This is the Xbox Series X SoC, a really powerful chip that bets on hybrid rendering
This is the Xbox Series X SoC,
a really powerful chip that bets on
hybrid rendering
Microsoft yesterday conducted a full technical analysis of the Xbox Series X SoC, a next-generation chip that uses a Zen 2 processor and an AMD Radeon GPU, supported the RDNA 2 architecture.
We have two vital pieces of data that confirm, on their own, that the Xbox Series X SoC marks an enormous evolution compared to the Xbox One X SoC. At the CPU level, the Zen 2 architecture represents a very large improvement in terms of IPC, something that alongside its higher working frequencies (2.3 GHz compared to three .8 GHz) confirms that we face one among the foremost important jumps within the system at the hardware level.
If we take a glance at the GPU we also find an excellent evolution both in quantitative and qualitative terms. The GPU that mounts the Xbox Series X SoC comes with 56 CUs, of which 52 are active, which leaves us with a count of:
- 3,328 shaders.
- 208 texturing units.
- 80 raster units.
- 12.15 TFLOPs of power in FP32.
In the image below we will see an entire diagram of the GPU. The presence of a unified geometry engine and support for mesh shaders is confirmed, also as hardware-accelerated ray tracing support, but we'll mention that later.
The processor has 8 cores and 16 threads, it works at a speed of three .8 GHz when it works with 8 threads and three .6 GHz when it handles 16 threads. Each core has 512 KB of L2 cache, and for every block of 4 cores, we discover 4 MB of L3 cache, which leaves us a complete of 8 MB of L3 cache.
In the attached image we will see an in-depth dissection with each of the weather of the SoC. The GPU occupies about 47.5% of the entire space of the chip. Right above it, we've got the interconnection system, the eight CPU cores, divided into two CCX units with four cores each, the multimedia elements, and therefore the GDDR6 memory. At rock bottom, we've got the I / O elements.
As expected, the Xbox Series X SoC uses the BGA interface, which suggests that it's soldered to the board and is manufactured in TSMC's 7nm + process, that is, it uses a more advanced process than the currently available Zen 2 processors. on PC.
Xbox Series X SoC and ray tracing: the key's within the texturing units
I have to mention that I used to be very surprised, as within the end AMD has chosen to implement ray tracing acceleration by following the method outlined during a patent that was leaked a short time ago.
AMD's approach is different from NVIDIA's with Turing. As our more advanced readers will know, Turing integrates RT cores grouped during a ratio of 1 for every active SM. once we activate ray tracing, the RT cores are responsible for calculating and handling all the workload related to the ray-triangle intersection loads and therefore the BVH transverse intersection tests. this suggests that the SM units don't need to exert any effort.
Each RT core is formed from two specialized units. the primary is that the one that performs the bounding box tests, and therefore the second is that the one that performs the intersection tests of rays and triangles, and is liable for sending the resulting information to every SM unit. As we've said before, all of this relieves SM drives from an enormous workload.
Well, within the Xbox Series X SoC, specifically in its GPU, we discover an approach focused on the texturing units since that's where the hardware responsible for ray tracing acceleration is integrated. Each of those accelerators is going to be ready to work with four texture operations or four ray operations per clock cycle.
I don't want to rush, but from what I even have seen thus far I feel we could find ourselves facing an implementation of ray tracing that's less than what NVIDIA used with Turing since it generates a transparent dependency on the texturing engines rather than a derivation filled with the workload to other elements, like RT cores, which relieve SM drives of the burden of ray tracing. we'll see how this translates into practice.
Regarding performance with ray tracing, unfortunately, we don't have data that permits us to form an immediate comparison between architectures, since Microsoft speaks of 380 Giga ray-box (bounding box tests) per second of peak speed and 95 Giga ray-tri (intersection of rays and triangles) per second of peak speed. their different values supported different units of measure than NVIDIA uses, so as we said, we cannot compare.
All in all, Microsoft's presentation made it clear that the Xbox Series X SoC will have details in common with Turing. As we will see within the accompanying image, the hardware-accelerated ray tracing is proposed as an "economic improvement of the normal system", and not as a definitive substitute. In other words, it's committed to hybrid rendering, and ray tracing will only be applied selectively to reinforce specific elements, even as NVIDIA did with Turing.
Although we cannot establish an immediate comparison in terms of performance between Turing ( RTX 20 series ) and therefore the RDNA 2 GPU of Xbox Series X, we do have another interesting reference that permits us to urge a rough idea. Microsoft talks about "3-10x faster acceleration with lower area cost" on the Xbox Series X SoC. Putting this in perspective can give us a thought of the performance we will expect.
In Quake 2 RTX, one among the foremost demanding games out there immediately thanks to its extensive use of ray tracing, a GTX 1080 achieves averages of 12 FPS in 1080p, while an RTX 2080 reaches 88 FPS. that massive difference is marked by the RT nuclei, which speed up ray tracing. As we will see, the performance improvement is gigantic and fits easily thereupon balance that Microsoft has given.
Advanced technologies to enhance performance
Power matters, but optimization is vital too. The Xbox Series X SoC features a strong CPU and next-generation GPU, paired with 16GB of GDDR6 memory and uses a state-of-the-art manufacturing process, but is additionally backed by a really robust ecosystem of APIs and technologies.
Microsoft has confirmed that the Xbox Series X SoC will feature ML (deep learning) inference acceleration with 3-10x performance improvement. this may allow the utilization of intelligent rescaling techniques, almost like NVIDIA's DLSS 2.0, to scale back resource consumption without a big loss of image quality.
It is not clear if the Xbox Series X SoC will have dedicated hardware (something almost like the tensor cores of the RTX 20 series) to figure with INT4 and INT8 operations, but a priori everything seems to point that these calculations are going to be administered through FP16 (single precision) via CUs (computing units). In any case, it's very interesting, and really important, especially if this technology is capable of reaching the extent that we've seen in NVIDIA's DLSS 2.0.
We have three other pending keys, the variable rate shader, the Xbox Velocity technology, and therefore the sound chip. Xbox Series X features a high-quality audio system that outperforms the Xbox One X's eight-core CPU. I'm unsurprised, after all, we're talking a few low-power, low-performance Jaguar chip.
Regarding the variable-rate shader, we already explained intimately what it consists of when talking about DirectX 12 Ultimate. consistent with Microsoft, this system could improve performance by 10-30%. great in the least, of course.
As for the Xbox Velocity technology, its operation is sort of simple. it's a system that permits the utilization of a part of the SSD drive as a "slow" cache memory that is support for the system memory, that is, it allows the utilization of the SSD's NAND non-volatile storage as a backup for the 16 GB of GDDR6 memory. This has allowed Microsoft to scale back the quantity of GDDR6 memory required to make its next-generation console, and has had a positive impact on manufacturing costs, and thus also on the asking price.
And speaking of costs, Microsoft has also recognized that the worth of the Xbox Series X SoC is far above that of Xbox One X. Again, it doesn't surprise us, in any case, we are talking a few very complex chip that comes, additionally, manufactured during a 7 nm + process and which is, therefore, much less mature than the 16 nm + process that used the chip of said console.
e don't yet have definitive information on the sale price of Xbox Series X, but consistent with the newest information it should be around 500 dollars.
Post a Comment