Thursday, March 02, 2006

Intel GMA 950: Terrible OpenGL performance

The first benches of the Intel Mac mini are out, and among them are some Cinebench scores. As suspected, the OpenGL performance of the integrated Intel GMA 950 graphics chipset leaves much to be desired:
Mac mini G4 1.5 GHz
Rendering (Single CPU): 152
Shading (CINEMA 4D): 159
Shading (OpenGL Software Lighting): 414
Shading (OpenGL Hardware Lighting): 506
OpenGL Speedup: 3.18

Mac mini Core Solo 1.5 GHz
Rendering (Single CPU): 213
Shading (CINEMA 4D): 259
Shading (OpenGL Software Lighting): 885
Shading (OpenGL Hardware Lighting): 441
OpenGL Speedup: 3.41
In the OpenGL Hardware Lighting test, the old G4 Mac mini with Radeon 9200 is actually 20% faster than the new Intel Mac mini. In fact, even the CPU-dependent OpenGL Software Lighting test on the Intel Mac mini is faster. While it's true that the Cinebench OpenGL test may not be representative of gaming performance, this does illustrate just how much of a limitation the new Intel Mac mini's integrated graphics can be.

The good news is that the other benches confirm the great superiority of the Core Solo CPU over the G4. Clock-for-clock, Core Solo is 40% faster. Furthermore, we must not forget about the 1.66 GHz Core Duo Mac mini, which scores 471 in the CPU rendering test, over 3X as fast as the old 1.5 GHz G4 Mac mini.

Tuesday, February 28, 2006

Apple releases Intel Mac mini

As widely predicted, Apple released the Intel Mac mini today.

Before the announcement I had wondered if Apple would choose to offer a cut down single-core Yonah in the Mac mini for cost reasons, but had cautiously hoped for the more expensive Core Solo. Well, we got the Core Solo (at 1.5 GHz), but we have to pay for it. Apple has raised the prices. The base model mini is now US$599, $100 more than before, although the new model does get other upgraded features to help justify the price increase. What surprised me however was Apple's choice to offer a 1.66 GHz Core Duo option as well. A Core Duo Mac mini will be one fast little machine in terms of raw CPU speed, and on average in multi-processor aware native applications, it will be faster than the fastest G5 iMac and capable of decoding most 1080p H.264 video cleanly.

The other huge improvement to the Mac mini's design is the inclusion of a second memory slot. This will make memory upgrades much less complicated. It's unfortunate though that Apple's base model uses two 256 MB sticks, not one 512 MB stick, so to upgrade the memory one must remove one stick. However, it is possible this is intentional, for dual-channel memory support.

Other nice improvements include Front Row with iTunes sharing support and the Apple Remote, 7200 rpm hard drives, analogue audio in, optical digital audio in/out, four USB 2.0 ports, Gigabit Ethernet, and built-in Airport Extreme and Bluetooth.

A big potential downside to the new Mac minis is the video. Apple has chosen to include Intel's GMA 950 integrated video chipset, which shares system memory for the video. On the other hand, GMA 950 does offer at least one major advantage (besides potential space & cost savings) over the Radeon 9200 that was in the previous G4 Mac mini, which is full CoreImage support (at least on paper). Unfortunately, the GMA 950's 3D performance is poor, and it's much slower than the Radeon 9550, which is in the current iBook G4 and which also supports CoreImage. Ironically, Apple summed it up quite nicely in its previous G4 Mac mini sales pitch:
Go ahead, just try to play Halo on a budget PC. Most say they’re good for 2D games only. That’s because an “integrated Intel graphics” chip steals power from the CPU and siphons off memory from system-level RAM. You’d have to buy an extra card to get the graphics performance of Mac mini, and some cheaper PCs don’t even have an open slot to let you add one.
However, if the Mac mini does indeed support dual-channel RAM, then perhaps graphics speed hit due to graphics memory swapping might not be as bad as it could be (if the system has a sufficient total amount of RAM). It would be interesting to do 3D gaming speed comparisons on the Mac mini with two sticks of 512 MB, vs. one stick of 1024 MB.

The other small issue is lack of HDCP support. Given that this unit has a fast enough CPU to decode most HD H.264 content, one would hope that it would be able to send the HD content (from an external Blu-ray drive) to a DVI screen. Unfortunately, without HDCP support, that HD stream will be downsamped to 540p. That's still better than DVD, but not ideal. Perhaps the next iteration of the Mac mini will support HDCP, especially considering that by then Blu-ray slot-load laptop drives should be available.

All in all, while not perfect, this new Intel Mac mini is a winner overall. It's a nice well-integrated design for someone with basic computing needs, and it can also fit quite nicely into many a home theatre system. I personally won't be buying one now, but perhaps I will at the next iteration or at least once I get a DVI-capable TV.

[Update 2006-03-01]

1) Apple has removed the references to 7200 rpm drives.

2) Blizzard Entertainment, who created World of Warcraft, is guessing that the new Mac mini may perform better overall compared to the old G4 Mac mini with this game. They too suggest installing memory in pairs (assuming the Mac mini supports dual-channel RAM) for optimal graphics performance:
After some examination of the Intel-based Mini's specs, it's my opinion that the ideal config for WoW will be to have a minimum of 1GB RAM installed, as a matched pair of DIMMs.

The CPU (either Core Solo or Core Duo) can use up to about 5GB/sec of memory bandwidth maximum; by having matched pairs of DIMMs installed, thhe machine has around 10GB/sec of total memory bandwidth to share between CPU and GPU tasks. At times when the CPU is executing out of its L2 cache (which is fairly large in the Core Solo and Duo), the GPU will have a larger percentage of that RAM throughput available.

A 1024x768 screen, refreshing at 60Hz and 24-bit color depth, will consume about 0.13 GB/second of bandwidth just for the pixel refresh, this is not too big of a deal.

At 5GB/sec, that should be enough fillrate to hit every pixel at 1024x768 res (assuming 24 bit color and 24 bit Z or a total of 6 bytes per pixel) over 1000 times in a second. Oh, but texture fetch bandwidth will eat into that too, so that is a very loose back-of-the-envelope number or "upper bound". Real games have more work to do than just sitting in a loop repeatedly erasing the frame buffer etc; they also tend to hit each pixel in the scene more than once.

(*) assuming the RAM "dual channel" mode is active and the DIMMs are 64-bits wide each, yielding a 128-bit wide memory system at 667MHz peak throughput.

Bottom line is that the new Mini is probably at its best for GPU speed when you have matched DIMMs installed. The base config has matched DIMMs but only 512MB total; we'd recommend 1GB minimum due to the shared RAM configuration.
Note however that they have not yet tested the Intel Mac mini.