Intel's integrated graphics solutions spawned out of a simple need from the business/corporate world: to have enough graphics power to host a desktop, surf the internet, and write up documents/view presentations. This was a simple task, and Intel found it had extra unused die space which could be utilized for it, but the graphics engineering team was never given more than what was deemed "unused space." This is why the power of the integrated graphics processor has been weak, but there are more fundamental differences between Intel's HD Graphics architecture and Nvidia's Kepler/Maxwell or AMD's Hawaii.

For simple instance, Intel's graphics architecture does not possess a lone multiply instruction in its SIMD units. Nvidia and AMD both implement the ability to simultaneously process 1 fused multiply-add instruction and 1 multiply instruction. To put this in perspective, HD Graphics 5300 possesses 24 EUs, each made of 2 SIMD units, each hosting 4 32-bit floating point units (also capable of integer calculations), each FPU capable of just 1 FMA instruction. To calculate FLOPS, we use the formula EUs * SIMD/EU * FPU/SIMD * (FMA + multiply instructions) * hertz. HD 5300 = 24 * 2 * 4 * (1 multiply + 1 Add) * 850 * 10^6 = 326.4 Gigaflops.

If Intel implemented a lone multiply instruction as well, the total flops would be 192 * 3 * 850*10^6 = 489.6 Gigaflops, or a 50% increase in throughput capability.

On more complex notes, Intel's graphics architecture does not possess a hardware-level tessellation or polygraph engine, technology very often used in games nowadays. Of course, Intel did not develop its integrated graphics for games, so it's not surprising for them to lack this technology currently.

Currently Intel's best graphics solution, Iris Pro 5200, possesses 40 EUs and runs at 832 Gigaflops. If Intel implements a simultaneous multiply instruction, the total calculating power of this SKU would rise to 1.248 teraflops, a hefty sum for an integrated graphics processor. The number for Iris Flops is somewhat startling, because this means the count of SIMD units per core and FPUs per SIMD is the exact same between Gen 7.5 and Gen 8. This is a bit annoying personally considering how much die area was gained in the shrink from 22nm to 14nm.

Intel has promised Gen 8 graphics (Broadwell, Braswell, and Cherry Trail iGPUs) will provide 20% more cores per tier (48 vs. 40 on the next Iris Pro 6200 SKUs for example), better tessellation performance, more cache per GPU core (not the 128MB eDRAM LLC, but the L1/L2 GPU cache), increased pixel fill rate, and some unknowns. This paragraph's information is courtesy of the Motley Fool.

Furthermore, Gen 8 has successfully provided double precision and half precision support. Shared virtual memory has also been provided (making programmers' lives much easier). The number of EUs per subslice (a compute group sharing a sampler and data port, where more EUs sharing reduces throughput) has decreased from 10 to 8. Intel has added a 3rd subslice to each slice (collection of computing groups), allowing more variations of GPU configuration. Lastly, local bandwidth from cores to L3 cache has been improved for better performance (The Compute Architecture of Intel Integrated Graphics Gen8 pg5).

While this is impressive and shows Intel's hard work, some unfortunate facts remain: Intel has next to no patents on 3D graphics rendering or architecture. The lion's share is possessed by Nvidia, AMD, Qualcomm, ARM Holdings, and Imagination Technologies. Currently Nvidia is allowing Intel access to the bottom of the barrel graphics intellectual property it has in exchange for some of Intel's CPU-based IP which is integrated into Nvidia's Tegra processors. However, it's far from likely that the discrete GPU duopoly of AMD and Nvidia will ever see another competitor in that realm, even from Intel, as long as their patents hold and are not forced to be made publicly usable. Even if you despise the idea of Chipzilla taking over the graphics market, it's hard to argue with the fact the GPU market has all but stagnated the last 3 years, and Nvidia could do with stronger, more well-funded competition. That is not to dismiss AMD, but rather to say AMD does not currently possess the finances to actually force innovation from Nvidia at this time.

Intel's integrated graphics solutions are not particularly strong for gaming, but they are very good for accelerating scientific computing and driving basic graphics needs for office workers and professional picture editors. In the most extreme case of Iris Pro Graphics, you have an okay gaming solution for 1080p titles, but the price is far too high for what is offered at this point in time ($600 for the 4950HQ). What Intel can do about this going forward is anyone's guess, but in my next blog entry I will lay out a few ideas I've mused about and figured the likelihoods of.