The rest of the media will focus on Koduri’s march, due to the fact that he is the person with the most media coverage. However, the departure of Blythe, who has been the chief architect of Ponte Vecchio and the ARC Alchemist, has become the beginning of the end for the Intel ARC. The problem? They were never good chips. And yes, we know that the departure of two managers does not have to affect the work of an entire team made up of hundreds of workers, but we find that Pat Gelsinger has beheaded two of the three top managers of the division and specifically those who they were in charge of designing the actual hardware.
The question about the future of the division is to know if they are going to develop a successor to the DG2-512 chip used in the ARC A770 and A750 or, failing that, they are going to reduce themselves to the work of making integrated graphics for the Intel Core and their corresponding drivers. . In any case, what has happened must be put into perspective and with information on the table to understand the end of the Intel ARC.
The case of the dwindling graphics card
The ARC Alchemist has had a curious history, to say the least, Intel initially set out to make it compete against NVIDIA’s RTX 3070, a graphics card that, before the price hike for cryptocurrency mining, had a recommended retail price of around 500 euros. , the RTX 3060, on the other hand, stood at 350 euros. The point is that the same graphics card, being sold for a higher price, obviously gives bigger benefits.
You can’t sell a graphics card with lower performance at a certain price, Intel’s first failure was to consider performance like the RTX 3070 and not reach it. Either because of the drivers or because of a bad design. Rather a combination of both. The reality is that Intel was forced to lower the price of its A770 below what was initially planned and it is not the only price drop they have had to endure.
That is, while AMD and NVIDIA can afford to increase the number of graphics cards to sell, Intel with its ARC Alchemist cannot, since it would make the hole even bigger. Hence, their distribution has been limited, due to the fact that they lose money with each ARC Alchemist sold and do not want the hole to be much bigger.
A disastrous GPU design in terms of costs
But the problem does not end there, you just have to compare the RTX 3070 chip, the AD104, and the ARC A770 to see how Intel needed a much more advanced node and almost twice as many transistors to get almost the same performance. or lower. It is said that with the latest drivers they have come close to the initial objective, but the reality is that Intel has to give them away and the design is not profitable.
| GPUs | Foundry | Technology | transistors |
|---|---|---|---|
| Intel ARC A770 and A750 | TSMC | 6nm | 21.700 million |
| NVIDIA RTX 3060 | samsung | 8nm | 12,000 million |
| AMD RX6650 | TSMC | 7nm | 11,060 million |
| NVIDIA RTX 3070 (Ti) | samsung | 8nm | 17.400 million |
| AMD RX 6750XT | TSMC | 6nm | 17.2 billion |
The reality is that the DG2-512 chip should have been back on the drawing board, the fact that NVIDIA was able to make a GPU with 50% more cores with nearly half the transistors and a smaller footprint should have meant going back to the drawing board. But Intel’s marketing had already kicked off the promotion of its graphics cards and there was no stopping it. Worst of all, bad blood between managers led to poor internal communication that meant that the hardware was for months in Intel’s warehouses without being able to sell it for mining, money that flew away before Intel’s eyes.
Ponte Vecchio, much ado about nothing
However, the decision of the Intel board of directors to decapitate Koduri and Blythe has nothing to do with the ARC Alchemist really, but the notorious failure is none other than Ponte Vecchio, A Frankenstein designed to combine various Intel manufacturing processes and TSMC to hide that the former’s foundries are having trouble making a large chip. On paper it is very good, but the execution, like the ARC Alchemist, with which it shares several elements, is disastrous. The fact is that in the midst of the AI boom, the sale of graphics cards for high-performance computing NVIDIA is getting tired of selling its H100, while Ponte Vecchio has a demand close to zero.
The cause of the problems? The fact that Koduri’s division used the same blocks to design their graphics architectures, a technique that all manufacturers do, but if we take into account the enormous number of transistors that their designs require, then it can be seen how these exceeded the limit of the lattice, causing it to split into chiplets and creating a fully disaggregated GPU.
Ponte Vecchio should never have existed, but the division had no viable product on the market, and the contract to put the Aurora Supercomputer on the table was lucrative. Not in vain had it been since 2017 and not achieving it after AMD will deploy the first supercomputer to reach ExaFLOP and that they did not achieve it was seen as a wound to Intel’s pride.
The future that was lost along the way
To understand part of the current problems of Intel ARC we have to go back to June 2021, two years ago, when Raja Koduri showed the first finished chip of the DG2-512 that equips the ARC A770, that is, Intel took a year in launching graphics cards with said chip. However, the design of a chip is not finished and it is given to be printed as a document. The reality is that since the year 2020 they had finished the design of the chip and were busy in not one, but two projects at that time.
Both designs were presented as Battlemage and Celestial on the roadmap, with the promise of Druid going forward, which was nothing more than a promise in a document. That is why the division for future development was divided into two groups. The goal was not to create a gaming GPU, but rather Tile GPUs that will be mounted on the interposer as small breakaway chips in Intel’s future desktop and laptop CPUs. So the goal was performance per watt and not raw power:
- Battlemage:
- It uses the same blocks as ARC Alchemist, not optimized for the number of transistors.
- TSMC 4nm node.
- 192 EU for the Tile GPU, that’s 1536 units on FP32 and a total of 12 cores.
- It will be implemented for the first time in Meteor Lake.
- Heavenly:
- Redesign of the different blocks of the GPU to be more efficient in terms of area.
- TSMC 4nm node initially with plans for a 3nm version.
- 320 EU as a target on a Tile GPU the same size as Battlemage. That is, 20 GPU cores.
- The goal is to implement it for the first time in Arrow Lake.
And what about future Intel ARCs?
This is where all the crux is, right now what would be an Intel ARC B770 seems to be in limbo waiting to see what Intel decides to do with it. Whether to launch it as they have it at the moment or simply move to the next generation. The reason for this is that the lower efficiency per area is repeated from one generation to another and this time worse, since we are not talking about Samsung 8 nm versus TSMC 6 nm, but a comparison of equal nodes.
In other words, Intel’s next graphics card should be based on Celestial and appear at the same time as the Arrow Lake desktop processors, which are being designed for 2024. The reality is that said chip is being designed as part of another project that is not It has nothing to do with the game market, but for the server market. We are talking about Falcon Shores, the next-generation Xeon that we were going to see in 2024, but which has been delayed to 2025.
So it is very likely that Intel will decide not to throw away the ARC brand to talk about its graphics cards, but it will do so in terms of those that are integrated into the processor and could disappear for three years as a rival to NVIDIA and AMD of stealthily and without making noise. In any case, they are nothing more than promises on a piece of paper.
