There are only a few corporations in the world capable of declare a historical past in the graphics market courting again to the “golden age” of the 90’s. Among the many handful of survivors of that period are in fact NVIDIA and AMD (previously ATI), however there may be additionally Imagination Applied sciences. The corporate has been round because the early days of graphics accelerators and has seen a shifting trade. Within the early to mid-2000’s the corporate moved away from the desktop house, as an alternative specializing in providing GPU IP optimized for cellular units.
The Cellular GPU IP Market Scenario
The cellular GPU IP market today is a fairly barren panorama; Imagination’s lengthy historical past together with Arm’s rise to a management place with the Mali GPUs has primarily created a duopoly, as smaller distributors resembling Vivante have exited the market. The issue of duopolies like these is that there’s a danger of stagnation if both vendor falls behind in phrases of know-how, and more often than not this comes at the price of the distributors who depend on licensing the IP.
During the last a number of years in explicit, we’ve seen a bigger shift in the cellular market dynamics as Qualcomm has more and more managed to dominate over its rivals. The extra vertically built-in firm produces its personal SoC designs utilizing its personal in-house GPU IP, and because of this isn’t reliant on applied sciences from Imagination or Arm (CPUs are an entire different story).
Extra not too long ago, we’ve seen Apple’s customized GPU efforts make enormous leaps in efficiency and effectivity. Whereas formally we now have no concrete data on the behind-the-scenes particulars of Apple’s design, it’s generally thought to be a design forked-off from Imagination IP again in 2015, which Apple continues to independently develop due to an architectural license. The connection between Apple and Imagination these days continues to be fairly unclear, however the frictions from again in 2017, which in the end result in Imagination placing itself to sale and being purchased by an fairness agency with Chinese language buyers, appear to have calmed down.
The extra necessary scenario to contemplate is that the 2 main GPUs in the cellular market in the present day – Apple and Qualcomm – are designs which might be internally developed by the SoC distributors themselves. This can be a bit troubling for everyone else in the market, as their reliance on IP licensing means they’re restricted in what they’re capable of do. MediaTek had suffered rather a lot in phrases of market share versus Qualcomm for instance (GPU being solely certainly one of many different causes). Different distributors resembling Samsung have evidently additionally realized their scenario, and launched improvement of their very own inner GPU architectures, with a more moderen shock improvement of truly licensing new GPU IP from AMD in order to compete with Apple’s designs.
Introducing Imagination’s new A-Series: “Fastest GPU IP ever created”
With that fairly lengthy foreword and context in thoughts, we come full circle to Imagination in addition to in the present day’s announcement. At the moment, the corporate is revealing its new A-Series GPU IP, describing the brand new product as Imagination’s most necessary GPU and IP launch in 15 years. The wording right here is extraordinarily robust, and the targets of the brand new product appear fairly clear: reverse the development of market share loss in opposition to customized GPUs in the market, and put licensed GPU IP again on the forefront of efficiency.
A-Series is the corporate’s 10th technology GPU structure and represents a giant leap for the corporate. Even one thing easy, just like the product names, have modified. Gone is the longstanding PowerVR branding, in favor of a extra instant give attention to Imagination the corporate. We’ll nonetheless see PowerVR round – Imagination is preserving it to explain their know-how, notably their tile-based deferred rendering method – however after over 20 years it is not a model in and of itself.
As a notice on the brand new GPU IP naming scheme: it’s a deliberate departure from the numerical numbering of the previous, and as an alternative is making an attempt to symbolize a brand new clear begin going ahead. I requested, however sadly the title it isn’t meant to be a intelligent twist on a “Series 10” with the numbering now represented in hexadecimal, a pity I say!
The brand new naming crosses the corporate’s entire GPU vary going from the low-end, mid-range to high-end. Right here we see the continuation of the XE, XM and XT suffixes; nonetheless as in the previous, there are architectural variations between the IPs, with in the present day’s information largely masking the brand new XT collection.
As talked about, the brand new A-Series guarantees to convey some distinctive leaps in efficiency in comparison with previous technology designs, though it’s additionally necessary to level out some footnotes on the matter.
An necessary metric that’s been showcased in the present day is Imagination’s declare that the brand new A-series XT design is 2.5x quicker. “Fast than what” you may ask? The determine represents the brand new structure’s efficiency leap in an ISO-area and course of node comparability in opposition to Imagination’s earlier technology Collection 9 GPU.
The factor to notice is that the corporate isn’t doing precise apples-to-apples comparisons between completely different generations of XT GPUs, however reasonably is making comparisons to the newest broadly out there “Series 9” in the marketplace. This was an unlucky, however but sensible option to make on condition that at the moment there’s no publicly introduced or out there chips which make use of the corporate’s 8XT or 9XTP designs.
The comparability information right here is probably going based mostly on MediaTek’s Helio P90 with the GM9446 – technically a part of the 9-series, but in addition architecturally based mostly on the older Rogue structure. Imagination needed to be extra clear in regards to its space and efficiency claims for the A-Series, and the comparability with the 9XM Rogue-based GPU gives a way more latest and available baseline in opposition to which distributors can independently confirm and evaluate metrics to.
Key enhancements of the A-Series embody a large overhaul over the GPU’s compute models and ALU structure, promising a 4x enhance over a 9XM Rogue GPU. AI is a first-class citizen in the A-Series and we’ll be seeing devoted compute sources ensuing into enhancements of as much as 8x in efficiency. Lastly, the brand new structure is alleged to convey a really massive leap in phrases of energy effectivity, with the brand new design consuming 60% much less energy for the same degree of efficiency.
The brand new structure coated in the present day spans 4 disclosed configurations in the XT collection, in addition to one XM configuration. The XE collection, whereas having seen enhancements, isn’t based mostly on the brand new structure, however reasonably a continuation from the earlier technology.
Imagination’s naming for the brand new A-Series is extraordinarily simple in regards to understanding their positioning a efficiency. The AXT-64-2048 for instance represents the very best finish configuration of the structure in the XT vary, attaining 64 pixels per clock and 2048 FP32 FLOPs per clock. AI TOP efficiency is quadruple the FLOPS, so in this case for an envisioned 1GHz design, the AXT-64-2048 reaches 2 TFLOPs in graphics and compute energy, Eight TOPs in INT8 inference throughput, and 64 Gigapixel/s fill charge.
The figures scale down the vary with corresponding efficiency with the AXT-48-1536, AXT-32-1024, AXT-16-512 and eventually the AXM-8-256. The AXE-1-16 is a separate, smaller extra personalized microarchitecture.
Imagination envisions the AXT-32-1024 to be the normal sweet-spot goal for a premium cellular GPU SoC in smartphones, whereas the bigger configurations may be used in bigger form-factor units.
The corporate calls the A-Series “the GPU of everything”, and there’s certainly a ton of adjustments and options that make up the brand new IP. We’ll be masking just a few facets of the corporate’s disclosed massive matrix of options above, beginning off with one of many greatest adjustments, the GPU’s ALU structure.