Ever since Intel conducted the first Haswell graphics explainer, there’s been talk about just how much additional performance the company would deliver with its new line-up of graphics processors. Ivy Bridge, after all, was a major improvement over Sandy Bridge – so much so that Intel talked up its “Tick Plus” model as a key component of the first-generation 22nm processor’s appeal.
We’re still barred from talking about Haswell’s GPU performance directly, but Intel has released new branding information and some early performance figures.
If you plan on buying a Haswell-based system, you’ll want to understand these performance levels and how the new brands are being used.
The top-end Intel graphics products are now known as Iris Pro and Iris. Intel Iris Pro systems are based on the company’s GT3e graphics solution. That’s the chip with an on-board eDRAM frame buffer (pegged at 128MB by RealWorldTech).
The slide below describes the differences…
As of the writing of this article, Intel’s Iris Pro 5200 GPUs will only be available in 48 Watt mobile SKUs. There are two further flavours of GT3 – Iris 5100 and Iris 5000. The Iris 5100 parts are intended for 28 Watt mobile processors while the 15 Watt chips all use the Iris 5000. Intel isn’t revealing clock speeds yet, but the difference between Iris 5100 and 5000 is going to be boost/turbo clocks.
How do these SKUs impact desktop parts? They basically don’t. Intel will offer one BGA-based solution (the R-series) with GT3e graphics. All of the other desktop chips will use the HD 4600-4200 SKUs. These top out at 20 EUs (Execution Units), compared to Ivy Bridge’s 16 EUs. Thus, the desktop performance increase in graphics should be 15-20 per cent. That’s reasonable, if not particularly exciting.
The performance per watt trade-off
Intel’s exceedingly limited performance numbers suggest the performance trade-off in mobile is significant.
We’re going to ignore 3DMark 6 (because no one cares), and look at the 3DMarkVantage and 3DMark11. Forget the far right bar for the 28 Watt part, for a moment, and focus on the middle. That’s the Intel Core i7-4650U, a 15 Watt part with (presumably) Iris 5000 graphics. Factor in the 2 Watt TDP decline, and the performance boost between the new Haswell chip and the older Ivy Bridge CPU is actually excellent. Moving to GT3 picks up around 25 per cent in one benchmark, 50 per cent in the other.
While Intel is obviously going to cherry pick results, these should map to real-world games. GT3 is more than twice the size of GT2 (40 EUs as opposed to 16) and one reason Intel went to the wider array is because it could down clock the graphics cores and save power as a result.
The i7-4558U is a different story. Forget Ivy Bridge for a moment and look at the 4650U vs 4558U comparison. The 4558U is only 1.32 times as fast as the 4650U in 3DMark Vantage, and 1.5 times as fast in 3DMark11 – despite drawing nearly twice as much power. This suggests that GT3′s power consumption benefits are narrowly defined – the array will sip power if you keep clock speed constrained, but raising the clock speed bar will quickly turn the tables on power consumption.
The below graph is important for the GT3e performance figures. The middle chip, the 4900MQ, is erroneously identified in Intel’s deck as using Iris Pro 5200 – according to Intel representative Dan Snyder, it actually has a GT2 solution. This makes a great deal more sense as far as performance is concerned.
Compare GT3 (at 28 Watts) against GT3e (at 48 Watts), and you’ll see that while the relative speed-up favours GT3e, the benefit of adding a cache is smaller than we expected. That could be quite different in shipping products, but it implies that Intel’s L3 cache and ring bus for Ivy Bridge did an excellent job of delivering high performance at a modest cost per Watt.
Here are Intel’s own performance figures for the new 48 Watt parts:
Comparison between graphs should be reasonably accurate. HD 4000 performance varied modestly between mobile and desktop chips due to GPU clock speed; 3DMark11 puts very little emphasis on CPU performance. The Core i7-3687U has a max GPU clock speed of 1.2GHz, the Core i7-3840QM tops out at 1.3GHz.
The other stuff
Intel also claims that Haswell offers far faster video encoding, support for 4K displays, “collage mode” display, and DX extensions. Faster video encoding is worthless without better video encode quality, and the current state of GPU encoding is awful. That’s not Intel’s fault, per se, but the video quality we saw with QS on Ivy Bridge was so wretched, it obviated any advantage of using the feature. We’ll definitely revisit this topic when Haswell launches.
We’re glad to see the DX extensions added, but the other features are fringe-ish. They’re certainly not bad, but they aren’t major purchase-drivers, either.
Clear as mud
It’s hard to give even a preliminary opinion on Haswell based on the limited data available thus far. The bottom line appears to be this: Haswell’s GPU will deliver 20-35 per cent more performance than the equivalent Ivy Bridge chips did. That’s a significant year-on-year jump. It’s impossible to tell, at this juncture, whether the higher-performing GT3 and GT3e parts will deliver superior performance-per-watt as compared to opting for a low-end AMD or Nvidia GPU.
That’s why we’re characterising the overall improvement as modest. It’s clear that the next-gen GT2 part improves on Ivy Bridge, but the higher performance parts have significantly higher TDPs. It’s telling that Intel makes nary a mention of AMD or Nvidia in these graphs – we’ll have to put systems side by side to know for certain, but it’s possible that the GT3e RAM integration doesn’t offer better performance than a lower TDP CPU and standalone GPU from Team Red or Green.