AMD Announces Their First 8000M GPUs
by Jarred Walton on December 17, 2012 4:21 PM ESTAMD held a press briefing today on their upcoming 8000M graphics chips, which they are calling the "second generation GCN architecture" parts. We’ll have more on that in a moment, but while we were expecting (dreading) a rebranding prior to the call, it appears we are at least partially mistaken; there will be at least one completely new GPU with 8000M. (If you want additional background material, you can see the previous generation high-end 7000M announcement from April 2012 for reference.)
I’m not going to get too far into the marketing aspects, as we’ve heard all of this information before: AMD has improved Enduro Technology, they’re continuing to improve their drivers, and APP Acceleration has a few more applications. There have been a few major titles released in the past couple of months with AMD Gaming Evolved branding (Far Cry 3 is arguably the most notable of the offerings, with Hitman: Absolution and Sleeping Dogs also scoring well amongst critics and users), and Bioshock Infinite is at least one future release that I'm looking forward to playing.
Cutting straight to the chase, at this point AMD has released limited information on the core specifications for some of their 8000M GPUs, but they coyly note that at least one more GPU announcement will be forthcoming in Q2 2013 (8900M by all appearances). Today is a soft launch of high level details, with more architectural information and product details scheduled for January 7, 2013 at CES. AMD did not share any codenames for the newly announced mobile GPUs, if you’re wondering, other than the overall family name of “Solar” for the mobile chips (replacing the outgoing “London” series), but we do know from other sources that the 384 core part is codenamed "Mars" while the larger 640 core part is codenamed "Neptune". Here are the details we have right now:
| AMD Radeon HD 8500M, 8600M, 8700M, and 8800M | ||||
|
Radeon HD 8500M |
Radeon HD 8600M |
Radeon HD 8700M |
Radeon HD 8800M |
|
| Stream Processors | 384 | 384 | 384 | 640 |
| Engine Clock | 650MHz | 775MHz | 650-850MHz | 650-700MHz |
| Memory Clock | 2.0GHz/4.5GHz | 2.0GHz/4.5GHz | 2.0GHz/4.5GHz | 4.5GHz |
| Memory Type | DDR3/GDDR5 | DDR3/GDDR5 | DDR3/GDDR5 | GDDR5 |
| FP32 GFLOPS | 537 | 633 | 537-691 | 992 |
| FP64 GFLOPS | 33 | 39 | 33-42 | 62 |
Obviously there are a lot of missing pieces right now, but what we immediately notice is that the core count on the 8500M/8600M/8700M means that we’re definitely looking at a new GPU. The only other time we’ve seen AMD do 384 cores is with Trinity, but that’s a VLIW4 architecture so we’re not seeing that again. Given the currently shipping Southern Islands chips (“London” on the mobile side) have 640 cores max for Cape Verde, 1280 max for Pitcairn, and up to 2048 for Tahiti, AMD has likely created a fourth SI derivative that drops down to two CU arrays, each with three CUs. (You can read more about the GCN/SI architecture in our earlier GPU coverage.) Performance is something of a wildcard with the new 384 core parts, and the choice of DDR3/GDDR5 memory will also influence the final result. We'll find out in the coming months how the 8500/8600/8700M stack up to NVIDIA's midrange "GT" offerings, which interestingly are also using 384 cores.
Also worth a quick note is that AMD is not discussing TDPs at this point in time—which is common practice for both AMD and NVIDIA. We expect the new "Mars" parts to be more power efficient than the outgoing Thames/Turks cores, thanks to the shrink to a 28nm process. However, AMD and NVIDIA typically stick to common power targets for laptops that are dictated by their OEM partners, which often means they'll play with clock speeds in order to hit a specific TDP. That's why all of the clock speeds listed in the above table have a qualifying "up to" prefix (which I omitted).
The final announced card is the one where we appear to have more of a rebrand/optimization of a previous generation chip. 8800M has the same 640 core count as Cape Verde/7800M, only with modified clocks this time. The earlier 7800M chips could clock up as high as 800MHz, so maximum core clock is actually down a bit, but they only ran the memory at up to 1GHz (4GHz effective) GDDR5. If AMD determined memory bandwidth was more important for that particular GPU than shader performance, the new 8800M would make sense. Also note that AMD isn’t including the boost clock speeds into the above chart; under the right circumstances, all of the new chips can run at higher clocks than the reference clock.
Radeon 7800M Left, Radeon 8800M Right
AMD isn’t calling the 8800M a rebrand, but we’re looking at the same core counts as Cape Verde and the same 28nm process technology, so we wouldn’t expect a substantial change in performance. There’s also the above chip shot as a point of reference. If the 8800M is substantially different from Cape Verde then the above images provided in AMD’s slides must be incorrect, as the new and old chips look the same. Minor tweaks to power use, caching, or other elements notwithstanding, we’re probably dealing with a die respin at most. But, there’s nothing inherently wrong with rebranding—AMD and NVIDIA have both been doing it for some time now. Don’t expect every “upgraded” GPU to be better; a 7400M isn’t faster than a 6700M, and likewise we expect 7700M and 7800M to be faster options than the 384 core 8500M/8600M/8700M and competitive with 8800M. Here’s a quick recap of the same core specs as above for the current 7700M/7800M parts:
| AMD Radeon HD 7700M/7800M Specifications | |||||
|
Radeon HD 7730M |
Radeon HD 7750M |
Radeon HD 7770M |
Radeon HD 7850M |
Radeon HD 7870M |
|
| Stream Processors | 512 | 512 | 512 | 640 | 640 |
| Engine Clock | 575-675MHz | 575MHz | 675MHz | 675MHz | 800MHz |
| Memory Clock | 1.8GHz | 4.0GHz | 4.0GHz | 4.0GHz | 4.0GHz |
| Memory Type | DDR3 | GDDR5 | GDDR5 | GDDR5 | GDDR5 |
| FP32 GFLOPS | 589-691 | 589 | 691 | 864 | 1024 |
| FP64 GFLOPS | 36.8-43.2 | 36.8 | 43.2 | 54 | 64 |
I’ll refrain from commenting too much more about performance of an unreleased part, but AMD indicated their 8870M should be substantially faster than NVIDIA’s current GT 650M GDDR5 (which isn’t too surprising considering clocks and core counts), and the 8770M should likewise be a healthy 20%+ bump in performance relative to the 7670M. I’d rather see comparisons with GTX 670MX and HD 7770M, respectively, but I suspect those wouldn’t be quite as impressive. Anyway, you can see AMD’s comparison charts in the complete slide deck gallery below. Availability of the new GPUs is slated for Q1 2013.
Facebook
Twitter
RSS
88 Comments
View All Comments
SetiroN - Monday, December 17, 2012 - link
It wasn't spelled out but to me it was pretty clear 8500/8600/8700M series are GCN parts in contrast with the older lower-end 7xxxM parts, which were a rebrand themselves.What should be made more clear is how *ABSOLUTELY RIDICULOUS* this all is.
After Nvidia rebranding Fermi parts to make them look like Kepler 6xxMs, now probably every single one of the new AMD mobile lineup is going to have lower performance than the respective 7xxxM SKU.
Although we don't know the details yet, if they're comparing 7670M to 8770M and 7590M to 8690M it is clear that they're pulling off the same 58xx->69XX stunt. It seems pretty clear, by projecting what the gimped&rebranded 8800M is going to be.
I could look the other way when it was limited to entry level SKUs, but rebranding relatively high-end 78xxM GPUs to 88xxM while also substantially lowering their performance (don't tell me 500 more MHz on the RAM are ever going to make up for 100-150MHz on the core, this is all done to lower power consumption) is absolutely unacceptable. They're trying to make people think "hey look, not only the 8800M is newer, it also has a lower TDP!" as if they were completely new generation parts.
Also I can't wait to see 8700M parts clocked as low as 8500M, as the graph suggests it will be possible. Disgusting.
mrdude - Monday, December 17, 2012 - link
It costs a pretty penny for a wafer at 28nm. If the yields are good, and all signs show that they are, it makes little sense for either company, nVidia or AMD, to make an entire product line using the architecture and process.Why bother to bin something like a GT610m or 7490m if it's going to sell for pennies anyway?
This is approach isn't seen only in the laptop area but in the same low end discrete desktop GPUs as well.
But this isn't a rebranding at all. In fact, this looks more like a refresh of the entire mobile lineup that was previously rebranded. nVidia will likely follow suit as well and hopefully we'll see both companies on 28nm throughout all of their discrete GPUs for both mobile and desktop - assuming the others reach EOL (end of life) sometime soon.
So what's up with the tone? and not just you, but Jarred as well. If you're sick of refreshes and rebrands, this is exactly the type of news you should be happy to see...
"We’ll have more on that in a moment, but while we were expecting a rebranding prior to the call, it appears we are at least partially mistaken—you can see the previous generation high-end 7000M announcement from April 2012 for reference."
You weren't partially mistaken, you seem to be completely mistaken.
JarredWalton - Monday, December 17, 2012 - link
Which part of the discussion of 8800M begin a rebrand did you miss? So everything 8700M and lower is using a new GCN chip, but at the same time we've just compressed the performance difference between 7500M, 7600M, and 7700M. The 8500M series is 384 cores at 650MHz (maybe lower on some models), the 8700M is 384 cores at up to 850MHz, so the range of performance is now about 30% increase.The second part of the discussion is that we're looking at new parts at the lower end of the market, so no that's not nearly as exciting as real improvements in the midrange and high-end markets. One performance comparison shown in the slide deck of AMD vs. AMD is 8770M vs. 7670M, rather than 8770M vs. 7770M. It looks like the 8770M is around 30% faster, give or take, but the 7770M is 512 cores clocked at 675MHz while 7670M is 480 cores clocked at 600MHz. So, 7770M is around 20% faster than 7670M, and 8770M would thus only be a ~10% upgrade to 7770M.
Another comparison is 8690M vs. 7590M, where 8690M looks to be about 20% faster in most games. Problem is, 7590M is 480 cores at 600MHz while 7690M is 480 cores at 725MHz -- that's a 20% difference right there! In my book, you compare the same family to show improvement, so 8690M goes up against 7690M, 8770M goes against 7770M, and 8590M would go against 7590M. That's not what AMD is showing right now.
Granted, these are all preliminary results and we won't have hardware for a month or more I'd guess. We don't know power or other aspects of performance that are important either. What we really need to know to judge value and performance is: how does 8870M compared to 8770M, how does 8870M compare to GTX 660M and GTX 670MX, etc. Right now, we only know a few base specs, and I'm not going to spend too much time praising an apparently new GPU at the low end of the market. The 384 core AMD parts are going to be a nice addition for budget laptops, but until we have hardware and independent benchmarks in hand there's not much to say.
lmcd - Monday, December 17, 2012 - link
I'd like to reference my above reply: it's irrelevant. It really is. Enthusiasts buy GPUs in a competitive market, where everyone looks at the GPU and overall price tag. Noobs buy looking at the CPU, price tag and MAYBE the AMD graphics sticker (though usually they don't look) requiring AMD to be just as competitive with pricing as if they made their model numbers match up with where you'd like them.Also, it's beneficial that they're compressed: we no longer have crap (equivalent to) the GT 610, HD 7350 and other junk that barely deserves to be on a CPU die, let alone one dedicated to a GPU.
It means AMD is waking up to Intel's improving iGPUs, too. That's equally important; AMD's marketshare depends heavily on beating Intel's iGPUs.
lmcd - Monday, December 17, 2012 - link
*compressed rangeAlso, half of these processors go onto machines with 1366x768 screens where most if not all are adequate for light-moderate gaming.
mrdude - Tuesday, December 18, 2012 - link
So let's get this straight...You're being pissy because AMD is providing incremental upgrades and bumping up the entire lineup to 28nm?
"Right now, we only know a few base specs, and I'm not going to spend too much time praising an apparently new GPU at the low end of the market. "
And when a majority of discrete GPUs in laptops are low end anyway, where does that leave us?
Right. The same place we started. You being completely unaware of what this new generation of products is meant to replace. Instead of realizing that a vast majority of mGPUs are sub GT660m - FFS, Jarred, have you seen the amount of Kepler rebrands "new" laptops have? These NEW GPUs are meant to replace everything that was lower than a 7730m, and even the first gen GCN parts might get a small architectural upgrade.
If you're sick of rebrands and want more efficient GPUs, whether it's lower end or not, this is the most welcome news we've seen in 2-3 years for mobile GPUs from AMD. And, as a writer for Anandtech, you COMPLETELY missed that point.
So well done. Really. Spot on article. While BSN* and TR actually at least partially addressed this point, you seem to ignore it altogether and go to the comments to defend your stance.
mrdude - Tuesday, December 18, 2012 - link
Sorry, that should read Fermi rebrands. Anything sub 640m is a Fermi product, and in fact even some post 640m are Fermi.mrdude - Tuesday, December 18, 2012 - link
Here are some actual articles worth reading that spell out exactly what these parts are meant to do and what Jarred felt wasn't important enough to mention:http://www.extremetech.com/computing/143579-amd-an...
http://www.xbitlabs.com/news/graphics/display/2012...
JarredWalton - Tuesday, December 18, 2012 - link
Just to be clear, you completely ignored all my commentary on the fact that the only benchmarks AMD has provided show performance comparisons between different classes -- 8690M vs. 7590M and 8770M vs. 7670M. We also have the lopsided HD 8870M vs. GT 650M comparison, which is like comparing the desktop HD 7870 with a desktop GTX 650 Ti and crowing about how AMD is over 50% faster. That's a big part of why this announcement isn't really impressing me, and until we know three things there's not much more to be said:1) We need to know the real performance numbers under a wide range of games.
2) We need to see how much the laptops with the various GPUs cost.
3) We need to know how much power the chips draw under various loads.
Until we know those items, to speculate and say that they're substantially faster and/or more power efficient is taking things too far.
Furthermore, regarding power use AMD and NVIDIA have been targeting (with their OEM partners) roughly the same mobile GPU TDPs for years: ~35W, ~45W, 60~75W, and ~100W (with that last being a more recent addition). Anything less than 35W is typically reserved for the stuff that no one actually recommends -- unless you think GT 620M and HD 7470M are great products? Right now, there's no reason the new parts would open up a lower power class than what we already have. I expect the new 384 core GCN parts will go head-to-head against NVIDIA's GK107 parts, but at the same time I don't expect to see any major improvements in overall gaming performance.
More important with regards to power, nearly all laptops are now using switchable graphics, so I don't really care how much power a dGPU uses at idle. It's either running full tilt to play a game -- in which case I'm plugged in and want as much performance as I can get -- or it's powered off and HD 4000 (or the appropriate Trinity iGPU) is doing the work. So at the lower end of the performance range we have GT 640M/650M/GTX 660M going against HD 8500M/8600M/8700M, and probably trading blows depending on the title. Competition is good, and NVIDIA really needs some in the mobile sector right now as they seem to be outselling AMD by a huge margin in laptops (until you add in numbers from Apple, but that's not entirely fair as Apple is a separate class and has been swapping between AMD and NVIDIA on a regular basis depending on who gives them the better deal), but going back to my three points above, it's way too early to be crowning a victor.
Finally, there's speculation -- and only speculation -- that the new GPUs will get into Ultrathins and Ultrabooks. I'm highly skeptical that will really matter, because right now I can name two Ultrabooks with dGPUs: ASUS UX32A/UX32VD, and Acer TimelineU M3/M5 -- and if you really want to stretch the definition, you have Razer's "first real gaming laptop" Blade, which is full of hyperbole, or the Apple MacBook Pro 15 Retina. Everything else is iGPU only, mostly because the majority of people aren't really playing games on Ultrabooks/Ultrathins. Plus, even if the new dGPUs use 25W, that's more than double what the rest of the system would use, and we already run into throttling and heat issues on many Ultrabooks with just a 17W CPU/APU. Considering the only viable and affordable gaming Ultrabook is a 14" Acer M5 that really pushes the limits of Ultrabook (it weighs about 4.5 pounds, for instance), I don't know that we'll see any real gaming class Ultrathin/Ultrabook laptops until the next process node -- which will bring Haswell and 20nm GPUs into the picture.
So go ahead and call this "the most welcome news we've seen in 2-3 years for mobile GPUs from AMD". I did't completely miss that point; I simply disagree with your (IMO unwarranted) optimism. I took a more or less neutral tone, as we don't have the information necessary to take a stronger stance. We have a Johnny-come-lately smaller Southern Islands GCN part, and we have a rebrand. If that's the most exciting news in 2-3 years, that speaks more about the unimpressive execution of AMD's mobile GPUs during the past 2-3 years than anything. And really, I'd say Turks at its launch was at least as impressive as the Mars "Solar System" GPU announced here.
mrdude - Tuesday, December 18, 2012 - link
Jarred, the price points are pointed out in your slides. Apparently you missed it:http://images.anandtech.com/doci/6520/slide-28_575...
The 8800 is competing with the 7700, and the 8600 is actually competing with the 7400 40nm parts as far as pricing and segmentation go.
We don't know the performance figures, but considering every GPU in recent history has been 10-30%, I'd say it's pretty darn safe to conclude that this will be along those lines. Furthermore, given the segments of the market these are intended to, we can also deduce what they'll likely cost as well. This isn't Bulldozer we're talking about, this is ATi. They're not going to backwards, Jarred. We know how GCN performs and we know the performance of the older 40nm parts. Is it really that difficult for you?
What we CAN conclude with near certainty is:
- they'll perform better than the 40nm parts. I think to assume otherwise would be asinine.
- The perf-per-watt will be substantially better.
- The parts will likely cost the same (maybe a bit higher) than those they replaced. AMD isn't going to refresh their entire lineup and move it to GCN + 28nm if they don't plan on it selling, which means reasonable pricing.
Despite what you think, the low end GPUs outsell the upper tier stuff by a very large margin, Jarred. Opting to ignore the improvements here - and you can argue % all you want, but I think we both KNOW that we'll see improvements in every single metric - is ridiculous.
And that's what you did. You ignored it, and then went to the comments to defend why you ignored it. The reason? Because apparently nobody gives a damn about lower/mid tier GPUs despite the fact that they outsell the other high end stuff many times over.
I don't care about AMD. They're not going to survive the next two years anyway. What I do care about, though, is that this article was exponentially worse than the others I've read on other sites and that this article included some seriously backwards thinking.
Only the high end matters?
Christ, Jarred. Get real.