AMD launched the long-awaited mobile Kaveri line yesterday along with two new branding lines. You can now get mobile parts with coherent GPUs in both FX and Pro forms.
The idea is simple enough, take a standard desktop Kaveri aka Ax-7xxx CPUs, downclock them, then slap them in laptops. This more than likely took six months from January’s launch for laptop refresh cycle reasons, not technical ones. One surprise is the branding of the new CPU/APUs, they are in the traditional A6/A8/A10 brands plus two new ones, FX and Pro. The lineup looks like this.
The SKUs so far
On the consumer side there are two variants, standard 35W and 19W ULV TDPs, with three models in each class. The biggest change is that the bottom of the stack is an A8 instead of an A6 here and the top model is now called FX. It is an unusual move but, well, marketing. Standard voltage parts come with a P suffix, ULV versions don’t have that.
Professional parts have a different nomenclature once again, instead of Ax- these three CPUs have the Ax followed by Pro-7xxx. See marketing comment above. Two of the Pro parts, A10 and A8, pull 19W but the lowest A6 only draws 17W. One thing to keep in mind is that Kaveri has the ability to set TDPs arbitrarily in firmware so these listed TDPs are only maximums, you can set them lower for a thin 15W design if you want.
Consumer parts are pretty simple to explain, they are the mainstream devices. Professional lines are not technically different in any way other than clocks and TDP but they come with a different set of support and drivers. As you would expect from an enterprise oriented line of chips they have more management features enabled in software but the real hooks are driver stability and longevity.
Pro drivers are frozen for a longer period of time, the normal refresh rates for gamers are costly and annoying for corporate IT staff. IT admins don’t care if a business laptop may not have the absolute max frame rate in the new hot AAA game, they more care that they don’t have to patch the damn thing every three weeks. If you are not part of this world, you may scratch your head over the usefulness of driver freezes but it is very important in the pro sector.
Similarly the devices have a longer life span than normal consumer parts. This means they will be on sale longer and have spares in inventory far longer than the normal 12-18 months that consumer parts live for. It is AMD’s way of saying that if you buy a Kaveri Pro device, you can get another or parts for the 3-5 years that corporate PCs live for. OEMs won’t touch a chip for a business oriented laptop without this feature.
If you look over the lineup there are a few things that pop out, the first being that the PCIe lanes in the ULV models are all Gen2 rather than Gen3 and are only x8, not x16. Why? Power. PCIe3 pulls more than PCIe2 and 16x pulls more than 8x. Also a PCIe2 8x link is more than enough for anything other than a GPU, discrete GPUs don’t do well in constrained form factors that ULV devices enable. Some may quibble with that idea but for the allowable TDPs in ~20mm Z-heights, Kaveri should be on par with any discrete GPU. DRAM is slower here for similar reasons.
Next is the fact that only the top 35W part, FX-7600P is fully enabled with 4 CPU cores and 4 GPU clusters. Yield problems anyone? It isn’t a binning issue either, most of the devices have a similar max turbo clock in the low 3GHz range and GPU frequencies are clustered very tightly by wattages too so that is unlikely to be the culprit. Time will tell here but it looks like the process optimization learning curve is still being scaled on Kaveri.
One interesting side note is on width vs clock for power savings. If you look at Nvidia’s Maxwell (GM107) parts they claim massive gains in performance per watt. If you look at the data however there is precious little gain over process improvements and downclocking allowed by the massive increase in shader width and die area. In short Nvidia traded off frequency for size/width and took the gains as power savings. It is a marketing win, not a technical one.
Looking over mobile Kaveri it is pretty clear that AMD didn’t do the same, the ULV parts have the same four or six shader clusters enabled as the 35W parts, FX and A6 notwithstanding. We asked AMD why they didn’t use the full eight clusters on the ULV parts and clock them down a bit more and the answer surprised us. In theory it should save power but the sweet spot for the power/performance curve is where the Kaveri ULV GPU sits.
One look at the charts says that this is 533MHz or so, essentially the knee of the efficiency curve. If you go up you can burn a lot of power for minimal gains, the 35W parts only gain ~100MHz on GPU frequency but pull almost twice the power. CPU base clocks go up by about 25% too and contribute a bit to the wattage gain but the GPU seems to be the main draw, pun intended. In short if you enable the two extra GPU clusters you pay the leakage cost and can drop the clocks by a bit, but the payoff is small or negative. If you only enable four but up the clock, power draw goes up far more, six shader groups at ~533MHz is the sweet spot.
In the end Kaveri is late and lower performance than initially promised but has a lot of really solid features. The new 19W parts should enable AMD to fit in any of the currently in vogue form factors and configurable TDPs will smooth over any bumps. The mainstream parts should be solid enough for mainstream consumer uses and mid-range gaming, about the best you can get without a discrete GPU. The real question is how well the Pro devices will be received by business customers, it is a market where AMD has almost no presence so it could have a very big impact for the company. It will be interesting to watch.S|A
Have you signed up for our newsletter yet?
Did you know that you can access all our past subscription-only articles with a simple Student Membership for 100 USD per year? If you want in-depth analysis and exclusive exclusives, we don’t make the news, we just report it so there is no guarantee when exclusives are added to the Professional level but that’s where you’ll find the deep dive analysis.
Latest posts by Charlie Demerjian (see all)
- A big ARM server project shut down in silence - Mar 14, 2023
- A bit more on AMD’s Genoa memory issues - Mar 13, 2023
- Want to know when Qualcomm’s Oryon taped out? - Feb 16, 2023
- Things go bump in the night for AMD’s Genoa - Feb 15, 2023
- AMD’s Instinct MI300 is much more than a GPU - Jan 6, 2023