ATI HD3850 & HD3870: Improved Acoustics & Power Efficiency

The Radeon HD3850 and HD3870 graphics cards finally put ATI in a competitive position against nVidia offerings in the all-important midrange market. They’re also said to offer much improved power efficiency. We explore just how much, and the all important (for us) related issue of noise.

March 10, 2008 by Lawrence
Lee

The ATI Radeon HD 3850
and Palit Radeon HD 3870 were launched late last year to near-unanimous applause from hardware reviewers. It’s the first time in recent memory that ATI has produced cards obviously competitive with nVidia offerings in the same price category. The midrange price category ($150~$200) of these cards may actually be more important than the high end, as it represents where the bulk of DIY graphics cards are sold. The HD 3800 series is the first to be DirectX 10.1
compliant and is geared towards 3D performance.

Both of our review samples use the stock ATI reference
coolers. Normally, gaming cards equipped with stock coolers would be dismissed as
inadequate for a well-cooled silent PC, at least not without significant modification.
However, these cards are not the power guzzling behemoths of GPUs past —
ATI is claiming lower power consumption, lower temperatures and
quieter fans. If this holds true, it may be possible for us to have our gaming
cake and eat it too.

We’ll also be looking at how well ATI’s Unified Video Decoder (UVD) performs
in a relatively low-end system (by modern standards). UVD is a piece of hardware
incorporated into the GPU core that helps offload a significant portion of video
decompression of H.264 and VC-1 content, allowing for relatively low CPU usage
during playback. UVD is available on all of ATI’s HD series graphics cards except
for the HD 2900 line.

ATI’s marketing emphasizes the HD 3800 series’ video playback capabilities,
low power draw, and cool / quiet operation. The two cards utilize the same
core though the HD 3870 is clocked significantly higher: 775Mhz vs. 670Mhz core/shader
frequencies, 2250Mhz vs. 1660Mhz memory frequency, and 72GB/s vs. 53GB/s of
memory bandwidth. Our HD 3850 sample has half the video memory (256MB) of the
HD 3870 but a 512MB variant is also available. Both are fairly lanky cards,
approximately 23.15cm long from one edge of the PCB to the other.

RADEON HD 3850 256MB VISUAL TOUR

Our sample came straight from ATI bubble-wrapped by itself, without any accessories or retail packaging, both of which will surely vary somewhat depending on the brand and price.

TEST METHODOLOGY

Our test procedure is an in-system test, designed to:

1. Determine whether the card’s cooler is adequate for use in a low-noise system.
By adequately cooled, we mean cooled well enough that no misbehavior
related to thermal overload is exhibited. Thermal misbehavior in a graphics
card can show up in a variety of ways, including:

• Sudden system shutdown or reboot without warning.
• Jaggies and other visual artifacts on the screen.
• Motion slowing and/or screen freezing.

Any of these misbehaviors are annoying at best and dangerous at worst —
dangerous to the health and lifespan of the graphics card, and sometimes to
the system OS.

2. Estimate the card’s power consumption. This is a good indicator of how efficient
the card is and will have an effect on how hot the stock cooler becomes due
to power lost in the form of heat. The lower the better.

3. Determine the card’s ability to play back high definition video, to see
if whether it is a suitable choice for a home theater PC.

Test Platform

• Intel
  Pentium D 930 Presler core processor. Official TDP of 95W.
• AOpen
  i945Ga-PHS motherboard – Intel i945Ga Chipset; built-in VGA.
• Gigabyte
  G-Power 2 Pro heatsink, modified with a Scythe
  Slip Stream 500RPM 120mm fan.
• Corsair
  DDR2 RAM, PC2-6400, 1024 MB
• Seagate Momentus 5400.3
  160 GB, 5400RPM, ATA/100, 2-platter notebook hard drive, suspended
• Seasonic S12-600
  ATX12V v2.0 compliant power supply, modified with a Scythe
  Slip Stream 800RPM 120mm fan @ 5V.
• Antec P180B case,
  modified in detail below.
• Nexus 120mm
  fan connected to a variable fan speed controller.
• Microsoft
  Windows XP Professional operating system
• ATI
  Catalyst 8.2 graphics driver
• QuickTime
  Alternative 1.81 codec package

Measurement and Analysis Tools

• ATITool
  version 0.26
  as a tool for stressing the GPU and to show GPU temperature
• CPUBurn
  P6 to stress the CPU
• SpeedFan
  version 4.33 to show CPU temperature
• Cyberlink
  PowerDVD 7 to play video.
• Seasonic
  Power Angel AC power meter, used to measure the power consumption
  of the system
• A custom-built variable fan speed controller to power the system
  fan
• Bruel & Kjaer (B&K) model 2203 Sound Level Meter. Used to
  accurately measure SPL (sound pressure level) down to 20 dBA and below.

The anatomy of our test platform is detailed here: Updated
VGA Card/Cooler Test Platform

Testing Procedures

Our first test consists involves recording the system power consumption using
a Seasonic Power Angel as well as CPU and GPU temperatures using SpeedFan and
ATITool (or just SpeedFan if a nVidia based card is used) during different states:
Idle, with CPUBurn running to stress the processor, and with CPUBurn and ATITool’s
artifact scanner running at the to stress both the CPU and GPU simultaneously.
This last state mimics the stress on the CPU and GPU produced by a modern video
game. The software is left running until the GPU temperature stabilizes for
at least 10 minutes. If artifacts are detected in ATITool or other instability
is noted, the heatsink is deemed inadequate to cool the video card in our test
system.

If the heatsink has a fan, the load state tests are repeated at various fan
speeds while the system fan is left at its lowest setting of 7V. If the card
utilizes a passive cooler, the system fan is varied instead to study the effect
of system airflow on the heatsink’s performance. A B&K Sound Meter is employed
to take system noise measurements at each fan speed.

Video Playback Testing

For our second test, we play a variety of H.264 and VC-1 video clips with PowerDVD. A CPU usage graph is created via the Windows
Task Manger for analysis to determine the approximate mean and peak CPU usage.
If the card (in conjunction with the processor) is unable to properly decompress
the clip, the video will skip or freeze, often with instances of extremely high
CPU usage as the system struggles to play it back. High CPU usage is undesirable
as it increases power consumption, and leaves fewer resources for background
tasks and other applications that happen to be running during playback. Power
draw is also recorded during playback.

Video Test Suite

1920×816 | 24fps | ~10mbps

H.264: Rush Hour 3 Trailer 1 is encoded with H.264. It has a good mixture of light and dark scenes, interspersed with fast-motion action and cutaways.

1440×1080 | 24fps | ~8mbps

WMV3: Coral Reef Adventure trailer is encoded in VC-1 using the WMV3 codec (commonly recognized by the moniker, "HD WMV"). It features multiple outdoor landscape and dark underwater scenes.

1280×720 | 60fps | ~12mbps

WVC1: Microsoft Flight Simulator X trailer is encoded in VC-1. It’s a compilation of in-game action from a third person point of view. While the source image quality is poor compared to the other videos in our test suite, it is encoded using the Windows Media Video 9 Advanced Profile (aka WVC1) codec — a much more demanding implementation of VC-1.

Estimating DC Power

The following power efficiency figures were obtained for the Seasonic
S12-600 used in our test system:

Seasonic S12-500 / 600 TEST RESULTS
DC Output (W)	65.3	89.7	148.7	198.5	249.5	300.2
AC Input (W)	87.0	115.0	183.1	242.1	305.0	370.2
Efficiency	75.1%	78.0%	81.2%	82.0%	81.8%	81.1%

This data is enough to give us a very good estimate of DC demand in our test system. We extrapolate the DC power output from the measured AC power input based on this data. We won’t go through the math; it’s easy enough to figure out for yourself if you really want to.

TEST RESULTS

#1 – Baseline, with Integrated Graphics: First, here are the results of our baseline results of the system with just its integrated graphics, without
a discrete video card. We’ll also need the power consumption reading
during CPUBurn to estimate the actual power draw of each individual card later.

#2 – Radeon HD 3850: The fan was very loud at bootup, but it quickly
throttled down to the point where it was below or at the ambient noise level
of our test system. Oddly, we could not coax the fan to spin up again, whether
automatically by stressing the GPU or toying with ATITool to do so manually.
Still, the cooler surprised us by keeping the GPU core temperature
at 89°C during extended load without artifacting or any other instability. Anything under 90°C
is acceptable, and the fact that it was inaudible amazed us, especially since
we were not impressed initially with the stock cooler design. Power consumption
was very low, indicating excellent efficiency — there wasn’t a lot
of heat to dissipate.

The fan problem may be specific to our test sample, as we know
of many users who are successfully controlling the fan on their HD 3850s. Actually
we don’t even regard it as a problem, since cooling was fine at the speed the fan was
stuck at.

#3 – Palit Radeon HD 3870: The fan control was automatic, but could also be over-ridden
with ATITool. Setting the fan between 0% and 30% resulted in the
same fan speed — what the card automatically used when idle. When stressed,
the fan vacillated up and down between 30% and 50%, trying to keep the GPU temperature
at around 90°C.

When we set the fan speed manually to stay at 30%, it was
not enough to keep the card cool. Artifacts appeared in ATITool as the GPU temperature
rocketed past 100°C. 40% seemed to be the sweet spot, keeping the GPU core
at 90°C with only a small increase in measurable noise (22 dBA). Incidentally,
90°C is how well our X1950XTX VGA heatsink test card’s stock cooler performs,
but with the fan speed at 43% and a noise level of 28 dBA.

The noise signature of the fan was poor. At 30% the overall noise
level was low, but the fan had a slight ticking and a low-pitched hum. At 40% fan
speed it registered 1dBA higher but the pitch of the sound also increased. At
50%, the ticking went away but the noise level became intolerable, thanks to
higher air turbulence and the motor running like a small vacuum cleaner.

40% is the minimum we would run the fan at, but at that speed, its 22 dBA noise
reading is deceptive as it’s abrasive rather than smooth. It’s too bad that ATI did not keep the heatpipe from the X1950XTX heatsink design or
increased the size and area of the fins and employed a fan with better acoustical
properties.

TEST RESULTS CONTINUED

Power

The power consumption of an add-on video card can be estimated by comparing
the total system power draw with and without the card installed on our test system. Our results
were derived thus:

1. Power consumption of the graphics card at idle – When CPUBurn is run on a system, the video card is not stressed at all, and stays in idle mode. This is true whether the video card is integrated or an add-on PCIe 16X device. Hence, when the system power under CPUBurn with just the integrated graphics is subtracted from the system power under CPUBurn with the add-on card, we obtain the increase in idle power of the add-on card. (The actual idle power of the add-on card cannot be derived, because the integrated graphics does draw some power — we’d guess no more than a watt or two.)

2. Power consumption of the graphics card under load – The power draw of the system is measured with the add-on video card, with CPUBurn and ATITool running simultaneously. Then the power of the baseline system (with integrated graphics) running just CPUBurn is subtracted. The difference is the load power of the add-on card. (If you want to nitpick, the 1~2W power of the integrated graphics at idle should be added to this number.) Any load on
the CPU from ATITool should not skew the results, since the CPU was running at
full load in both systems.

Both cards drew much less power than the 18-month old X1950XTX
but the HD 3850 stood out with less than half the consumption on load. Idle
power consumption was excellent, again especially the HD 3850. +11W at idle over the integrated graphics is no
small achievement.

Video Playback

The video playback was equal as expected, both cards taking full advantage
of UVD. It has been our experience that the core, memory, and shader speeds
do not affect playback. CPU usage during H.264 decoding was especially low —
at about 3% one could easily mistake the system as being idle. Considering the
test system uses an old dual core Presler processor, these results are excellent.

HDMI OUTPUT

The Palit Radeon HD 3870’s HDMI ouptut (via DVI adapter) offered all the
standard and widescreen resolutions between 800×600 and our monitor’s native
resolution of 1440×900. Unfortunately the image quality was poor — the
text was fairly blurry at all resolutions. It was completely unacceptable for
2D use. Video on the otherhand, looked satisfactory.

The audio functionality did not work at all. ATI claims the HD 3800 series
has its own 5.1 audio controller built into their cards (Palit lists it in its
specifications as well) and their DVI to HDMI adapter is designed to carry audio,
so it’s unclear what is the cause of this issue. It could simply be that Palit
unwittingly included a generic adapter that isn’t wired to carry audio.

MP3 SOUND RECORDING

This recording contains ~10 seconds of the video card
fan at each speed tested. The recording begins, ends, and is interspersed with
the ambient noise of the test system (with no video card installed). No recordings were made of our ATI HD3850 sample, as its contribution to the test system noise was not audible. (See #2 Test Results on page .)

• Palit HD 3870 at 30% (21 dBA@1m), 40% (22 dBa@1m),
  and 50% (26 dBA@1m): One meter (For best results, save to your own PC, then listen.)

This
recording was made with a high resolution, studio quality, digital
recording system, then converted to LAME 128kbps encoded MP3s. We’ve
listened long and hard to ensure there is no audible degradation from
the original WAV files. It represents a quick snapshot
of what we heard during the review.

The
recording is intended to give you an idea of how the subject
of this review sound in actual use — one meter is a reasonable
typical distance between a computer or computer component and your ear.
The recording contains stretches of ambient noise that you can use to
judge the relative loudness of the subject. For best results, set your
volume control so that the ambient noise is just barely audible. Be
aware that very quiet subjects may not be audible — if we couldn’t
hear it from one meter, chances are we couldn’t record it either!

FINAL THOUGHTS

Gaming: This is not our forté, so for an overall sense of how well the
HD 3850 and HD 3870 perform in games, we recommend the reviews
at The Tech Report, [Hard]OCP,
X-bit
Labs, and AnandTech.
The general consensus is that the HD 3850 represents the best value in mid-range
gaming performance today. The HD 3870’s extra horsepower helps for those
wanting to play at higher resolutions, and offers fair value to those who cannot
afford a Geforce 8800GT, which is said to do a touch better in some games.

Video Playback: The Unified Video Decoder built into the HD 3800 series
made video playback a breeze, blowing through our test clips with ease despite
our antiquated processor.

Stock Cooling: Both stock coolers were adequate in our opinion, but
the type of noise emanating from the HD 3870’s fan was a little too grating
for our liking. It was a huge improvement over previous dual-slot ATI coolers,
but it’s still a prime candidate for a third party heatsink such as the Accelero
S1. The HD 3850’s stock cooling surprised us — it was akin to a single slot passive heatsink as the fan was effectively inaudible in our
test system. We’d be happy with this card’s default acoustics in any PC. (We certainly hope its fan behavior is typical!)

Power Consumption: It’s very low for both cards, especially compared to the
previously top-of-the-line Radeon X1950XTX. During idle, the HD 3850 and HD
3870 added only 13W and 20W, respectively, to our IGP system’s AC power draw. Considering the 3D performance, those are
impressive numbers. No doubt there’s some automatic 2D undervolting
involved — of which we whole-heartedly approve! It’s about time we had
the equivalent of Cool’n’Quiet or SpeedStep for graphics cards. Perhaps as with CPUs a couple of years ago, a peak has been crested, and the industry will be moving towards high efficiency and lower power GPUs. The next generation or two of GPUs will tell.

Of the two cards, we highly recommend the HD 3850. For gaming at lower resolutions,
it offers solid performance, very low power draw, and it’s going to be completely inaudible
in most systems. The price is also very attractive, especially taking into account
ATI’s recent price cut on the entire HD 3800 line. The HD 3850 512MB variant
is now US$169 while the HD 3870 512MB price has been lowered to US$199. Rumor
has it an additional price cut for the HD 3850 is on the way. For gaming at
higher resolultions, the Palit HD 3870 is the better choice. It also presents
a good value, though its stock cooler is not quite up to our standards and
we advise replacing it if you are serious about silence. You should also seek other opinions about the Palit HD 3870’s HDMI performance.

Our thanks to ATI
and Palit for
the video card samples.

* * *

Articles of Related Interest
Arctic Cooling Accelero S1
VGA Cooler
Updated VGA Card/Cooler
Test Platform
Zalman VF1000 LED Graphics Card Cooler
Asus EN8600GT Silent/HTDP/512M Graphics Card
Asus EN8600GT OC GEAR graphics card

***

Discuss this article in the SPCR forums.