this is, it appears to perform almost 4x as much as the Tegra 3 (given that their resolutions are very similar), which does not fullfill NVIDIA's promise of a 6x improvement. Not quite what NVIDIA promised, but still, quite an improvement. 

Normalize the resolution to 1080p, and the iPad 4 takes the lead again. Still, this test paints a good picture for Tegra 4/4i. Compared to the Mali-T604 in the Nexus 10, Wayne could process 17 frames more (716 vs 699). Not a very significant improvement, but it at least shows that this Tegra can be as good as the Mali-T604. This test shows almost a 3x improvement over Tegra 3. Still not quite living up to NVIDIA's promise.

Now, here's a test we're more familiar with. At the low resolution of 1366x768, the Tegra 4/4i can really shine. Once again, it outperforms the iPad 4, both of which come closer to the Microsoft Surface Pro and its PC-class Intel HD 4000 graphics. This test shows a 2.5x improvement over the Tegra 3.

Normalize resolution, and the Tegra 4/4i not only falls behind the iPad 4, but also the Nexus 10. Now that is embarassing! Actually, the margin between Wayne and the Nexus 10 is quite big, and even bigger between Wayne and the iPad 4. Tegra 4 is only performing about 2.6x as much as the Tegra 3. And the two Tegras (4/4i and 3) sit embarrasingly at the bottom of the benchmark results.

Of course, these are probably results of a Tegra 4 device that either isn't running on final clocks or has premature drivers, if this is Tegra 4 at all. The weird thing is, Tegra 4 is about to be released this quarter of the year. By this point, NVIDIA should have Tegra 4 running on final clocks and drivers. If this turns out to be the true prowess of Tegra 4, I must admit my disappointment. Firstly, NVIDIA promised 6x the performance of Tegra 3, and hardly half of that was fulfilled, according to these tests. A 72-core GPU looked very promising, but if this is it, then Tegra 4 turned out to be far below my expectations. Still, chances are that this is Tegra 4's little brother, the Tegra 4i, which would explain the underachieving GPU (Tegra 4i only has 60 GPU cores and less memory bandwidth) and the lack of fill rate. Actually, I'm really hoping this is Tegra 4i, and not Tegra 4.