Pages: 1 2
The really interesting question is what happens to future products at Intel that were related to Larrabee, and that really deals with three areas: HPC, discrete graphics and integrated graphics (IGP). Larrabee 2 was already in the works, and was largely a shrink to 32nm with quite a few improvements. However, the scheduling for that project was unknown and Intel’s roadmaps for Larrabee related products are likely to be a bit chaotic internally, let alone externally.
In HPC, Larrabee is still very attractive. Cache coherency is a big plus compared to existing GPUs, and supporting x86 means a nice familiar environment for many programmers. Larrabee’s performance is likely to be much more competitive, since the software stack is not a critical element in the same way that graphics drivers are. Additionally, Nvidia is really the only game in town for GPUs in HPC; ATI’s software stack is not really mature enough, so the competition is a less fierce than in graphics where ATI has an excellent product line. The tricky part here is that Larrabee 1 was not planned with ECC support, which Intel will want for the HPC market. The other downside is that graphics really are the volume driver for GPUs, and Intel will not have that volume on Larrabee products until it is a graphics product.
Nonetheless, Intel will probably go after the HPC market, since that is where GPUs actually pose a threat to the server business. In theory, they might be able to use a derivative of Larrabee 1 if they were sorely pressed, but Larrabee 2 is definitely a safer bet – or perhaps another product that emerges from Intel’s roadmaps.
For discrete graphics, the reality is that Larrabee and its software stack are yet unproven. An optimistic view would be that given a bit of time, the graphics software stack can mature and be ready to achieve high performance for Larrabee 2. Since the scheduling for Larrabee 2 is still unknown, that may be dicey. A more skeptical (and perhaps realistic) view that Charlie Demerjian is advocating is that it may take till the third generation to really have a suitable high performance graphics product. Either way, Intel will seed out Larrabee 1 development kits to the game industry, so that they can become familiar with Larrabee and pave the way for future products. Hopefully early next year, Intel will be able to clarify their plans for discrete GPUs and lay this issue to rest.
The one area that may benefit from the culling of Larrabee 1 as a graphics product is Intel’s IGPs. Recall that Intel will have an IGP on the same package as their 32nm CPUs shortly, and the IGP will be on the same chip in Sandy Bridge, the second generation 32nm CPU. Assuming that Intel will not have a discrete GPU till 2011 or 2012, then it would behoove them to increase the performance of their IGP offerings. By increasing the performance of their IGPs, Intel can effectively cannibalize the low-end of the discrete graphics market, where most of the volume lies.
Intel has two avenues to improve performance for IGPs, allocating more die area and thermal headroom and improving the drivers. Whether it is feasible to make hardware changes or not is unclear. In general, Intel tries to avoid last minute changes to CPU designs. However, allocating more die area for the GPU may be low-risk change, especially if it does not impact the CPU. It is almost certain that Intel will increase the investment in drivers for their IGPs which can have a substantial impact on performance. Improving an existing software base is low risk (compared to starting from scratch) and should not disrupt the scheduling of products, not interfere with any hardware design.
At the end of the day, the cancellation of Larrabee 1 was due to a number of factors, but mostly it comes down to overly aggressive scheduling, available resources and perhaps an underestimation of the complexity of the software that is involved on a high performance graphics card. Larrabee 1 took longer to get to market than expected and the window of opportunity closed. Perhaps one lesson that Intel may take from this is that the target design cycle time should be more in-line with the practices at ATI and Nvidia, rather than the CPU design cycles that Intel, AMD, Sun and IBM set – which are typically a bit longer.
Going forward this represents a fork in the road for Larrabee. On one side, Intel will focus and push ahead full steam for HPC products based on Larrabee – which is a very logical first step and avoids much of the complexity in graphics. On the other side, there will be substantial delays (at least a year or two) for graphics products based on Larrabee. In the interim, Intel will focus and improve their IGPs as best they can – within the constraints of their design and manufacturing flows.