You are only looking at a short-term perspective of the consumer winning on price/performance temporarily. If the firm cannot sustain this strategy financially or gain substantial market share using it, it's a failure long-term.
This strategy ONLY works until your competitor catches up or beats you in perf/watt at which point their 550-600mm2 die chips and 400mm2 chips pummel your 325-350mm2 chips into the ground. Secondly, selling small- to mid-die chips against superior large die competing products means lower average selling prices, lower margins, and worse brand equity.
I am well aware of all the NV and AMD launches in the last 15 years. It was only a resounding success for the consumers but for the GPU market as a whole overall, it was a horrible outcome because it meant once NV caught up and exceeded AMD in perf/mm2 and perf/watt, they could charge flagship prices for mid-range chips = this all started with Kepler. Kepler proved AMD's small-die strategy was not sustainable long-term.
There is no way selling HD4870 for $299, HD5870/6970 for $369-379 produced high margins for AMD. NV sold 256-bit small die chips for $500 with GTX680, and $550 for GTX980. Those are high margins mid-range chips. Also, the small die strategy has no contingency plan. If you are late with a mid-range die, you lose the ENTIRE market until next generation. With the large die strategy, even if your large die chip is late, you still have solid low-end and mid-range chips.
And you seem to not get this point at all - the main reason AMD hang on to market share in the HD4000-6000 generations was because NV had much more inefficient perf/watt and perf/mm2 architectures. Once NV fixed that, it's game over for that AMD strategy.
-> If your large die chip is 500-600mm2+, then you can build mid-range 350-400mm2 chips. If AMD's flagship chip is not even 400mm2, how do they expect to even compete in the highest segments?
They won't.
HD4000 vs. GTX280 generation
HD4870 =
256mm2
HD4890 =
282mm2
vs.
GTX280 =
576mm2
HD4870 was only as fast as NV's 3rd top card - GTX260 216.
AMD's small die strategy didn't work in terms of profits or large market share gains. They
started off with 38% market share vs. 61% for NV, but market share collapsed as low as 31% and went up as high as 43.8%. By the beginning of HD5800 generation, AMD was at about 40% to 60% for AMD, or barely a
2% increase after releasing a $199 4850 and $299 4870. Ouch. :hmm:
HD5000 vs. GTX400 generation
HD5870 =
334mm2
vs.
GTX480 =
526mm2
HD5870 did beat GTX470 initially but with time, lack of geometry performance meant a GTX470 OC would crush a 5870. AMD started off that generation at about 40% and by the time Barts (6800) and Cayman (HD6900) came out, AMD had 38.8% market share. OK so going from HD4870 with 38% market share all the way until the end of HD5870 generation, AMD was still at just 38.8% market share. Are you starting to see a trend now? And all this time AMD is not making a lot of $ on GPUs.
During this generation, HD5850 significantly undercut NV's 470 and 5870 the GTX480. It took 9 months for GTX460 to come out to match a 5850 but AMD
still lost market share. :thumbsdown: AMD basically beat NV by 6 months with GTX470/480 and 9 months with GTX460 and below. This huge head-start, and amazing price/performance didn't
even make a dent in the market share by the end of the Fermi vs. HD5000/6000 generation; and neither did this make AMD a lot of money!
HD6000 vs. GTX500 generation
HD6970 =
389mm2
vs.
GTX580 =
520mm2
This is where AMD is starting to realize you can't have your cake and eat it too. You can't magically perform next generation DX11 effects like geometry/tessellation and global illumination for "free". You need to grow your die size with more functional units of you'll get run over.
Shockingly, despite HD6950 selling for $299 and unlocking to an HD6970, basically undercutting GTX570 and having the option of effectively almost buying dual 6950s for the price of a single 580 1.5GB, AMD doesn't gain market share.
At the start of that generation AMD was at 38.8% and by HD7970 launch, AMD is at 37.8%,
another 1% loss of market share. :thumbsdown:
All this time a gamer could use HD4000/5000/6000 for mining which made those upgrades all free.
HD7000 vs. GTX600 generation
HD7970/7970Ghz =
352mm2
vs.
GTX680 =
294mm2
Now, the turning point. Finally what many have said would become true did. Once the competitor catches up or beat AMD on perf/watt and perf/mm2, AMD's small die strategy would come crashing down. AMD even beat GTX680 to launch by 2.5 months, and even after GTX680 came out, AMD recaptured the performance crown as of June 2012 and held on to it until the Titan. HD7770/7850/7870/7950 all beat every single Fermi GTX500 chip in perf/watt and VRAM and performance but still failed. NV easily coasted until GTX660Ti and below launched 6+ months later.
AMD started this generation with a market share of 37.8%, and it went down as low at 33% (!). By the time R9 290/290X launched in Q4 2013, AMD's market share
fell to just 35%, or
another 2.8% fall in market share from the end of the HD6970 generation. :thumbsdown:
Recall the green market share number of 38% - that's where AMD started off with HD4870, and by the end of HD7970Ghz generation, they ended with just 35%. Therefore, the small die strategy absolutely failed what it set out today -- it failed to make $ and it failed to gain market share. :\
You pretty much know today the market share is close to 70%/30%, which means AMD lost even more since R9 290/290X.
Data:
This is not reality. It only had the market cornered from a price/performance point of view but since most of the market didn't respond to price/performance, the strategy didn't work. I compiled market share numbers for you but I won't spend time compiling the financial because you know AMD's GPU division performed horribly for the last 4 years.
A lot of people must have paid extra for NV because AMD's price/performance & small die strategy did not result in market share conversion long-term. Those are the facts. NV's large monolithic die strategy prevailed, despite NV being late with both Fermi and Kepler.
NV's strategy is nothing like AMD's. AMD's focus is on GPGPU compute, OpenCL and small to mid-size die chips, using dual-chips for uber flagships. HD5870, 6970 and 7970 all had a massive amount of double precision compute, not just gaming performance. NV has tried all it can to make efficient gaming cards and has continued to neuter compute performance to maximize gaming efficiency. Maxwell just barely caught up to GCN in some areas of compute and that's only because it's 1 full generation ahead of GCN, but NV's DP ratio is now usually 16:1, 24:1 or even 32:1 vs. SP. Even their Kepler products had a fixed function compute unit which used way less power.
NV only continues to grow the die size of their flagship chips, and it's now able to use a mid-range chip as "flagship" for 2 consecutive generations because they have surpassed AMD in perf/mm2 and perf/watt. GK104 and GM204 are priced at $500 and $550 because gamers keep paying those prices for next-gen mid-range chips and because NV is now so far ahead of AMD in perf/watt and perf/mm2.
This was not the original AMD small die strategy. NV simply releases the smaller die chips first as flagships because of how much more ahead they have gotten in terms of efficiency. However, they have their high-end GK110/GM200 series of cards on top of the mid-range. AMD's mid-range strategy never had any large die chip sitting above; instead they used a dual-chip card to fill-in that niche.
These 2 sentences are almost contradictory. The whole point AMD should go back to large monolith dies (if financially possible) is exactly because you can't use 2 smaller dies to try and fight the uber-high-end market segment. Not only will AMD improve its brand equity, it would be able to raise prices, increase profit margins, and provide competition so that we are not being sold $550 mid-range next gen chips. That's why a 500mm2+ die is the best thing to happen to the GPU industry if true. It would mean having the old ATI-style mentality of wanting the single chip performance crown back.
Also, another consequence of having large die chips other than strong pricing competition, is we as gamers get a lot more performance. Imaging taking a perf/watt efficient Maxwell, Volta, Pascal architecture and extrapolating that performance to a 550-600mm2 chip. If both AMD and NV are onboard for superior perf/watt and perf/mm2, we will get huge performance increases in the next 5 years and if AMD executes on time, hopefully next gen mid-range chips won't be sold at $500-550. Although I think it's very likely that both AMD and NV will bifurcate a generation from now on since yields on newer nodes are going to get worse and well financially it's more beneficial for both of those firms to have 2 flagships every 2 years since it makes them more $.
Maybe it's going to become this new trend:
1st revision on a new node = 350-400mm2
2nd revision on that node = 500mm2
3rd revision of that node = 550-600mm2
NV is more likely to be gunning for a 500mm2 earlier on since they have a proven track record of great execution with large monolith die chips and AMD is only now supposedly making their first such chip!
Based on all the information I've read so far, I am inclined to say yes. I've never read any rumour of R9 390X card being anything other than a 4GB HBM product.
Right now we don't have any info to say if AMD will use HBM for laptops or not this round. I am going to say I just don't know than guess. My point was that I didn't believe AMD would having nothing better than R9 290M/295M Tonga but I don't know if they will continue with GDDR5 or HBM for 2015 high-end laptop GPUs. I don't have any info to support either case.