2022 is a year for GPU history. After years of NVIDIA dominance and AMD’s attempts to compete on an equal footing, Intel fulfilled its promise to re-enter the graphics card market. Now we have NVIDIA, AMD and Intel competing for a pie that’s much bigger.
If we compile what has happened so far in this 2022, we have NVIDIA with its RTX 4000 series, how expensive and powerful they are; We have AMD with their RX 7000 graphics card due to appear this month; And then we’re left with Intel, which launched the Intel Arch as a mid-range proposition.
After watching this, let’s put all the makers on the table and delve into ArchitecturePeeling Back the Layers of Marketing to See What’s New graphics cardWhat do they have in common and what does it mean for the average user.
In this article you will find the most important information about each company, which will allow you to differentiate between each brand, its architecture and the chip that makes it possible. Without further ado, we get to the case that has a lot of graphics cards to bite:
AMD Y RDNA 3: EL Chip Navi 31
The Navi 31 is the largest chip ever built under the RDNA 3 architecture. Compared to the Navi 21, the new SoC promises to compete head-to-head with NVIDIA and its RTX 4000.
Shader Engines (SE) host fewer Compute Units (CUs), 16 versus 200, but there are now 6 SEs in total, two more than before. This means that the Navi 31 has up to 96 CUs with the Radeon 7900 XTX’s total of 6,144 stream processors (SPs).
Each shader engine includes a dedicated raster unit, a primitive engine for triangle configuration, 32 render output units (ROPs), and two 256KB L1 caches.
AMD hasn’t changed much in the raster engine and primitives either: the 50% claimed improvement is for the entire chip, as it has 50% more shaders than the Navi 21 chip.
The most obvious change is the one that sparked the most buzz and gossip leading up to the November release: Chiplet’s approach to GPU packages. With many years of experience in this area, it is only logical that AMD would have chosen this, but it is for cost/manufacturing reasons rather than purely performance.
In the Navi 31, the memory controllers and their associated end-level cache partitions are placed in separate chiplets (called MCD or memory cache dies) surrounding the primary processor (GCD, graphics compute die).
With more SEs to feed, AMD has also increased the number of MCs by 50%, so the total bus width of global GDDR6 memory is now 384 bits. There’s less Infinity Cache overall this time around (96MB versus 128MB), but the higher bandwidth makes up for it.
Intel and Su Chip ACM-G10
We move on to Intel and the ACM-G10 matrix (formerly known as the DG2-512). While this is not the biggest GPU Intel has ever produced, it is their biggest graphics matrix for customers like you and me (not regular users and servers).
The block diagram is a fairly standard layout, although it looks more Nvidia than AMD. There are 4 Xe-Cores each, with a total of 8 render slices, for a total of 512 vector engines (intel’s equivalent to AMD’s stream processors and Nvidia’s CUDA cores).
In addition, each render slice consists of a primitive unit, a rasterizer, a depth buffer renderer, 32 texture units, and 16 ROPs. At first glance, this GPU seems pretty big, with 256 TMUs and 128 RoP more than the Radeon RX 6800 or the GeForce RTX 2080.
However, AMD’s RNDA 3 chip has 96 compute units, each with 128 ALUs, while the ACM-G10 has a total of 32 Xe cores with 128 ALUs per core. Thus, in terms of the number of ALUs alone, Intel’s Alchemist-powered GPU is one-third the size of AMD’s.
Compared to the earlier Alchemist GPUs that Intel released through OEM vendors, this chip has all the signs of a mature architecture in terms of component count and structural arrangement. Unfortunately, they are still light years away.
NVIDIA and its AD102 chip, the RTX 4000’s jewel
We end our review of the different designs with NVIDIA’s AD102, their first GPU to use the Ada Lovelace architecture. Compared to its predecessor, the Ampere GA102, it doesn’t look that much different, just a lot bigger. And for all intents and purposes, it is.
NVIDIA uses a component hierarchy of a graphics processing cluster (GPU) consisting of 6 texture processing clusters (TPCs), each with 2 stream multiprocessors (SMs).
In the full AD102 die, the number of GPCs has been increased from 7 to 12, so there are now 144 SMs in total, giving a total of 18,432 CUDA cores on the RTX 4090 and slightly less on the RTX 4080. That may seem like a ridiculous number compared to the Navi 31’s 6,144 sm, but AMD and NVIDIA count their components differently.
Although this is oversimplifying things, an NVIDIA SM is comparable to an AMD CU: they both have 128 ALUs. So, While the Navi 31 is twice the size of Intel’s ACM-G10, the AD102 is 3.5 times larger.
So it’s unfair to compare the performance of chips when they differ so markedly in terms of scale. However, once you’re inside a graphics card, things change with its pricing and marketing.
With all this data on the table, Now you should be able to differentiate how each brand approaches its approach to gaming and technology Which makes it possible to play video games in 4K and with retracing. The world of GPUs is not easy, but it is very interesting.