We first observed differences between the Radeon R9 290X cards that AMD sent out for review and the ones being sold online just before our R9 290 coverage went live. After additional testing, we have answers, feedback from AMD, and more questions.
Leading up to the Radeon R9 290X launch, I had two cards in the lab, both supplied by AMD. They ran at slightly different core frequencies in single-card configurations. No big deal, right? That was just PowerTune doing its job. But paired up in CrossFire, even with plenty of space between them, heat build-up forced the pair to drop to even lower clock rates. Right away, I was able to establish that AMD’s Hawaii GPU operates within a range of clock rates, determined by a number of variables. For R9 290X, that range starts at 727 and ends at 1000 MHz. As an aside, I do have an issue with vendors simply advertising this as a 1000 MHz GPU.
The discussion that so badly requires clarification, however, is a derivative of Hawaii’s inherent behavior. Because the GPU is always trying to run as fast as possible, and then adjusting down to obey certain power, temperature, and fan speed settings, every card is going to be just a little bit different. This is expected, and applies also to Nvidia’s GPU Boost-capable cards. But when I got my hands on a retail card purchased from Newegg, it was way slower. Like, 20% slower in many cases.
So that left me wondering: were the press boards just lower-leakage parts able to sustain higher clocks? Without a definitive answer, and with AMD insistent that my results couldn’t be right, I approached the Radeon R9 290 review much more cautiously, presenting data from both the sampled and store-bought 290X cards.
A Break In The Story
Then, on Tuesday, I received word from AMD that it had a smoking gun.
All of my numbers were generated using Quiet mode—AMD’s name for its 40% PWM ceiling, designed to keep acoustics in check. That’s a signal representing a percentage of maximum voltage into the cooling fan. So, from one card to another, 40% should yield roughly the same fan speed. Variance in fan speed is limited to the mechanical components, and should be very small.
But as it turns out, 40% on the cards AMD sent and 40% on the retail cards do not equal the same fan speed. To confirm, I compared the press card to a retail boards from Sapphire and Asus, using our Battlefield 4 benchmark at 2560x1440 and Ultra detail settings.

Right out of the box, the Asus card tops out around 1850 RPM, the Sapphire card peaks just under 1950 RPM, and the press board is sitting around 2050 RPM. “Big deal,” you say. “It’s a couple-hundred rotations per minute.” Well, let’s see how that corresponds to operating frequency.

At 1850 RPM, the Asus card cannot keep Hawaii cool enough, and the GPU hovers at exactly 727 MHz. In fact, at times, 40% isn’t even enough, which is why you saw it bump up in two spots on the previous chart—the card has to violate its 40% ceiling in order to stay within spec at 727 MHz.
Just 100 RPM more makes a massive difference. Rather than getting stuck at 727 MHz, Sapphire’s board averages 809 MHz through our 100-second Battlefield 4 run. Add another 100 RPM, and the press card yields an average of 917 MHz.
How does this relate to average frame rate?

And there we have it. The card that AMD sampled to press is 18% faster than the slowest retail sample in our lab, an Asus board. It’s nine-percent faster than the Sapphire card, which falls roughly in the middle.
Obviously it is impossible to produce milions of GPU dies in same quality.
Due to marketing all producers sending best possible samples to press for reviews.
But at the end of the day producers need to sell also the not 100% perfect chips with maximum margin.
1st way is lockout part of the chip/lower clock rate (aka 290x->290 same on NV side) and sell it as cheaper model.
This is fair bacause customer know what is paying for.
2nd way both NV and AMD started recently is dynamicaly modulate chips operational parameters to the edge of thermal limits.
Combined way SW frekvency trotling and/or fans speed modulation.(NV BOOST &AMD power tune).
IMHO 2nd is quite customers cheating.Average ordinary customer do not spend hours researching product details to make informed purchase decision.
All shops advertise just one frekvency with very small wording at the bottom about possible operation variations.
I would feel cheated if AUDI sell me car marked 2,5L/100KW motor but my piece whould be running only 90 KW in the summer.Just because colling is undersized=cost reduced -> not able to maintain advertised/payed performance level.
However I still don't understand why there is a 6℅ difference even after FAN speeds were equalised? Considering that all fans have the same RPM I would expect all of the GPUs to have the same average frequency but press samples still manage to outperform. Could thermal interface between die and heatsink be a reason?
It would be great to see temperature and power consumption graphs in addition to RPMs and frequencies. We know that PowerTune takes all of these into account when controlling RPM and frequency.
However I still don't understand why there is a 6℅ difference even after FAN speeds were equalised? Considering that all fans have the same RPM I would expect all of the GPUs to have the same average frequency but press samples still manage to outperform. Could thermal interface between die and heatsink be a reason?
It would be great to see temperature and power consumption graphs in addition to RPMs and frequencies. We know that PowerTune takes all of these into account when controlling RPM and frequency.
I have no pity for them, since they had no respect for COOLING & NOISE.