AMD’s Past and Future CPUs (Updated with Author’s Note)

Less than half a decade ago, if you had walked up to someone in the industry and said that in 5 years’ time AMD would have the fastest CPUs, you would have been laughed straight out of the room—but here we are.

At the time of writing, AMD does have the fastest CPUs on the market with Zen 3, so let’s see where AMD is heading in the future.

The Ugly Past

It’s not a controversial statement for me to say that the Bulldozer family of CPUs was a failure. Not just a commercial failure in a period that eventually saw AMD nearing bankruptcy, but also a technical failure. There are however parts of the Bulldozer core that are good such as the FPU and branch predictor unit. Weaknesses in its shared front-end and poor cache/memory systems meant a 4-module Bulldozer CPU lost to a 4-core Sandy Bridge CPU in most workloads, despite the higher clocks. Insult to injury for AMD was the the fact that a 4-module Bulldozer CPU also consumed more power than a 4-core Sandy Bridge CPU while being slower—in short the architecture was almost universally panned.

However, the failure of Bulldozer was not all because of poor CPU design. A lot of Bulldozer’s failures can be traced to its underperforming 32nm node from what was AMD at the time, before its fabs were spun off into Global Foundries. Broadly, the original design goals for Bulldozer were 33% clock speed lift over Phenom on the same 45nm node with no IPC decrease and a release date in 2009. As history shows, this did not happen and so nearly two years later in October 2011, Bulldozer became the product we all know it as: late, hot and slow.

Thankfully for AMD, when the original Zen architecture released in Q1 2017 it finally showed they weren’t out of the CPU game. Zen 1 had an extremely bullish 40% IPC increase design goal over Excavator and ultimately delivered an unheard of 52% IPC gain. AMD was back. Following up with Zen 2 and Zen 3, though they were not the same monstrous uplifts as Zen 1, they represented a respectful 15% and 19% IPC gain over their predecessors respectively. But y’all know that already, so let’s talk about the future of AMD’s CPUs.

A Bright Future

AMD looks poised for even more continued success with Ryzen, assuming of course that the information and sources that both I and the rest of Chips and Cheese’s staff have access to are accurate. Please take all the following information in this section with the usual cautionary dose of salt. While both the Chips and Cheese team and I have full belief that this information is accurate, that does not mean that this is 100% confirmed and you should not take this as unsalable truth. With this disclaimer out of the way, on to the juicy bits: Zen 3+, Zen 4 and even Zen 5.

AM5 & DDR5 – New IOD and Future Nodes

Zen 3+ looks to be a small IPC gain on base Zen 3, having been told “It’s more than Zen+ was [over Zen 1] but not much” which I interpret to mean around a 4 to 7% IPC gain along with customary clock gains moving to the smaller N6 node from TSMC. N6 is a variant of N7 using 5 layers of EUV and is not a true “new node”, more of a refinement. However, the most interesting thing to me is that Zen 3+ on desktop may be the first AM5 CPU. I was told that the IO die for Zen 3+ desktop is using “Not quite [the same] IOD as Zen 4 but uses Zen 4 IP” which I take to mean that it will be using DDR5 and it will be on the same node as Zen 4’s IOD. That’s all I have on Zen 3+, so now on to Zen 4.

Zen 4 is what a lot of people are waiting for, and, if the info I have is accurate, that wait will prove to be even more worth it. It is important to note that the one common thread in all Zen 4 chatter I have heard is resounding positivity. From IPC gains over 25%, a total performance gain of 40%, and even possibly (finally) 5GHz all-core thanks to the new (full node) N5 fabrication at TSMC! Now, I can’t say what is true and what is an over-exaggeration, however I was told from a trusted source that a Genoa engineering sample (Zen 4 server chip) was 29% faster than a Milan (Zen 3) chip with the same core config at the same clocks. Factor this in with what I have heard about the possible clock gains that N5 will enable over N7 and Zen 4 sounds like it is going to be a monster of a CPU.

Now I said I had Zen 5 info, unfortunately this comes from a different, less-proven source than my Zen 4 info, however they have said that the jump [to Zen 5 from 4] from will be about as much as Piledriver to Zen 1 design goal, which if you recall to earlier in this article was 40%. I was told from a 3rd source that Zen 5’s original design goal was 2.5 to 3 times the IPC of Zen 1 which roughly lines up with the perspective of a “Piledriver to Zen 1”-like jump.

Looking Forward… AMD’s Problem is Success

If this info is all true, then that puts AMD in a very good position from a performance standpoint. However, the biggest problem for AMD of late hasn’t been one of performance but of wafer supply. With Apple moving off of Intel CPUs and onto their own silicon for (gradually) their entire lineup, the supply of N5 wafers from TSMC for AMD will be less than if Apple had stuck with Intel. Being that AMD is always 2nd in line for new nodes at TSMC behind Apple, this may be a cause for concern moving forward.

Bright Spot: Consoles Out of the Picture with TSMC N5

However, perhaps an even bigger factor for availability on TSMC N5 will be that neither future AMD CPUs nor GPUs will need to compete with console chips as they will remain on N7. Couple this with the rumors that AMD GPUs are also going MCM with RDNA3, and it is possible that Zen 5 and RDNA 3 supply will be better than what we know of the current shortages for Zen 3. This is especially important for the GPU side seeing the current availability issues around RDNA2; the wafer allocation of which has been said to represent low-single digits of TSMC’s N7 supply. This is all of course, wishful thinking. At this point in time, who knows.

One thing is certain however: AMD’s future looks bright.

Author’s Note (2/8/21)

There was a miscommunication between myself and the editor with regards to the performance claims of Zen 4. The only thing that we are claiming with any level of certainty is the claim of a Genoa ES sample being 29% faster then Milan at the same clocks.

We apologise for any confusion this may have caused.

6 thoughts on “AMD’s Past and Future CPUs (Updated with Author’s Note)”

  1. Whada ya think about future zen CPU’s with HBM on die like intel’s SPR? They could be like some sort of L4 cache on package, and with the zen 4 desktop CPU for AM5 Raphael rumoured to have both zen4 cores and navi2, I think we’ll definitely be seeing a chiplet based design where one of those chiplets is graphics.

    I mean if you just look at Navi 10, it’s 40CUs @ 251mm^2. Meanwhile Navi12 (40CUs but w/ HBM 2) comes in at ~210mm^2, so removing all the memory controllers, i/o and central command processor, you get 40CUs withe the ROPs, RBEs etc at 140mm^2. So a 5nm shrink (x1.84) should put that to around 80mm^2, the same size as a zen 3 chiplet. This obviously ignores RDNA2 and RDNA3 and the ‘infinity cache’ taking up more transistors ofc, but I think it’s not an unreasonable assumption that we could see some sort of desktop processor with 8 cores and 40CUs for desktop gaming.

    1. Apex – Passionate about writing, hardware, cars and racing. As an avid PC enthusiast for nearly 20 years, I love understanding what makes computers go 'zoom'.
      Apex says:

      It is certainly a very exciting time, with lots of possibilities for future APUs to move to many configurations, including CPU+GPU+HBM chiplets as you describe. There are talks of some exciting things happening with Zen 4’s I/O die, though I wouldn’t go as far as saying it’s going to have HBM.

      In regards to a “console-like” APU with 40 CU IGP, the main limitation is indeed memory bandwidth. This is the main reason AMD hasn’t bothered to upgrade IGP much with current platforms [Renoir and Cezanne]; ROI just isn’t there when a Vega 7 is still memory bandwidth limited on standard DDR4. DDR5 might change this; notice that RDNA2-powered APUs will finally be arriving with it.

      The idea of an HBM-powered APU is certainly on AMD’s radar; it will be a question of the target audience for such a “high-end” APU and ultimately profitability for them to justify bringing such a product to market.

      1. Well it doesn’t necessarily HAVE to be a HBM chiplet APU so to speak (although that would be pretty cool), but AMD can use gddr or lpddr for more bandwidth in future. Case in point here: the Subor Z+, which had a custom apu for the chinese market with 4 cores/8 threads of zen and 24 Vega CUs on a 256-bit 8GB gddr5 interface. Alternatively, a 256-bit LPPDR5 6400Mhz memory system would achieve comparable bandwidth to the series S, but bloating the i/o die size dedicated to PHY’s. However I would be remiss not to mention large caches as a away to make up for bandwidth. Although this would certainly bloat die sizes further, it can massively reduce the need for bandwidth. TSMC 7N does you about 100mm^2 for 128MB of slower game/infinity cache at 1.6TB/s. One pretty smart way to circumvent that die bloat would be to fab the cache on an older node and just 3D stack it on top of the die. This only makes sense since 5nm only gets a 1.35x cache shrinkage and you’d also have to find a way to deal with the heat generation that using 3d stacking with microbumps presents. While you can deal with microbumps using wire bonding as an alternative, i’ve yet to see any company produce this for high volume manufacturing in a major product so far, then again I think I read on digitimes that Google and AMD are gonna be the first users of TSMCs InFo or SoIC stacking tech, along with apple’s M2 or M3, soooo…. .Just goes to show there’s plenty of options, for working around bandwidth, I just hope the costs are low and the price is right.

      2. Separate cache from the die has actually already been done with Intel’s older Iris parts, eDRAM is a great technology for it. The main problem is price. You can get more out of a wafer that you can sell with separate dies instead of a single APU, especially when you’re packaging extra cache with it. There’s no point in making a high end APU when you can instead make 2 separate products that have a higher ASP and lower overall cost.

  2. Yeah, it was just a theory about cache and you do have a point on costs. I never said it would be necessarily cheap, but it’d certainly be cool. In fact Xe-HPC already does this in a way with the ‘Rambo Cache,’ a bunch of slices of SRAM fabbed on intel’s 10ESF node (same as Alder Lake). I still think the idea of a 40CU GPU chiplet could deffo be a thing. It could easily be paired with an 8 core CPU chiplet for APUs or ganged together with another 40CU chiplet for a gaming GPU. Besides AMD could prob charge $600 for an 8 core/ 40CU combo and easily make decent profit margins whilst stealing intel’s and Nvidia’s marketshare. Though they better act quick, it’s not much of a secret that Nvidia are gonna bring arm+RTX to the PC space eventually. So we’ll have to wait and see I guess.

Comments are closed.