At the J.P. Morgan's 14th Annual Tech Forum at the 2016 International CES. presented January 6th, 2016, Gregory Bryant, Intel’s new vice president and general manager of desktop clients platform was interviewed. As a result of that interview, a slew of stories appeared over the following days with headlines like this:
Mainstream gamers no longer require discrete GPUs, says Intel
Intel: For Mainstream Gamers, Our IGPs Are Equivalent to Discrete GPUs
Integrated CPU graphics are catching up with discrete graphics cards, says Intel
Intel Says Iris and Iris Pro Graphics Can Outperform 80% of Discrete GPUs – Casual and Mainstream Users Don’t Need dGPUs
Intel claims most don’t need a GPU
Intel's Integrated Graphics Could Hurt Sales Of Radeon And Nvidia Cards
Except that’s not what he said. I listened and transcribed his talk because I couldn’t believe Intel would make such a claim. They didn’t.
Here’s what Bryant said (11 minutes 20 seconds into the interview) “We believe that the performance of integrated today, what we offer in the products as built into the SoC is Equivalent to the performance of about 80% of the discrete installed base.”
Here’s the point: eighty-percent of the installed base of GPUs are three to five year old midrange mainstream graphics add-in boards (AIBs) typically in enterprise system and some PCs used for casual gaming. This is a low bar for Intel, but they’re justifiably proud of making the grade. The installed base Bryant refers to are AIBs no serious gamer would want in his or her house, let alone in their PC.
The other comments being made from Bryant’s interview are things like:
“Integrated graphics have nearly caught up”—no they haven’t, you think Moore’s law only works for Intel’s processors?
“If you’re a casual or mainstream gamer, you don’t need a discrete graphics card, Intel says”—Intel never said that. What Bryant said was: (10:56) “For the mainstream and the casual gamer, we’ve improved IRIS Pro, IRIS graphics integrated graphics tremendously. If you look at the installed (11:06) base of PCs, . . . . (11:11) we’ve improved graphics performance like 30x from what they were 5 years ago.” That's true and maybe for some people Intel’s integrated graphics are “good enough,” but Bryant didn’t come anywhere close to saying integrated graphics are equivalent to discrete graphics.
Most of the “stories” put out were copies of another story. Several even used a block diagram of IRIS graphics to make the point. A couple even included some benchmark data.
What do we think?
Intel has 14nm process, which is great for jamming a lot of transistors into a small space, and also for being able to run them at high-speed.
But . . .
IRIS graphics has with 48 and 72 execution units (shaders) with peak performance of 883 and 1152 GFLOPS, respectively. AMD’s R9 Fury has 4,096 shaders—almost 100 to 1.
IRIS is a shared memory GPU. It shares and is in contention with the CPU for a slower DDR3 or 4 memory, whereas a discrete GPU has a dedicated, large (4 to 12 GB) super high-speed memory of GDDR5, and in the case of AMD an enormous bus width (1024 bits).
AMD and Nvidia come out with a new multi-billion transistor GPU every other year. They too enjoy Moore’s law, albeit not at quite the same process node as Intel.
It’s beyond absurd to suggest an integrated GPU could “catch up” or replace a discrete GPU, it is the epitome of ignorance verging on stupidly to suggest such a thing. People hoping to influence a share price of one of the companies, or just trying to keep their name in the headlines put out such un- and ill-informed statements. I’ve been hearing them since 1992.
Intel is not to blame for this; it’s the sensationalizing press trying to sell headlines.
Can you imagine how stupid a headline would sound that said: Volkswagen says the new Beetle can outperform Ferrari? Of course not. It always helps to go back to the source. If it was once true, you can’t always believe what you read in the papers, it’s even more true that you can’t believe what you read on the Internet.
Recommended reading: http://journal.sjdm.org/15/15923a/jdm15923a.pdf